Summaries Of Past Works Biology Essay

Published: Last Edited:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Many researchers have done their research on fault detection and diagnosis of induction motor and there are many fault detection techniques that have been developed. This section will present a brief discussion on the past works that are done on fault detection and diagnosis of induction motor.

Traditionally, techniques such as the measure of the tangent of the delta angle, the measure of the polarization index or the measure of the insulating strength with the use of a megohmmeter to establish dielectric features in the insulators of electric winding machines. All of them are characterized by the capacity to submit the winding insulator to a voltage above the nominal. In this way, fault currents can be measured and the dielectric capacity of the insulating material can be settled. The impulse testing has become commonly used recently. It consists in the use of high tension pulses on the windings of a machine and the analysis of a transitory response. As a result, a fault in a winding stator can be found when there are differences among the responses of each coil or phase in the machine. All these techniques are very effective and are capable of establishing the estate of the insulator and estimate its useful life. However, its use is quite limited because the diagnostic has to be done with the machine out-of-service. With respect to fissures or cuts in bar rotors, the detection was done through the study of motor vibrations or observing fluctuations of low frequency in stator currents. In both cases, the fault must be found in an advanced estate to be seen ostensibly. Bearing faults are generally detected by the study of vibrations. If an accelerometer is used, it is possible to control the intensity and frequency of vibrations in the motor. (Verucchi, Acosta & Benger, 2008)

Apart from this, there are some on-line faults detection methods. These include the Complex Park Vector. The well-known Park transformation allows showing the variables of a three-phase machine through a system of two quadrature shafts. The components of the stator current in a reference system formed by two octagonal shafts which are fixed to the stator (shafts D and Q) are obtained by the following reports:



where, and are the currents of the phases A, B, and C of the stator. (Marques Cardoso, 1993) In normal conditions, when a motor without faults works through a three-phase system of sinusoidal currents, balanced and of positive sequence, the components of the Park vector form a circle centered in the origin of the plane D-Q with a constant radius, as it is shown in Figure 2.1. In case of a short circuit in winding stators the motor behaves as an unbalanced load and the stator currents stop being a balanced system. Such unbalances cause an oscillation in the radius of the Park vector and turn into elliptical shapes. Figure 2.2 shows the results obtained in a motor with two coils of the a phase (over 16), in short circuit. The bending on the main shaft of the ellipsis shows the phase in which the fault occurred (Verucchi et al. 2008).

Figure 2.1: Geometric locus of the Park vector for normal working condition.

Figure 2.2: Geometric locus of the Park vector for coils in short circuit.

Another way of finding a fault consists in watching the value that the radius of the vector takes through time. As the radius moves between its minimum and minimum value twice in each cycle of the power net, its analysis in Fourier series, shows a component in twice the frequency of the power net. The amplitude of this component shows the relevance of the fault. In Figure 2.3 is shown the example of a machine with faults in the stator coils.

Figure 2.3: Harmonic analysis of the vector Park module.

Some authors propose the use of the Park vector method to detect eccentricities in the rotor. In this case, what is obtained is a double circle with the centers displaced, as it is shown in Figure 2.4. This is because the geometric locus of the vector shows a complete circle for each cycle of the net. In machines with more than one pair of poles, the circles that correspond to different angular positions of the rotor are overlapped. In a four pole machine, for example, the vector will present a circle for each half rotor turn. If the gap of the machine is not the correct one due to eccentricities, two consecutive circles will not match exactly and the difference between both will show a faulty state. (Verucchi et al. 2008)

Finally, faults in shorting bars are present when there is an overlapping of concentric circles with an oscillating radius as it is shown in Figure 2.5. This is the case of a rotor with 3 cut bars, over a total of 58 working at half load. The frequency with which the current stator vector moves between its maximum and minimum radius is the same as 2.s.f and it can be seen in the frequency spectrum of the radius in the vector.

Figure 2.4: Geometric locus for the Park vector currents for eccentricities of the rotor.

Figure 2.5: Geometric locus for the Park vector currents for two cut bars and ½ of the nominal load.

Besides, axial flow is another method to detect the induction motors fault on-line. In any induction machine, even in normal working conditions, there are small unbalances in the currents. These appear due to manufacture imperfections and the unbalances of the power net. This causes negative current sequences in the rotor and, besides, the unbalance of the currents on the end of the coils cause a flow in the axial side of the motor. This flow, which is the result of the stator currents, has the same harmonics as those, and as a consequence, it allows identifying the faults. A coil placed on the end of the motor in a concentric way with its shaft, allows seeing the axial flow and diagnosing faults (Joksimovic & Penman, 1994). So, through the analysis of the frequency spectrum of the axial flow of the motor, short circuits in the winding stator, eccentricities and cut bars in the rotor can be detected. Although the method is not totally non-invasive because it requires the use of a sensor on the back part of the motor, it has advantages with relation to the study of stator currents to detect cut bars in the rotor and short circuits between loops. In motors of half power it is more complex to detect the stator currents and in those cases, it is easier to use a coil of axial flow. This technique has been successfully tested even with motors that work with frequency variations. (Verucchi et al. 2008)

Another on-line method is using Wavelet Packet Transform (WPT) and Multiple Support Vector Machine (SVM). In this method, both vibration and motor current signature analysis are performed to detect the mechanical and electrical faults. Multi-scale decomposition process using WPT is performed on the obtained signal to extract the features. The extracted features are given to a classifier to identify whether a fault has occurred. If a fault exists, it identifies the fault location and isolates it. Multiple SVM using the one-against-others approach is used to obtain multi-class classification of fault. (Aravindh et al. 2010)

Figure 2.6: Schematic representation of the WPT and Multiple SVM method.

Figure 2.6 represents the flow of the signals from motor to classify the fault and proceed with the condition monitoring. The signals acquired from the motor parts with the help of sensors are analyzed with WPT to extract the features. These features are given to the trained classifier to state the fault location if fault exists. The sensor used for vibration analysis is the accelerometer and for MCSA is the Hall Effect sensor. When induction machine runs at fault conditions, some special components occur at the stator current. When the rotor bar breaks, the feature components in the motor stator line current are


where is supplied frequency and s is the slip. When rotor eccentricities occurs, the feature frequencies in the motor stator line current are


where R is the number of rotor bars, p is the pole pairs, =0 is in static eccentricity, =1 is in dynamic eccentricity and =1, 3, 5, ..

From equation 2.4, it is evident that the fault conditions in induction machines can be easily detected by monitoring the stator currents.

Wavelet transform is known as multi scale decomposition process. Decomposition of the signal provides frequency information about the discrete time signal. This is done by convolution of signal with high pass and low pass filters. The convolution of signal with high pass and low pass filter provides two vectors such as the detail coefficient and approximation coefficient. The detail coefficient provides high frequency information whereas approximation provides low frequency information. In wavelet trasnform, multi scale decomposition is basically done only on approximation coefficient. So time-frequency localization of any fault is not more accurate. Hence, WPT is proposed to obtain more flexible and wider base result of the fault location. Wavelet packet transform performs multi scale decomposition on both approximation and detail coefficient. Thereby good localization is obtained. (Aravindh et al. 2010)

SVMs are binary classifiers, which are designed to separate only two classes from each other. But for the fault detection in induction motor we are in need of multi-class SVM. Such multi-class SVM is obtained by decomposing the multi-class problem into several number of binary class problems. Then classifiers are trained to solve the problems assigned to each binary SVM. Finally the classifiers are coupled to reconstruct the solution of the multi-class problem from the outputs of the individual classifiers. The multi-class classification structure is basically given by one-against-others approach. This approach can also be defined as one-versus-all. In this method, one class is compared with all other classes in multi- class structure. Classification of new instances using one-versus- all method is done, in which the classifier with the highest output function assigns the class. (Ruiming Fang & Hongzhong Ma, 2006)

Apart from the methods stated above, higher order spectrum (HOS) is used for fault detection and diagnosis of induction motor. Any fault either in the stator or the rotor may distort the sinusoidal response of the motor RPM and the main frequency so the MCSA response may contain number of harmonics of the motor RPM and the mains frequency. Hence the use of the HOS, namely the bispectrum of the MCSA has been proposed here. Additionally, the bispectra with the unwrapped phase angle along its frequency has been applied to analyze motor faults. It has been observed that it not only detects the early fault but also indicate the severity of the fault to some extent. (Juggrapong Treetrong, 2010)

Figure 2.7 The experimental setup for Higher Order Spectrum Method.

Figure 2.7 shows the experimental setup for the data sample collection in HOS method. The setup consists of an induction motor (4kW, 1400RPM) with load cell with a facility to collect the 3-phase current data directly to the PC at the user define sampling frequency. The experiments were conducted for these 3 different conditions - Healthy, Stator Fault and Rotor Faults at different load conditions. The data were collected at the sampling frequency of 1280 samples/s. The stator fault was simulated by the short circuits these include 5 turn shot circuit, 10 turn short circuit and 15 turn short circuit whereas the rotor fault by the broken rotor bars. (Juggrapong Treetrong, 2010)

The other fault detection technique of induction motor is by using set-membership filtering and Kalman filtering. Estimation techniques are more and more used in fault detection and diagnosis of induction motors. Most state estimation problems are solved via a stochastic approach. Measurement noise, disturbances and model errors are assumed to be a realization of a random process. These estimation techniques require the knowledge of stochastic characteristics of the different disturbances and a Kalman filter is often used to solve such a problem. However, in some situations, it can be more natural to consider a geometric approach, assuming only that the perturbations belong to known bounded sets with no hypothesis on their distributions inside these sets. This bounded-error approach describes the set of all the states that are consistent with the model, the data and the error bounds. All elements of this feasible set are then candidate solutions for the estimation problem. The set thus obtained may become extremely complicated. In order to be computed in real time, this feasible set is recursively characterized by the smallest ellipsoid that encloses it. Usually, the size of an ellipsoid, characterizing the state estimation uncertainty, is measured by its volume, which is proportional to the square of the product of the semi-axe lengths and corresponds to the determinant criterion. However, this criterion presents some disadvantages; this is why an alternative criterion, namely the trace criterion which corresponds to the sum of the squares of the semi-axe lengths of the ellipsoid, is also considered. (Durieu, Loron, Sedda & Zein, 1999)

The fault detection is based on the electrical model of the induction motor. Therefore, the Kalman filter and the set-membership will only detect faults that have a significant effect on the electrical behavior of the motor. Both approaches take into account model approximations and measurement errors to estimate the motor state. To obtain fault detection with a high sensitivity, these uncertainties must be minimized. Moreover, the working conditions have a large influence on the fault detection sensitivity. For instance, a broken bar cannot be detected if the motor torque is null. Nevertheless, the proposed approaches offer practical advantages: no supplementary instrumentation is required and the state estimation can also be used for the torque control of the motor. The results indicate that both the bounded approach and the Kalman filtering are able to detect electrical faults of an induction motor. These algorithms require an adequate noise characterization to be efficient. To take into account natural variations of the electrical parameters (essentially due to the temperature influence) adaptive schemes have to be implemented. This can be obtained with an extended Kalman filter or by a similar extension of the se-tmembership filtering. (Durieu et al. 1999)

The next method is by using impedances of inverse sequence (IIS). Depending on the theories of symmetric components, all three-phase unbalanced system can turn into two three phase unbalanced systems of different sequence plus a group of monophasic phasors. The former are systems of direct and inverse sequence and the latter, a system of zero sequence. So, with the complex values of voltages and the currents of a three-phase system, the components of the sequence systems can be found starting from the relation given by the Equation 2.5 and 2.6.



The sub indexes A, B, and C, refer to each one of the components of the phase of the real system, while 0, 1, and 2, show the components of the systems of zero sequence, direct and indirect respectively. The constant a is given by:


The relations between the sequence currents and voltages are determined by the impedances of direct, inverse and zero sequence as follows:


Taking separately each one of the systems, the impedances of direct, inverse and zero sequence can be defined. In the case of induction motors, taking into account that they are generally connected in triangle, or in star with a disconnected center, the component of the zero sequence is null. In this way, the asynchrony motor will be identified by the impedances of direct and inverse sequence as follows:



While the impedance of the direct sequence is independent of the load state of the motor, the impedance of the inverse sequence is practically independent and very susceptible to short circuits in winding stators. Consequently this one is the most suited for the diagnosis of this kind of fault. Figure 2.8 shows an example of application (Verucchi et al., 2008), where there are consecutive measures of the impedance of the inverse sequence of the motor, first in normal conditions and then with a minor fault in one of the stator coils. The accuracy with which the value can be measured depends on the unbalance level of the power. This technique needs to count with a great variety of values for the different slippings. With that data the value of the inverse current sequence can be calculated and then compared with its mean value. An important difference between both currents shows a fault in the winding stator. (Verucchi et al., 2008)

Figure 2.8: Application of Impedances of Inverse Sequence (IIS)

2.2 Fundamental Background

2.2.1 Artificial Intelligent (AI)

AI is one of the areas of computer science focusing on inventing machines that can resemble on humans intelligent behavior. In the past, the ability to create intelligent machines has intrigued humans. Today, with the development of computer technology and more than 50 years of research in AI, the dream of smart machines is becoming a reality. As a result, machines which can mimic human thought, play football and even beat the best human chess player are created. The applications of AI include machine vision as well as expert system.

At 1950, Turing test was introduced by Alan Turing in order to test on a machine ability to demonstrate intelligence. Turing test is a test designed to determine whether a machine can pass a behavior test for intelligence. Turing defined the intelligent behavior of a computer as the ability to achieve human level performance in cognitive tasks. During the test a human interrogates someone or something by questioning it via a neutral medium such as a remote terminal. The computer passes the test if the interrogator cannot distinguish the machine from a human.

Figure 2.9 Turing Test

Machine learning has been central to AI research from the beginning. Unsupervised learning is the ability to find patterns in a stream of input. Supervised learning includes both classification and numerical regression. Classification is used to determine what category something belongs in, after seeing a number of examples of things from several categories. Regression takes a set of numerical input/output examples and attempts to discover a continuous function that would generate the outputs from the inputs. In reinforcement learning the agent is rewarded for good responses and punished for bad ones. These can be analyzed in terms of decision theory, using concepts like utility. The mathematical analysis of machine learning algorithms and their performance is a branch of theoretical computer science known as computational learning theory.

2.2.2 Artificial Neural Network (ANN)

In general, machine learning involves adaptive mechanisms that enable computers to learn from experience, learn by example and learn by analogy. Learning capabilities can improve the performance of an intelligent system over time. Machine learning mechanisms form the basis for adaptive systems. ANN is one of the types of machine learning.

ANN can be defined as a model of reasoning based on the human brain. The brain consists of densely interconnected set of nerve cells called neurons. The human brain incorporates nearly 10 billion neurons and 60 trillion connection, synapse, between them. By using multiple neurons simultaneously, the brain can perform its function much faster than the fastest computers in existence today.

Although each neuron has a very simple structure, an army of such elements constitutes a tremendous processing power. A neuron consists of a cell body, soma, a number of fibers called dendrites, and a single long fiber called the axon. While dendrites branch into a network around the soma, the axon stretches out to the dendrites and somas of other neurons. Figure 2.10 shows the schematic drawing of a neural network.

Figure 2.10: Biological neural network

Signals are propagated from one neuron to another neuron by complex electrochemical reactions. Chemical substances released from the synapses cause a change in the electrical potential of cell body. When the potential reaches its threshold, an electrical pulse, action potential, is sent down through the axon. The pulse spreads out and eventually reaches synapses, causing them to increase or decrease their potential. Neurons demonstrate changes in the strength of their connections. Connections between neurons leading to the "right answer" are strengthened while those leading to the "wrong answer" weaken. As a result, ANN has the ability to learn through experience.

An ANN consists of a number of a number of interconnected processors called neurons. The neurons are connected by weighted links passing signals from one neuron to another. Each neuron receives a number of input signals through its connection. Neurons never produce more than a single output signal. The outgoing connection splits into a number of branches that transmit the same signal. The neurons are connected by links and each link has a numerical weight associated with it. Weight are the basic means of long-term memory in ANN. They express the strength. ANN "learns" through repeated adjustments of these weights. Table 2.1 shows the analogy between biological and artificial neural network. Figure 2.11 shows the architecture of ANN.

Table 2.1 Analogy between biological and artificial neural network

Biological Neural Network

Artificial Neural Network









Figure 2.11: Architecture of Artificial Neural Network

ANN offers several advantages over conventional computing:

Trainability and generalization: ANNs can be trained to form associations and learn underlying connections between any input and output pattern. This experiential knowledge can then be used to generalize to new cases previously unseen by the classifier.

Non-linearity: ANNs can compute nonlinear, nonparametric functions of their input, enabling them to perform arbitrarily complex transformations of data.

Robustness: ANNs are tolerant of both physical damage and noisy data.

2.2.3 Multilayer Perceptron (MLP)

MLP is a subset of ANN, defined as a system of massively distributed parallel processor (consisting of simple processing units called neurons) that have natural tendency for storing and utilizing experiential knowledge. Generally, the MLP learns the relationship between a set of inputs and outputs by updating internal interconnections called weights using the back-propagation algorithm. (Nadiah, Fatimah, Aiman, Ihsan & Azlee, 2010)

In MLP, the units are arranged in interconnected layers: one input layer, one (or more) hidden layers, and one output layer. (Nadiah et al., 2010)The numbers of input and output units are typically fixed, since they depend on the input and desired output. However, the training algorithm and the number of hidden units are adjustable, and can be set so that it maximizes the performance of the MLP.

The interconnections between the MLP layers (weights) are typically initialized at random prior to training. The initialized weights represent the initial points in which the MLP begins the search for the solution. It is because of this, the value of the random numbers affects the network convergence. A too large or too small initial weight values would slow down or prevent convergence. (Nadiah et al., 2010)

Learning in a multilayer network proceeds the same way as for perceptron. A training set of input patterns is presented to the network. The network computes its output pattern and if there is error or in other words a difference between actual and desired output patterns, the weight are adjusted to reduce the error.

2.2.4 Fuzzy Min-Max (FMM) Neural Network

FMM neural network is a supervised learning neural network classifier that utilizes fuzzy sets as pattern classes is described. Each fuzzy set is an aggregate (union) of fuzzy set hyperboxes. A fuzzy set hyperbox is an n-dimensional box defined by a min point and a max point with a corresponding membership function. The min-max points are determined using the FMM learning algorithm, an expansion-contraction process that can learn nonlinear class boundaries in a single pass through the data and provides the ability to incorporate new and refine existing classes without retraining. The use of a fuzzy set approach to pattern classification inherently provides degree of membership information that is extremely useful in higher level decision making. (Simpson, 1992)

FMM neural network creates classes by aggregating several smaller fuzzy sets into a single fuzzy set class. FMM neural network are built using hyperbox fuzzy sets. A hyperbox defines a region of the n-dimensional pattern space that has patterns with full class membership. A hyperbox is completely defined by its min point and its max point, and a membership function is defined with respect to these hyperbox min-max points. The min-max (hyperbox) membership function combination defines a fuzzy set, hyperbox fuzzy sets are aggregated to form a single fuzzy set class, and the resulting structure fits naturally into a neural network framework; hence this classification system is called a FMM classification neural network. Learning in the FMM classification neural network is performed by properly placing and adjusting hyperboxes in the pattern space. FMM classification neural network recall consists of computing the fuzzy union of the membership function values produced from each of the fuzzy set hyperboxes. (Simpson, 1992)

There are several properties that a pattern classifier should possess. Each of these properties has motivated a portion of the development of the fuzzy min-max classification neural network. Some important properties are:

On-Line Adaptation: On-line adaptation or on-line learning refer to the ability of a pattern classifier to learn new classes as well as refine existing classes without affecting old class information. Some pattern classifier use off-line classification. Off-line learning can increase the training time. This is because when new information is added to the classification scope, a complete retraining is required to classify both the old and new information.

Overlapping Classes: Pattern class tends to overlap. A pattern classifier should have the ability to form a decision boundary that minimizes the amount of misclassification for all of the overlapping classes. The most prevalent method of minimizing misclassification is the construction of a Bayes classifier. Unfortunately, to build a Bayes classifier requires knowledge of the underlying probability density function for each class. (Simpson, 1992) This information is often unavailable. For the on-line adaptation case, constantly tuned is required to represent the current state of the data being received.

Training Time: All researchers are concern with the time required for the pattern classifier to classify the information. A very desirable property of a pattern classification approach able to learn decision boundaries is a short training time. (Simpson, 1992)

Fuzzy sets are used to manipulating data that are not precise or called fuzzy. A fuzzy set is formed from the union of a collection of fuzzy sets. This fuzzy set is used to describe each class. The fuzzy class is formed by aggregating the collection of hyperbox fuzzy sets that belong to each class. Fuzzy sets and neural network can be effectively merged by utilizing neural network nodes as fuzzy sets and using fuzzy set operations during learning.

For FMM neural network, hyperboxes, defined by pairs of min-max points, and their corresponding membership functions are used to create fuzzy subsets of the n-dimensional pattern space. The majority of the processing is concerned with finding and fine-tuning the boundaries of the classes.

The FMM classification model is formed using hyperbox fuzzy sets. The size of a hyperbox is controlled by θ, which varies between 0 and 1. When θ increases from a small to large value, the number of hyperboxes created is reduced. (Simpson, 1992)The membership function is set with respect to the minimum and maximum points of a hyperbox, and to the extent which a pattern fits in the hyperbox. For an input pattern of n-dimensions, a unit cube In is defined, and the membership value ranges between 0 and 1. The definition of each hyperbox fuzzy set Bj is:


where Vj is the minimum and Wj is the maximum points. Figure 2.12 illustrates the minimum and maximum points of a three dimensional box.

Figure 2.12 A min-max hyperbox Bj = {Vj,Wj} in I3

The combined fuzzy set that classifies the Kth pattern class, Ck, with the definition of a hyperbox fuzzy set, is:


where K is the index set of those hyperboxes associated with class k. The training process is concerned with finding and fine-tuning the boundaries of the classes. An example of the decision boundary in a two-class classification problem is shown in Figure 2.13.

Figure 2.13 : An example of the FMM decision boundary of a two-class problem

In FMM, the learning algorithm allows hyperboxes of the same class to have overlapping, while overlapping among different classes is to be eliminated. The membership function for the jth hyperbox bj (Hh), 0 ≤ bj (Hh) ≤ 1, measures the extent to which the hth input pattern Ah falls outside hyperbox Bj. This can be considered as a measurement of the extent on each component is greater (or lesser) than the maximum (or minimum) point value along each dimension that falls outside the min-max bounds of the hyperbox. The function that meets all these criteria is the sum of two complements-the average amount of the maximum point violation and the average amount of the minimum point violation. (Simpson, 1992) The resulting membership function is:


where, is the hth input pattern, is the minimum point for Bj,is the maximum point for Bj, and γ is the sensitivity parameter that controls how quickly the membership values decrease when the distance between Ah and Bj increases. (Simpson, 1992)