Intelligent Fault Recognition and Diagnosis for ...

International Journal of Software Science and Computational Intelligence, 3(4), 67-83, October-December 2011 67

Intelligent Fault Recognition and Diagnosis for Rotating Machines using Neural Networks Cyprian F. Ngolah, Sentinel Trending & Diagnostics Ltd., Canada Ed Morden, Sentinel Trending & Diagnostics Ltd., Canada Yingxu Wang, University of Calgary, Canada

ABSTRACT Monitoring industrial machine health in real-time is not only in high demand, it is also complicated and difficult. Possible reasons for this include: (a) access to the machines on site is sometimes impracticable, and (b) the environment in which they operate is usually not human-friendly due to pollution, noise, hazardous wastes, etc. Despite theoretically sound findings on developing intelligent solutions for machine conditionbased monitoring, few commercial tools exist in the market that can be readily used. This paper examines the development of an intelligent fault recognition and monitoring system (Melvin I), which detects and diagnoses rotating machine conditions according to changes in fault frequency indicators. The signals and data are remotely collected from designated sections of machines via data acquisition cards. They are processed by a signal processor to extract characteristic vibration signals of ten key performance indicators (KPIs). A 3-layer neural network is designed to recognize and classify faults based on a pre-determined set of KPIs. The system implemented in the laboratory and applied in the field can also incorporate new experiences into the knowledge base without overwriting previous training. Results show that Melvin I is a smart tool for both system vibration analysts and industrial machine operators. Keywords:

Artificial Neural Networks, Data Acquisition, Industrial Applications, Intelligent Fault Recognition, Remote Machine Condition Monitoring, Vibration Analysis

1. INTRODUCTION Various studies have been done on the application of neural networks to provide intelligent solutions for machine condition monitoring and fault diagnosis (Samhouri et al., 2009; Peng, 2004; Srinivasan, 2003; McCormick & Nandi, 1996; Umesh & Srinivasan, 2005; Gerlad et al., 2000; Aravindh et al., 2010; Yangwen, 2009; Tetsuro & Wang, 2008; Balakrishnan & Honavar, 1995; Saxena & Saad, 2004; Shiroishi et al., 1997). For example, Srinivasan (2003) described a neural network to identify the approximate location of damage due to cracks through the analysis DOI: 10.4018/jssci.2011100105 Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.

68 International Journal of Software Science and Computational Intelligence, 3(4), 67-83, October-December 2011

of changes in the neural frequencies. McCormick and Nandi (1996) described neural network methods for automatically classifying machine conditions from the vibration time series. Umesh and Srinivasan (2005) carried out studies on experimental data simulation of faults such as parallel misalignment, angular misalignment, unbalance, crack, light and heavy rubs, looseness and bearing clearance. Saxena and Saad (2004) carried out research on fault diagnosis in rotating mechanical systems using Self-Organizing Maps. Despite these theoretically sound findings on developing intelligent solutions based on neural networks for machine condition based monitoring, there are virtually no commercial tools in the market that can be readily used. This paper reports the design and implementation of a tool that uses neural network technology for deriving intelligent fault recognition based on observation of changes in various fault frequency indicators. An Artificial Neural Network (ANN) is an intelligent information processing paradigm that imitates the biological nervous systems. In the brain, electrochemical signals pass between neurons through the synapses. In neural network analysis, the signal between neurons is simulated by interlinked circuits and software, which apply weights to the input nodes and use an activation function to scale the neuron’s output to an acceptable range. Thus, the basic element in an ANN is the artificial neuron node, which receives and combines signals from many other neurons through input paths in the same way as the biological neuron receives and processes signals via axons. The output of a neural network is therefore a linear combination of inputs, determined by weights that simulate synapses in the biological neural system. The weights are usually selected such that when the element is presented with input data, the output is as close to the desired output as possible. A typical neuron in an ANN has two modes of operation - the training mode and the using mode. In the training mode, the neuron can be trained to fire (or not) for particular input patterns, while in the using mode, when a taught input pattern is detected at the input, its associated output becomes the current output of the ANN. If the input pattern is not yet present in the trained set of the input patterns, a predetermined firing rule is used to determine whether to fire or not. Neural networks make use of two types of values – weights and thresholds. Weights define the interaction between the neurons, while thresholds define what it takes to get a neuron to fire. In condition based monitoring systems, data is acquired from various types of machines via sensors. By analyzing the collected vibration data, possible faults such as unbalance, bent shaft, shaft crack, bearing clearance, rotor rub, misalignment, and looseness can be identified. Because a lot of industrial machines operate in remote areas, they are often inaccessible at certain periods of the year. A system that remotely collects vibration data in real-time, processes it, detects possible faults and makes recommendations on predictive maintenance will certainly reduce downtime and associated consequences. This paper is an extended version of Ngolah et al. (2011) and presents the development of a mechanism for fault recognition and classification based on fault frequency indicators using the neural network technology. In this paper, Section 2 gives an overview of the set of tools that have been developed to provide intelligence on online machine condition based monitoring. Section 3 describes the design and implementation of a neural network for recognizing machine faults. Section 4 presents experimental results of the system and Section 5 draws a set of conclusions and presents future directions.

Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.


2. THE OVERALL STRUCTURE OF THE MELVIN I SYSTEM The intelligent fault recognizer known as Melvin I is a component of a large system (Melvin) developed to offer an open and expandable online machine condition monitoring system for rotating machinery. The overall system is divided into four subsystems, which perform data acquisition, signal conditioning and processing, fault recognition and prediction, and fault reporting. Figure 1 presents the top level view of the four subsystems of Melvin. The next two sub-sections present a detailed discussion on the data acquisition and signal conditioning/processing mechanisms used in this work. A detailed discussion on the design and implementation of the fault recognition component is presented is Section 3.

2.1. Data Acquisition for Melvin The work reported in this paper has been carried out on reciprocating compressors used for processing natural gas. A reciprocating compressor uses pistons to compress gas. On each compressor, there is a central crankshaft that can drive up to six pistons inside cylinders. The crankshaft is generally driven by an external motor. As the pistons draw backwards, gas is injected from an intake valve in the compressor. This gas is in turn injected into the cylinders of the pistons, and is then compressed by the reciprocating action of the pistons. The compressed gas is discharged either to be used immediately by a pneumatic machine, or stored in compressed air tanks. To monitor the health of these machines, data must be collected from suitable locations on it for processing. Figure 2 shows a typical reciprocating compressor, where each of the possible data collection points (C4A, C4V, C2V, 1VC, etc.) is indicated. Common bearing locations include drive end horizontal, drive end vertical, drive end axial, outboard horizontal, outboard vertical,

Figure 1. System overview of Melvin



Figure 2. Data collection points on a reciprocating compressor

and outboard axial. Depending on the types of machines and faults being monitored, data may be collected from all or only a subset of these points. Vibration signals are collected using piezo-electric transducers attached to locations on a rotating machine based on the VDI 2056 standard guidelines (Kumaraswamy et al., 2002). In reciprocating machinery, many mechanical characteristics change with the machine’s speed. As the rotational speed changes, the frequency bandwidth of each individual harmonic gets wider and some frequency components can overlap. Thus, a simple Fast Fourier Transform power spectrum analysis may not be adequate to identify characteristic vibration components. To take care of this situation, a tachometer is mounted close to the rotating machine and used to measure the number of pulses per revolution. The measured value is then used to calculate the speed of the machine in real-time as the machine rotates.

2.2. Signal Processing and Conditioning for Melvin Raw time waveform data is collected in the time domain through the piezo-electric accelerometers. This has to be converted into the frequency domain so as to isolate the various frequencies of interest. To be able to acquire data, suitable data acquisition parameters such as sampling rate (fS) and the number of samples (N) in a given window must be determined. The faster the sampling rate, the more accurate the data acquired. However, a higher sampling rate entails intensive processing time of the processor. The sampling rate used in this work is 12,800 samples per second, but different sampling rates can be used depending on the problems being monitored. Ten key performance frequency indicators (KPIs) were calculated based on the nature and type of faults being monitored by our system. One of the most common analysis methods to analyze noise and vibration signals is the fast Fourier Transform (FFT) analysis (Loan, 1992). This method identifies and quantifies the frequency components of noise and vibration signals in a time domain signal. However, in reciprocating machinery, many mechanical characteristics change with speed. As such, machinery noise and vibration tests also require a run-up or coastCopyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.


Table 1. KPI Setup for Melvin I

down test to be conducted as well. When the rotational speed changes, the frequency bandwidth of each individual harmonic also gets wider, thus causing some frequency components to overlap. This can cause the resulting FFT power spectrum to become blurred. In such a situation, the FFT power spectrum approach will no longer help in identifying characteristic vibration components. In this paper, the order analysis technique has been used to calculate some characteristics frequency indicators that can be affected by the machine’s speed. Order analysis (Gade et al, 1995) is a technique for analyzing noise and vibration signals in rotating or reciprocating machinery. Such machinery typically has a variety of mechanical parts such as a shaft, bearing, gearbox, fan blade, belt, etc., where each mechanical part generates unique noise and vibration patterns as the machine operates, with each part contributing a unique component to the overall machine noise and vibration. In this work, depending on the type of faults, the key performance frequencies are calculated either in frequencies or orders. The first order is usually the machine’s rotational speed and order n is defined as n times the rotational speed of the machine. Table 1 shows the different KPIs used in this experiment and their associated types. Certain frequency components (shown in column 1 of Table 1) such as the overall vibration (OVR), 120 Hz, and the high frequency component (1-5KHz) are measured in frequency while others such as SubHarm, 1X, 2X, 2X, 3-5X, 5-10X, 10-20X, and 20-50X are in orders. Table 1 shows the different parameters used in deriving information about each key performance indicator. Column 1 is the name of the KPI, Column 2 is a description of what is being measured (displacement, velocity or acceleration), Column 3 shows the type (order or frequency), Column 4 shows the units of measurements, Columns 5 and 6 show the band limits for each frequency component, Columns 7 and 8 show respectively the alert and alarm limits for each KPI, Column 9 shows the database field name on



which data for that KPI is stored, and Column 10 shows possible faults associated to that KPI when it is either in alert or alarm. The collected data is stored in a control server called the Object Linking and Embedded (OLE) for Process Control (OPC) to display real-time trends, and also on an SQL server for fault diagnosis and prediction by Melvin. OPC (http://www.opcfoundation.org/) is a software interface standard that permits Windows programs to communicate with industrial hardware devices. This paper only reports the development of the fault recognition component (Melvin I), which recognizes and classifies faults based on fault frequency indicators.

3. THE NEURAL NETWORK FOR FAULT RECOGNITION IN MELVIN I On each reciprocating compressor (Figure 2), there may be many data collection points. Each data collection point requires a sensor. As described in Section 2, each sensor collects ten data points which are stored in a database and subsequently used to diagnose possible faults. Thus, if there were 1000 machines with data being collected on ten different points for each compressor, then there will be a total of 10x10x1000 (100,000) numerical values collected simultaneously. The number gets bigger if the number of machines to monitor increases. With this large number of numerical vales being collected, it becomes very difficult to track changes on individual values in real-time. There is therefore, a need for a solution that simultaneously tracks changes in patterns rather than individual numerical values. The approach adopted in this paper is the use of the neural network technology to implement a pattern recognition algorithm which receives KPI data from all the sensors and processes it to give accurate diagnosis of possible faults. The next sections present the design of a neural network mechanism developed for fault diagnosis and prediction in Melvin.

3.1. Design of Melvin I The kernel architecture of Melvin I is designed as a multilayer perceptron with feedforward neural network using the Backpropagation training algorithm (Rumelhat et al., 1995; Heaton, 2008). The neural network has three layers: the input, hidden and output layers. The input layer contains ten neurons that receive inputs corresponding to ten KPI values as described in Table 1. The hidden layer has five neurons, while the output layer has a single neuron. Figure 3 shows the layout of neurons for a single sensor. Input neurons receive KPI values from an SQL database. The activity of each hidden neuron is determined by the activities of the input neurons and the weights on the connections between the input and hidden neurons. In Figure 3, Wijk represents weight k for neuron j in layer i. In this architectural design, the following factors were taken into consideration: 1. Rapid and stable training: The system is designed to be capable of incorporating new experiences with minimal training time, and without overwriting previous training experiences. 2. Hypothesis generation capability: The system is designed to be capable of suggesting additional fault possibilities in order to fully identify a fault. This is realized by having an interactive component of the system where experts/users can provide further input in situations when the system is unable to diagnose a certain fault. The knowledge obtained from such interaction is incorporated into the knowledge base for future reference.



Figure 3. Network layers in Melvin for a single sensor

3.2. Implementation of Melvin I Melvin I performs three major tasks: data acquisition, feature extraction, and fault identification. Data acquisition involves the collection of data that indicates the health conditions of machines. The data acquisition mechanism has been described in Section 2. Feature extraction and the techniques for implementing Melvin I are described in the following subsections. 1) Feature Extraction Feature extraction is the process of detecting features hidden in the acquired data using signal transformation technologies. According to empirical data, machine faults are associated with certain frequencies based on each machine’s rotating speed. With frequency domain signals,



it is necessary to extract the amplitudes corresponding to these frequencies. By observing the changes in amplitudes, conclusions can be drawn on possible machine faults. The extraction of the different features of interest was discussed in Section 2. The next section describes how the extracted features can be used in the development of Melvin I. 2) Training of Melvin I The training of a neural network requires a suitable training set and a training algorithm. The training algorithm used in the implementation of Melvin is based on the Backpropagation algorithm (Rumelhat et al., 1995; Heaton, 2008) and is fully described below. Three training sets are employed in the training of Melvin as described below. (i) Data Normalization In this implementation, we have used the alert values in Table 1 to normalize the input values into each neuron, which is set to 1 if the value is above the alert threshold and 0 otherwise. Thus, an input pattern of {1,0,0,0,0,0,0,0,0,0} indicates that the value of the first KPI (OVR) is above the alert threshold, a pattern such as {0,1,1,0,0,0,0,0,0,0} shows that the values for the SubHarm and 1X key performance indicators are above the alert thresholds, and a pattern such as {0,0,0,1,0,0,0,1,1,1] shows that the values of 2X,10-20X,20-50X and 1-5KHz are above the alert limits. (ii) The Training Set The training data are represented by three arrays: 1) The Pattern_Input Set: This is a 2-D array, initially holding input patterns with single fault indicators for each of the ten KPIs. Each row of this array holds a separate pattern corresponding to the ten input neurons, while each column holds respectively KPI values for OVR, SubHarm, 1X, 2X, 120Hz, 3-5X, 5-10X, 10-20X, 20-50X and 1-5KHz in that order. This set initially contains 11 patterns with the first pattern representing no fault, and the remaining ten patterns corresponding to single faults for the ten KPIs. However, more training patterns will be added dynamically to the training set as a new fault is identified during the network’s operation. These patterns will be used to provide initial training for the neural network to recognize the indicated faults. 2) The Fault_Code Set: This is a 2-D array that holds 11 different codes corresponding to the 11 input patterns in (1) above. This array grows dynamically. As new faults are recognized and further trainings performed, new fault codes can also be added into this set. 3) The Fault_Set: This is a 1-D array that holds corresponding fault types as shown on the last column of Table 1 for the 11 fault codes as described in (2) above. This array also grows dynamically, which is used to report a given fault once its corresponding fault code is identified. (iii) The Training Algorithm



As each pattern of training data from the Pattern_Input Set is presented to the neural network, an error is calculated between the actual output of the neural network and the expected output specified in the training Fault_Code Set. The Backpropagation algorithm can be modeled by a set of equations to accomplish its task as follows: 9

Yi = A(∑ x i wi )

(1)

i =0

4

NetOutj = A(∑Yi w j )

(2)

j =0

Equation (1) is used to calculate each neuron’s output (Yi) in the second layer, where xi is the input value into a neuron i and wi the weight assigned to the connection into neuron iEquation (2) is used to calculate the entire neural network’s output (NetOut) for a given pattern, where Yi is the input value from neuron i in the second layer into neuron j in the third layer, wj the weight assigned to the connection into neuron j, and A the sigmoid activation function used to scale the neural network’s output. The training of the ANN is accomplished in two phases: In the first phase, the inputs and the initial weights are propagated forward through the different layers using Equations (1) and (2) to calculate the network’s output (NetOuti) for each training pattern i. This is used to calculate the error (erri) between the expected output (ExpectedOutputi) and the actual output using Equation (3). The Initial weights for training Melvin I were randomly assigned using a random number generator function in C-sharp. erri = ExpectedOutputi − netOuti

(3)

In the second phase, the error (erri) is propagated backwards through the network in order to adjust the initial weights for the next iteration using the delta rule (Rumelhat et al., 1995) in Equation (4), where µ is the learning rate, and ∆wij the change in weights between neurons i and j. This change is added to the weight used in the previous iteration to get the new weight that will be used in the next iteration. ∆wij = 2µxi(errk)j

(4)

To get an overall network error after each iteration (epoch), the average network error (netError) is obtained by calculating the mean square root of all errors in the training set using Equation (5), where n is the number of patterns in the training set. The NetError acts as the global rate of error for the entire network and is used to advance to the next iteration if the target error level has not been reached. NetError =

1 n −1 ∑ (erri )2 n i =0

(5)



(iv) Coding Melvin I The programming language used in implementing Melvin I is C-Sharp. It is selected because of its flexibility in handling matrix mathematics in order to program the functions as defined in Equations 1 through 5. C# is also found useful in implementing the feedforward neural network architecture and the Backpropagation algorithms. Two important parameters used in the implementation of a feedforward neural network using Backpropagation algorithm are the learning rate as shown in Equation (4) and the momentum rate. The former specifies the degree to which the weights are modified in each iteration, while the latter specifies the degree to which the previous training iteration influences the current one. The learning rate and momentum rates used in the implementation of Melvin I are 0.7 and 0.9, respectively. The desired error tolerance is 1%.

4. RESULTS 4.1. Experimental Setup To evaluate the performance of Melvin I, a Pro-point 6” Bench Grinder was used to simulate a rotating machine. Five sensors were connected to different locations of the grinder. These were then connected to the TCP/IP data acquisition card (Melvin DAQ). The sampling rate (fS) was set as 12,800 samples per second and the window size (number of samples) was 4,096 samples. The Pro-point 6” Bench Grinder had a machine speed of 3,599 rpm. Using a data acquisition and signal processing software developed in our Lab, data was extracted from the raw vibration signals for each of the ten KPIs per sensor and stored in an SQL database every second.

4.2. System Evaluation There are two operation modes of Melvin I known as the “train” and “use” modes. 1) The “Train” Mode The “Train” mode is a one-time operation used during initial training of the network. Melvin I is trained according to the algorithms, training data sets, and the learning parameters as described in Section 3.2. Figure 4 is a screenshot of the training results for one of the sensors (MIH). Training begins with the specification of the server on which the acquired data are stored. Based on the selected server, databases for all machines processed by the server are automatically generated as shown on the top left window of Figure 4. The trained neural network can be duplicated for machines with similar characteristics. However, for machines with dissimilar characteristics, the training parameters must be individually adjusted to reflect the special features of the machine. For example, alarm and alert limits are machine and fault dependent. The training sets for each machine must therefore take this into consideration. Also, depending on the type of machines, the number of measurement points may vary. As can be seen on the bottom left window of Figure 4, the Edson machine had five sensors namely CIA, CIA2, MIA, MIH and MOH. The training results shown in Figure 4 are for the MIH sensor. The same faults were being monitored in all five sensors. Consequently, the training for the MIH sensor was used to recognize faults in all of the sensors. The total network training took 14 minutes on an Intel Core Duo processor of 2.53 GHz running Windows 7 operating system. The upper part of the machine state window in Figure 4



shows the last few iterations (epochs) during the training. It can be seen that the target error of 1% was attained after 21,332 iterations (epochs). The bottom part of the Machine State window in Figure 4 shows the actual and expected outputs for each of the training sets. After the initial training, the state of the trained network and the corresponding training data sets are saved for future applications and Melvin I is ready for the “Use” mode. The next section shows test results obtained during the “Use” mode for different faults and their combinations. 2) The “Use” Mode In the “Use” mode, users select from a database the machine from which to diagnose possible faults. Once a machine is selected, the user specifies which sensors to use in diagnosing faults by clicking on an appropriate sensor from the sensor information window as shown in Figure 5. The user interface of Melvin I provides a flexibility for navigating between sensors and machines modeled in the configuration database. Once in the “use” mode, the saved state of the trained network and the training sets are loaded for Melvin I to run. To enable the network to build on the existing knowledge base, the “use” mode has facilities for adding new fault information to the existing trained set. Figure 5 shows a fault reporting user interface for Melvin I. In this experiment, the server used for holding acquired data was located on the local machine (localhost). KPI data from the data acquisition card for five sensors CIA, CIA2, MIA, MIH, and MOH was processed using the signal processing/conditioning mechanisms described in Section 2. This data was stored in database tables corresponding to the sensors. The KPI data can be accessed using the tree structure on the bottom left window of the user interface. The results reported in this Figure 4. The training mode of Melvin I



paper were carried out on the Edson database. To check for faults on a given machine, the user first selects the “use” mode as shown in Figure 5. The user then selects the machine, the sensor, and the period on which to check for faults. The fault recognizer is then run against the set of parameters. If a known fault is recognized, Melvin I displays the fault type, time of occurrence and information about the fault in the machine state window. If no fault is found, a message is displayed indicating that the fault diagnosis operation completed without any fault. Figure 5 shows the diagnosis results with no faults. If an unknown fault is encountered, an interactive window is presented, which allows the user to add new information about the fault. This is then added automatically to the existing trained set for future use. To evaluate the system, Melvin I was used to test for faults between the times 2/10/2011 12:00:00 and 2/12/2011 12:00:00. As can be seen on the machine state window in Figure 5, the system reports that the diagnosis has completed and no fault is reported. This is because no known fault was found within the specified period. Such a situation occurs when the machine is in good health within the specified period. Applying the acquired data from the Bench Grinder saved in the SQL database, a number of single faults were seeded by introducing faults at the KPIs OVR, 1X and 120Hz at times 2/10/2011 14:22:07, 2/10/2011 19:36:26 and 2/10/2011 19:36:40 respectively. The tool was run with the same parameters as in Figure 5. Figure 6 shows the results of this operation. The newly introduced faults are recognized and reported at the indicated times corresponding to the seeded faults. To test for a fault or multiple faults that have not yet been presented in the trained set, combined faults were introduced at the KPIs 2X, 3-5X and 20-50X at time 2/10/2011 14:22:35. The tool was run against the same set of parameters as in Figure 6. Figure 7 shows the results after this operation. A dialog box is presented to the user to describe the new fault. In order to facili-

Figure 5. Results with no faults



Figure 6. Recognition of single faults

Figure 7. Dialog for newly encountered fault



Figure 8. Recognition of new faults after training

tate training, Melvin I also gives an indication on the severity of changes on different KPIs. This is done through the calculation of a KPI dominant factor that determines which KPI deviates most from its expected threshold. For example, for the newly introduced set of faults, Melvin I shows that the change in the amplitude corresponding to the fourth KPI (2X) is greater than for all the other KPIs. During initial training, this provides useful information that can be used by the knowledge experts to give more accurate information on faults corresponding to a given pattern. To test for the newly-acquired knowledge, Melvin I was run against the same data set as in Figure 7. Figure 8 shows results of Melvin I against the same input data at a different time with the newly added faults. It is seen that the faults that were introduced in Figure 7 are now recognized and reported without any further training. This shows that the knowledge acquired during the training in Figure 7 has been added to the knowledge base without overriding any previous training.

4.3. Discussion In Figures 5 through 8, a set of screenshots of the Melvin I user interface is demonstrated. Figure 5 shows the results when there was no fault within the given period (2/10/2011 12:00:00am to 2/12/2011 12:00:00am). In Figure 6, the simulated faults (OVR, 1X and 120Hz) are recognized and reported. Figure 7 shows the case when a new fault that is not present in the trained set is encountered. The system interacts with the user to add more information about the new fault. This new fault, together with information about it, is in turn added to the knowledge base of the system. Figure 8 shows results run on the same data set after information on the new fault was added. It is noteworthy that the system is able to report the newly identified fault without



the need for additional training. This is because the newly added information has updated the system knowledge base without overriding the initial knowledge acquired during initial training. As all the fault classes cannot be known during initial training, this option is very important as it allows the knowledge base to grow dynamically until enough knowledge has been acquired.

5. CONCLUSION Despite theoretically sound findings on developing intelligent solutions for machine conditionbased monitoring, it is hard to find commercial tools in the market that can be readily used. This paper reports research work done on the development of a component of an intelligent system (Melvin I) that remotely diagnoses an industrial machine’s health based on the observation of changes in vibration fault frequency indicators. The paper first presents an overview on the architecture adopted on the implementation of Melvin I as well as a detailed discussion on the data acquisition and signal processing mechanisms used. The development of a 3-layer neural network, which uses input from ten key performance indicators to classify faults is also reported. Melvin I is trained using an initial training set of 11 patterns. An expandable knowledge base is designed for Melvin I that is capable of accommodating new experiences during the system’s operation. As a fault not yet known to the system is detected, new information on this fault is updated to its knowledge base. With sufficient training, the neural network is able to classify faults not yet in its knowledge base. Experimental results demonstrate that Melvin I is a powerful tool for both vibration analysts and industrial machine operators. Melvin I has been tested on data remotely collected from industrial machines on site. The results so far are encouraging. Further work on Melvin I will be to improve the fault recognition success rate. Another future direction area will be to develop Melvin II, which extends the system’s functionality from fault diagnosis to fault prediction and prevention.

ACKNOWLEDGMENTS The authors would like to acknowledge the Alberta Government through Alberta Innovates Technology Futures for sponsoring this project. They would also like to acknowledge the International Institute for Cognitive Informatics and Cognitive Computing (IICICC), University of Calgary, for the industrial collaboration towards the realization of this project. They would like to thank the anonymous reviewers for their valuable comments on this work.

REFERENCES Aravindh, K. B., Saranya, G., Selvakumar, R., Swetha, S. R., Saranya, M., & Sumesh, E. P. (2010). Fault detection in induction motor using WPT and multiple SVM. International Journal of Control and Automation, 3(2). Balakrishnan, K., & Honavar, V. (1995). Evolutionary design of neural architecture – A preliminary taxonomy and guide to literature (Tech. Rep. No. 95-01). Ames, IA: Artificial Intelligence Research Group, Iowa State University. Bracewell, R. N. (1986). The Fourier transform and its applications (2nd rev. ed.). New York, NY: McGraw-Hill. Gade, S., Herlufsen, H., Konstantin-Hensen, H., & Wismer, N. J. (1995). Order tracking analysis (Tech. Review Note No. 2). Naerum, Denmark: Bruel & Kjaer.



Gerald, M., & Javadpour, K. R. (2000). An ARTMAP neural network-based machine condition monitoring system. Journal of Quality in Maintenance Engineering, 6(2), 86–105. doi:10.1108/13552510010328095 Heaton, J. (2008). Introduction to neural networks in C (2nd ed.). Chesterfield, MO: Heaton Research. Kumaraswamy, S., Rakesh, J., & Nalavade, A. K. (2002). Standardization of absolute vibration level and damage factors for machinery health monitoring. In Proceedings of the Second International Conference on Vibration Engineering and Technology of Machinery (pp. 16-18). McCormick, A. C., & Nandi, A. K. (1996). A comparison of artificial neural networks and other statistical methods for rotating machine condition classification. In Proceedings of the IEEE Colloquium on Modeling and Signal Processing for Fault Diagnosis, Leicester, UK (pp. 2/1-2/6). Ngolah, C. F., Morden, E., & Wang, Y. (2011). An intelligent fault recognizer for rotating machinery via remote characteristic vibration signal detection. In Proceedings of the 10th IEEE International Conference on Cognitive Informatics & Cognitive Computing (pp. 135-143). Peng, Y. (2004). Intelligent condition monitoring using fuzzy inductive learning. Journal of Intelligent Manufacturing, 15, 373–380. doi:10.1023/B:JIMS.0000026574.95637.36 Rumelhat, D. E., Durbin, R., Golden, R., & Chauvin, Y. (1995). Backpropagation: Theory, architectures, and applications (pp. 1–33). Mahwah, NJ: Lawrence Erlbaum. Samhouri, M., Al-Ghandoor, A., Ali, S. A., Hinti, I., & Massad, W. (2009). An intelligent machine condition monitoring system using time-based analysis: Neuro-fuzzy versus neural network. Jordan Journal of Mechanical and Industrial Engineering, 3(4), 294–305. Saxena, A., & Saad, A. (2004). Fault diagnosis in rotating mechanical systems using self-organizing maps. In Proceedings of the Conference on Artificial Neural Networks in Engineering, St. Louis, MO. Shiroishi, J., Li, Y., Liang, S., Kurfess, T., & Danyluk, S. (1997). Bearing condition diagnostics via vibration & acoustic emission measurements. Mechanical Systems and Signal Processing, 11(5), 693–705. doi:10.1006/mssp.1997.0113 Srinivasan, K. S. (2003). Fault diagnosis in rotating machines using vibration monitoring and artificial neural network (Unpublished doctoral dissertation). Indian Institute of Technology, Delhi, India. Tetsuro, M. H., & Wang, P. C. (2008). Fault diagnosis and condition surveillance for plant rotating machinery using partially-linearized neural network. Computers & Industrial Engineering, 55(4), 783–794. doi:10.1016/j.cie.2008.03.002 Umesh, K. N., & Srinivasan, K. S. (2005). Study of effects of misalignment on vibration signatures of rotating machinery. In Proceedings of the National Conference Mechanical Engineering, Mangalore, India.

Cyprian F. Ngolah received a PhD in Software Engineering from the University of Calgary, Canada in 2006, an MSc in Control Engineering from the University of Bradford, England in 1989, and a BSc in Mathematics and Computer Science from the University of Essex, England in 1988. He taught several computer science and software engineering courses at both the graduate and undergraduate levels for thirteen years at University of Buea, Cameroon. He is currently a senior software engineer in the Research and Development Department of Sentinel Trending & Diagnostics Ltd, Calgary, carrying out research on the development of a neural network for machine condition monitoring and predictive maintenance using vibration analysis. His main research interests are in real-time process algebra and its applications, tool support for formal specification languages, real-time operating systems, formal methods in software engineering and real-time systems, and artificial neural networks. Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.


Ed Morden is the President and CEO of Sentinel Trending & Diagnostics Ltd, a Calgary based company he founded in 1997. He received a BSc in Environmental Sciences in Brandon University, Brandon in 1980 and a Diploma in Mechanical Engineering Technology in Kelsey Institute, Saskatoon – 1983. He has worked as a vibration analyst for over twenty years and has extensive knowledge in predictive machine maintenance. Yingxu Wang is professor of cognitive informatics, cognitive computing, and software engineering, President of International Institute of Cognitive Informatics and Cognitive Computing (IICICC), and Director of the Cognitive Informatics and Cognitive Computing Lab at the University of Calgary. He is a Fellow of WIF, a P.Eng of Canada, a Senior Member of IEEE and ACM, and a member of ISO/ IEC JTC1 and the Canadian Advisory Committee (CAC) for ISO. He received a PhD in Software Engineering from the Nottingham Trent University, UK, and a BSc in Electrical Engineering from Shanghai Tiedao University. He has industrial experience since 1972 and has been a full professor since 1994. He was a visiting professor on sabbatical leaves in the Computing Laboratory at Oxford University in 1995, Dept. of Computer Science at Stanford University in 2008, and the Berkeley Initiative in Soft Computing (BISC) Lab at University of California, Berkeley in 2008, respectively. He is the founder and steering committee chair of the annual IEEE International Conference on Cognitive Informatics (ICCI). He is founding Editor-in-Chief of International Journal of Cognitive Informatics and Natural Intelligence (IJCINI), founding Editor-in-Chief of International Journal of Software Science and Computational Intelligence (IJSSCI), Associate Editor of IEEE Trans on System, Man, and Cybernetics (Part A), associate Editor-in-Chief of Journal of Advanced Mathematics and Applications, and Editor-in-Chief of CRC Book Series in Software Engineering. Dr. Wang is the initiator of several cutting-edge research fields or subject areas such as cognitive informatics, abstract intelligence, cognitive computing, cognitive computers, denotational mathematics (i.e., concept algebra, inference algebra, system algebra, real-time process algebra, granular algebra, and visual semantic algebra), software science (on unified mathematical models and laws of software, cognitive complexity of software, and automatic code generators, coordinative work organization theory, built-in tests (BITs), and deductive semantics of languages), the layered reference model of the brain (LRMB), the mathematical model of consciousness, and the reference model of cognitive robots. He has published over 110 peer reviewed journal papers, 220+ peer reviewed full conference papers, and 16 books in cognitive informatics, software engineering, and computational intelligence. He is the recipient of dozens international awards on academic leadership, outstanding contributions, research achievement, best papers, and teaching in the last three decades.
