Chapter 1: Technical Background on Centrifugal Pumps Free

Recently, condition monitoring has been very vital for industrial machines due to the good results that have been produced using this technique for diagnosing machine problems. There is an increasing demand for improved machine lifetimes in industry and production organizations due to the need for manufacturing advanced products with complex requirements while adhering to tight schedules. Under these conditions, it is necessary to balance the wear of machines from prolonged use with the need for continuous production. Frequent machine faults may incur expensive maintenance costs or, if neglected, can lead to catastrophic failures, production downtime, and failure to supply. The reliance on complex machinery means that any breakdowns will affect profitability. This risk arises for many reasons, such as loss of availability, cost of spares, cost of breakdown labor, cost of secondary damage, and risk of injury to people and the environment.

As centrifugal pumps are critical process equipment, failure free operation is of vital importance to production. Moreover, the centrifugal pump is one of the most common rotating machines, and it is widely used in various industries for different processes and purposes, including pumping and shifting liquids like water and even heavy and sensitive liquids like oil and chemicals. Therefore, the detection of faults and monitoring the health condition of centrifugal pumps are considered to be important in oil and gas industries. A centrifugal pump can experience different possible faults, many of which can be categorized as hydraulic and mechanical faults. Hydraulic faults include cavitation, water hammering, and turbulence. Mechanical faults include misalignment, imbalance, damaged impeller, mechanical looseness, damaged bearing, and damaged seal. Though most of the previous works in the area of fault diagnosis of centrifugal pumps have considered a number of faults, there are possibilities for designing a better experimental rig for investigating pump faults. Therefore, the current research investigates different numbers and types of faults using a centrifugal pump experimental rig built specifically for the work.

There are four maintenance methods: breakdown, which is also known as run to failure maintenance; preventive or time based maintenance, where the maintenance of a machine is planned and scheduled; predictive or condition based maintenance, where the machines are monitored and maintenance is done wherever is necessary and required; and proactive or reliability based maintenance, which works as a root cause analyzer. Condition based maintenance is preferred due to its various methods and precision, where the proper time to do the maintenance for a machine based on the condition can be decided.

It is essential to have an effective and a powerful diagnostic system that can be used particularly for centrifugal pumps and even for any rotating machine due to the complexity of such machines and their frequent failures that can lead to very expensive maintenance. Consequently, the complexity of centrifugal pumps and their frequent failures lead to the need to consider pumps for this work. An effective diagnostic system should be able to diagnose and interpret the different conditions (faults) of a machine. It is also important to consider vibration in machines as a vital parameter that contains most of the various dynamic machine behaviors and information.

There are various vibration analysis techniques that can be used in analyzing the machine signals; they can be particularly used to extract the features from the original signals, such as

1. Time domain, which gives information about the time response using different statistical parameters, such as Root Mean Square (RMS), peak, mean, kurtosis, etc.

2. Frequency domain, which provides information about the frequency response, but faces difficulty in recognizing non-stationary signals that most rotating machines can experience (Al-Badour et al., 2010). A popular frequency method is the Fast Fourier Transform (FFT).

3. Time–frequency domain, which has the ability to provide information on both time and frequency and can overcome the drawback of frequency domain with analysis of non-stationary signals. Time–frequency domain includes different methods such as Short Time Fourier Transform (STFT), Wavelet Transform (WT), etc.

As highlighted, the time–frequency domain is considered to be more attractive in tackling the rotating machines’ fault diagnosis due to its good reputation in dealing with non-stationary signals, which most of the rotating machines' vibration signals, including those from centrifugal pumps, consist of. Therefore, the second motivation of this book is to consider time–frequency analysis methods and, in particular, the powerful technique Wavelet Transform (WT) and also to investigate three different WT methods, namely, Continuous Wavelet Transform (CWT), Discrete Wavelet Transform (DWT), and Wavelet Packet Transform (WPT). Moreover, from the need to provide a powerful diagnostic system for pumps, the pre-processing stage of the pump's signals has to be taken into account. Therefore, this book is also motivated to consider WT as a pre-processor method to be used for the feature extraction, and WPT is considered and compared with the other two WT methods (CWT and DWT).

The other main elements of the diagnostic system are the fault classification methods and procedures such as the artificial intelligence (AI) classifier types, including training algorithms, the number of input features, and optimization methods for the neural network architecture. It has been noted that the fault diagnosis of centrifugal pumps has been applied using different classifiers, such as Multilayer Feedforward Perceptron, with its traditional training algorithm “Back Propagation” (BP) and Support Vector Machine (SVM). Such classification methods have mostly shown positive performance on centrifugal pumps, but further investigations and developments are required. Therefore, the writing of this book is also motivated to investigate the application of Genetic Algorithm (GA) based optimization and selection of neural network architecture and also the training method combined with BP. There is also a need to investigate the performance of such classifiers with different numbers of inputs (features).

Generally, this book is motivated by its significant importance at various scales of industries and machines involving centrifugal pumps. There has been limited research on the application of wavelet analysis using different Wavelet Transform (WT) methods, optimization methods such as genetic algorithms, and artificial intelligence systems to centrifugal pumps’ fault diagnostics. The main motivation of this book is to develop novel processing methods by applying vibration techniques in fault detection and to extract input features for artificial intelligence classifiers for the purpose of automatic diagnosis of these centrifugal pumps. Throughout the literature review and personal viewing of applied condition monitoring based techniques in industries, conventional techniques are still applied, which results in increasing errors made by human interpretation. The proposed approaches in this book would provide a solution in the form of automatic fault diagnosis that aims to reduce diagnosis errors and increase accuracy.

According to Oman vision 2020, Oman is moving toward boosting its industrial development and growth (Strolla and Peri, 2016), and one of its key drivers for this growth is increasing the number of machines including centrifugal pumps; this energizes the need for reliable fault classification of centrifugal pumps in Oman. Thus, the research and investigations discussed in this book can be further investigated and tested locally with centrifugal pumps to evaluate the diagnostic performance in real time. Moreover, it is very important to prevent centrifugal pump failures and understand their modes of failure by carefully studying and monitoring the conditions and failure causes so as to make centrifugal pumps highly reliable (Sakthivel et al., 2012).

There is an essential need to find an effective technique for the fault diagnosis of centrifugal pumps as there are limitations in the available diagnosis methods. Centrifugal pumps can experience many different types of faults, which without an early diagnosis may lead to catastrophic failures. There are various methods of vibration analysis, such as time domain analysis (Al-Tubi and Al-Raheem, 2010) and frequency domain analysis, where the Fast Fourier Transform (FFT) method is applied (Mehala and Dahiya, 2008; and Al-Tubi et al., 2012). More recently, a powerful multi-resolution technique called wavelet analysis has been applied to fault detection in rotating machines and has demonstrated the ability to analyze non-stationary signals (Peng and Chu, 2004; Mehala and Dahiya, 2008; Prakash et al., 2014; and Al-Tubi and Al-Raheem, 2015). Artificial intelligence (AI) systems for automatic fault diagnosis and classification have been investigated by numerous researchers (Wang and Chen, 2007; Rajakarunakaran et al., 2008; Farokhzad, 2013; Muralidharan et al., 2014; and Sakthivel et al., 2014). Such automatic fault classification can be used to improve precision and reduce mistakes that might otherwise be caused by human misinterpretation.

Throughout this book, some current research gaps were filled and answers were provided for the following questions:

1. What are the existing techniques/algorithms used to ensure better centrifugal pump fault diagnosis through an artificial intelligence system?

2. What improved techniques/algorithms can be used to ensure better centrifugal pump fault diagnosis through an artificial intelligence system?

3. What failure modes do centrifugal pumps exhibit?

4. What monitoring and diagnostic techniques are applied to similar vibrating rotating machinery?

5. What are the existing techniques/algorithms used to ensure better feature extraction?

6. What improved (new) techniques/algorithms can be used to ensure better feature extraction?

In this book, the following objectives are considered:

1. To identify the main faults of centrifugal pumps in a real industrial environment and use the Fast Fourier Transform (FFT) to verify and confirm different pump conditions.

2. To apply wavelet transform methods, such as CWT, DWT, and WPT, for signal processing and pre-processing and extract various parameters, including input vector features to be forwarded to the Artificial Neural Network (ANN) classifier.

3. To apply and investigate the artificial intelligence classifier of multilayer perceptron (MLP)-ANN using a hybrid training method, which is combined with GA and BP algorithms and used as an optimizing algorithm.

4. To apply and investigate the artificial intelligence classifier of Support Vector Machine (SVM) along with a proper kernel function for the centrifugal pump fault classification.

5. To investigate the impact of wavelet mother function selection on AI classifiers' performance.

This book provides in-depth investigation based on case studies and experiments. The overall methodology is divided into three stages, namely, data collection, data pre-processing, and condition classification and diagnosis. These stages are applied in order, such that each stage serves as the next one. The three stages are briefly described in Secs. 1.4.1–1.4.3, respectively. Figure 1.1 illustrates the work methodology through the three stages.

The purpose of this stage is to collect the vibration data from the centrifugal pump rig. The vibration signals can represent two conditions, namely, healthy conditions and faulty conditions. First, the signal of the normal condition is acquired when the pump is in a healthy condition without any faults. Second, the faulty conditions are divided into two main categories: five mechanical faults (bearing, misalignment, imbalance, impeller, looseness) and one hydraulic fault (cavitation). These faults are created and simulated one by one. The technique of acquiring the signals from the pump uses an accelerometer, which is mounted on the centrifugal pump and specifically on the bearing housing. This sensor transfers the vibration data to the Data Acquisition Device (DAQ) where the signals are amplified and noise filtered out, and then they are transmitted to a computer that is provided with a digital/analog converter card in order to convert the analog signals to digital signals. Finally, these signals are captured via LabVIEW, saving the raw signals to be used later in the second stage for further processing.

The purpose of this stage is to process the acquired vibration signals using MATLAB software. This stage is a pre-processing implementation for the signals using three different wavelet transform methods, namely, CWT, DWT, and WPT, and it extracts a number of features. The three wavelet transform methods are tested and investigated individually. CWT is tested with the Morlet wavelet mother function, and two different wavelet functions are used and tested with DWT and WPT, namely, db4 and rbio1.5. A number of features are extracted from each condition. The genetic algorithm is applied for the first time for optimal performance (for a centrifugal pump fault diagnosis) with a WT neural network hidden layer and neuron optimization and for the neural network training along with its traditional algorithm (Back Propagation). MATLAB software is used to implement the pre-processing including feature extraction for the acquired signals; the code is provided in Appendix B.

The purpose of this stage is to classify and diagnose the conditions of centrifugal pumps using two artificial intelligence classifiers, namely, MLP and SVM. MLP is implemented along with its traditional learning algorithm (Back Propagation) and with GA for rotating machinery to investigate and compare the two learning algorithms. SVM is also used as a classifier and its performance is compared with MLP. MATLAB software is used to implement the classification stage; the code is provided in Appendix B. Classifiers consist of two main processes: training of the data and testing, where an automatic classification is implemented for different conditions. The performance of the AI classifier is measured and evaluated according to the classification accuracy rates (%) by calculating Mean Square Error (MSE).

GA and BP algorithms are tested and investigated for the training of MLP where its weights have to be modified and updated. It is remarked that this hybrid training method is applied with ANN since reports on this are limited with any fault diagnosis research on centrifugal pumps. Generally, the principal operations of MLP-ANN with learning algorithms of GA and BP are depicted in Fig. 1.2.

MLP is trained using GA and BP where the following steps are applied during the training process with GA:

1. [W1, W2, Wn] are bias weights of the neural network, and these weights are coded into individual populations (chromosomes).

2. Training and evaluating each chromosome (previously weighted) individually based on the fitness.

3. Ranking the chromosomes according to fitness.

4. Selecting the best and top ranked chromosomes to survive.

5. Using genetic operators of crossover and mutation to recreate a new generation of chromosomes where the weak chromosomes at step 1 are replaced with new and strong ones.

6. Repeating step 2 by training and computing the new chromosomes, and continuing the process until reaching the desired Mean Square Error (MSE).

BP algorithm is then used after the optimization of GA. BP is based on the Levenberg–Marquardt (LM) method that works to modify the neural network weights, which have been optimized through GA.

This book proposes a novel technique that works as an automatic fault diagnosis for the centrifugal pump based on the combination of artificial neural network systems, wavelet transforms, and genetic algorithms. Wavelet transform methods are used as pre-processors where the input features of the artificial intelligence classifiers are extracted. A genetic algorithm (GA) is applied as an optimizer and trainer for the artificial neural network classifier, where the learning algorithms, including the conventional BP as well as GA, are tested and investigated for the fault classification. This book considers different types of centrifugal pump faults, including both mechanical faults and a hydraulic fault.

The following approaches are applied for novelty considerations:

1. A new rig has been designed, built, tested, and used to investigate mechanical and hydraulic faults in a centrifugal pump.

2. WPT has been applied for feature extraction from centrifugal pump data using two mother wavelet functions: db4 and rbio1.5, and the results are compared with CWT and DWT.

3. Signal decompositions (approximations and details) have been investigated based on features extracted from centrifugal pump data using DWT and WPT.

4. GA based selection of the hidden layer and neuron has been applied for the neural network of the centrifugal pump condition classification.

5. A hybrid training method combining GA and BP algorithms has been applied for the condition classification of a centrifugal pump.

This chapter emphasized the aim, objectives, and the main methodology that this book would employ. There are many novel contributions that have been identified from the proposed work that attempt to provide a new approach on the integration between the mechanical part that is represented by the condition monitoring of the centrifugal pumps and artificial intelligence. This approach also presents a unique emphasis on the possibility of various applications of advanced methods like artificial intelligence and genetic algorithm. The next chapters of this book will provide further details on such applications.