Sensor Fusion for Machine Condition Monitoring

Machinery maintenance accounts for a large proportion of plant operating costs. Compared with the conventional scheduled maintenance strategy which is to stop the machine at pre-determined intervals, modern condition-based maintenance strategy stops the machine only before there is evidence of impending failure. With the development of cheaper sensors, more and more sensors are designed for machine condition monitoring. It is now possible to use multi-modal sensor input to monitor machine condition in a collaborative and distributed manner. In this paper, three categories of methods for condition monitoring are reviewed – 1. knowledge based, 2. model based 3. data based methods. Knowledge-based systems are derivations from expert systems that use rules and inference engines to determine failures and their causes. Data-driven methods use machine fault data, typically derived during experiments to train a monitoring system. Pattern recognition algorithms then attempt to classify actual sensor data using the results of the training phase. However it is often impractical to obtain data for every type of fault. Model-based techniques on the other hand use mathematical models to predict machine performance. We propose to combine the model based and data based method for machine condition monitoring. The data is used to train the model and derive its parameters. The various fault modes are then identified and simulated. The output is then input to the classification schemes that can be then used to identify and classify real-time data. We apply the technique to condition monitoring of electrical motors. INTRODUCTION Machinery maintenance accounts for a large proportion of plant operating costs. It has been clearly demonstrated that the use of appropriate condition monitoring and maintenance management techniques can give industries significant improvements in efficiency and directly enhance profitability [1]. Condition monitoring or CBM (Condition Based Maintenance) is an effective form of predictive maintenance (PdM) where, as the name stated, people monitor the condition of specific areas of plant and equipment. CBM involves the observation of a system over time using periodically sampled dynamic response measurements from an array of sensors, the extraction of fault-sensitive features from these measurements, and the statistical analysis of these features to determine the current state of the system [2]. It is also referred to fault detection, fault isolation and identification. The use of CBM allows maintenance to be scheduled, or other actions to be taken to avoid the consequence of failure, before the failure occurs. It is typically much more cost effective than allowing the machinery to fail [1]. Over the past sixty years major improvements have occurred in the technology, practice and systems used for equipment condition measurement. Since 1939, vibration measurements have been used to judge the condition of machinery [3]. However, wireless technology [4] has only recently been developed and deployed for vibration-based condition monitoring. Besides sensor and signal processing technology, there have been significant developments in the architectures and methodologies to perform condition monitoring. Current existing techniques can be classified into three categories: knowledge based, model based, and data based methods. These methods are reviewed in the next section, followed by a review of popular frameworks of sensor fusion. A new combined method is proposed for machine condition monitoring. REVIEW OF CURRENT METHODS Knowledge based methods Knowledge based methods mainly perform automated reasoning to carry out situation assessment, consequence prediction and analysis [5]. One of the first machinery expert systems (Amethyst) was introduced by IRD in the mid 1980s. Also in the mid 1980s, Predict DLI developed an expert system for use with the US Navy aircraft carrier Condition Based Maintenance program for which they provided data and analysis [3]. This knowledge-based method is still popular for making expertise available to decision makers and technicians who need answers quickly. The most common form of expert systems is a program made up of a set of rules that analyze information (usually supplied by the user of the system) about a specific class of problems and recommend a course of action to implement corrections [6]. A typical representation of knowledge would be presented by rules that have the form: If (evidence exists for X) then do Y, where Y may involve performing a computation or updating a database. The rules, however, need to programmed into the system based upon opinions of human experts and can therefore be prone to subjectivity. Model based methods The basic principle of a model-based fault detection scheme is to generate residuals that are defined as the differences between the measured and the predicted variables. Ideally, these residuals are only affected by system faults and are not affected by any changes in the operating conditions, such as power quality changes or load variations [7, 8]. So the key point of this method is to find a variable which can be well modeled and measured to generate the residual signal which is not always available in some systems. Data based methods Many modern approaches to fault diagnosis and prognosis are based on the idea of pattern recognition (PR). In the broadest sense, a PR algorithm is simply one that assigns to a sample of measured data a class label, usually from a finite set. The appropriate class labels would encode damage type, location etc. In order to carry out the higher levels of identification using PR, it will almost certainly be necessary to construct examples of data corresponding to each class [9]. Classical Bayesian classifier, Linear Discriminant Analysis (LDA), Artificial Neural Network (ANN) and Support Vector Machine (SVM) are the major algorithms for pattern classification [10]. A feature is some characteristic of measurements (such as averages, dominant frequencies etc.) that provides information to discriminate between various classes of input data. In fact, an ideal feature extraction can make the job of classifier trivial If the extracted features are good then simple methods would do a good job in classification. Good features are however usually domain dependent. Different applications will have different features of interest. SENSOR FUSION ARCHITECTURE A complex machine consists of many components which could be the potential fault sources. When a single sensor is not able to identify all the faults in a machine, multiple sensors are needed to fulfill this task. Multi-sensor based condition monitoring system collects data from different sensors. Data fusion is necessary to make good use of all the sensors’ data. Bedworth and O’Brien [11] described a popular framework called the Omnibus model. Figure 1 gives the general layout of this framework, which consists of four main modules and can be executed along the clockwise loop. These modules are used to address the various tasks in sensor fusion and its functional objectives. Fig. 1 Omnibus data fusion model In pattern recognition module, there are two kinds of learning method which are supervised learning and unsupervised learning. Supervised learning is a type of learning algorithm in which the diagnostic is trained by showing it the desired label for each data set [10]. Unsupervised learning doesn’t require the labeled faults data, but it can only be used for detection which is called anomaly detection methods. If supervised learning is required, there will be serious demands associated with it; data from every conceivable damage situation should be available. The two possible sources of such data are computation or modeling, and experiment. In order to accumulate enough training data, it would be necessary to make copies of the system of interest and damage it in all the ways that might occur naturally; in reality, this is simply a waste of money. Modeling requires understanding of physics of the system. The more physics we know about the system, the less faults data we need for the training process. PROPOSED METHOD As the purpose of condition monitoring, not only the diagnosis of the machine faults but also some prognosis of the machine life are expected. Although the prognosis task of life time prediction is difficult at present due to the lack of the understanding of complex system and the lack of access to the faults data, the method used now should have the potential to achieve the prognosis task in the future. Supervised learning methods have the potential to predict the time to failure. The difficulty of this kind of methods is always the access to the faults data. Modeling of the machine measurable signal under normal condition and faults condition is vital to achieve this method. With the development of the physics which describes the complexity of the machine system, more and more measurable signal can be modeled mathematically. Figure 2 shows an ideal system diagram for machine online condition monitoring which uses the measurement of the system operation under normal condition to first identify the system parameters. The model with its parameters thus identified is used to simulate the system under faulty conditions. The simulated data can then be used to extract features and train classifiers. When the system is deployed and operational, sensor data from the machinery is used to extract features that are then classified by the classifiers obtained from the simulated results. The concept is applied to fault monitoring of 3-phase induction motors. Motor Current Signature Analysis (MCSA) is a commonly used technique for fault monitoring of large induction motors. It is a noninvasive, on-line monitoring technique for the diagnosis of problems in large induction motors. Specific harmonic components are located to detect different faults such as broken rotor bars, shorted turns in low voltage stator windings, and air gap eccentricity [12]. However with multiple faults or different varieties of drive schemes, MCSA can become an onerous task as different types of faults and time harmonics may end up generating similar signatures. Thus other signals such as speed, torque, noise, vibration etc., are also explored for their frequency contents [13, 14]. The most popular PR method for induction motor faults is Artificial Neural Network (ANN) algorithm [15-17]. Neural network algorithms however do not employ any physics of the process and do not contribute to a deeper understanding of the cause of failures. A lot of mathematical models of simulation for rotor bar broken faults and experiments are also available [15-18]. However, few people deal with multiple classes of motor faults under different load conditions using model based methods. The proposed method uses mathematical models of induction motor developed in previous research [19-23] and pattern classification techniques. Multi-Sensor Measurement Machine Normal Condition Simulation Fault Condition Modeling & Calculating Parameters Feature extraction & training Online Monitoring Faults modeling under different load conditions Fig. 2 Multi-Sensor Condition Monitoring Diagram