Development of a classification scheme for fault detection in longwall systems
Daniel R. Bongers · The University of Queensland · 2004
This thesis addresses the development of a classification scheme for training data, for the purpose of generating a data-driven fault detection and isolation function. Such a scheme is necessary when mathematical models of a dynamic system cannot be determined, and the grouping of sensor data into discrete classes is non-trivial. The development of data-driven systems for classification relies on collected data with known class information, ie. for each observation (either a single piece of information or an observation vector of information), there is an associated class. It is the set of observations and known classes that is used to develop 'rules' (statistical or otherwise) to classify new observations of unknown class. In the case of fault detection and isolation (FDI), the class of an observation may be the state of the system being observed, which if known, can indicate the presence of a system fault. If a sufficient quantity of system data is collected, with known state (or class), a classifier may be developed that can determine the state of the system from and unclassified observation, and hence detect and isolate faults in the system. The success of a classifier developed from such training data is almost solely dependent on the correct classification of each observation in the training set. Although the level of signal noise will affect the classifier performance, correctly classed observations are the key. All literature regarding the development of classifiers from example data focus on a new or improved method of using the information in the data to develop classification rules. To this point in time, no author has focussed on producing training data in a non-trivial situation. The training data is either computer generated from a mathematical model, or taken from a system where the classes of observation are simple and/or distinct. The problem with the development of an FDI system for a longwall from example data is twofold: 1. The maintenance log is a record of what happened on the longwall, when it happened, and how long production was delayed. The documented time is quoted from shift personnel, and is rarely congruent with the computer time associated with the condition monitoring data. As such, it is necessary to determine the 'computer time' of each recorded fault, using the available condition monitoring data and information stored in the maintenance logs. 2. Since the events recorded are those that have caused the longwall to cease production, there is no information as how long the fault was present in the system before catastrophic failure. Also, there is limited knowledge of how the fault developed in the longwall, and the nature of the transition from normal operation, to operation with the fault present. As such, even with the ability to determine the computer time of each fault recorded, we have no way of classifying the observations immediately before longwall shutdown. These problems, that have not been dealt with in the past, are ones that must be overcome if a fault detection scheme is to be developed for a longwall, short of creating a detailed mathematical model of the longwall system and making vast assumptions about the distribution of the recorded data. In addition to this, the development of a method to determine the classes of observation vectors from maintenance records may be useful in the development of FDI systems in other applications, and could be used to determine more accurate maintenance records and statistics. Validation of this classification scheme is in the form of successful fault detection and isolation of major longwall faults. Linear discriminants and artificial neural networks were generated using the developed training dataset, which demonstrated the timely detection and characterization of faults in more than 85% of instances. Improvements in FDI performance were achieved by modifying the performance function of the back-, propagation algorithm used to train the various networks. In addition, a rigorous degradation study of the classification scheme showed it to be self-consistent.