Sensor Calibration Monitoring and Fault Detection Using Neural Network Based Techniques

The reduction of maintenance costs through the use of condition based maintenance practices have become a primary goal of industrial maintenance managers. Systems that can monitor the calibration of sensors can help maintenance managers meet this goal. Recently, the use of autoassociative neural networks (AANNs) to perform on-line calibration monitoring of process sensors has been shown to be not only feasible but practical. This paper investigates the input correlation requirements for an AANN based sensor monitoring system and compares the time delay neural network (TDNN) architecture with the standard multilayer perceptron (MLP) architecture. The results indicates that TDNNs may train more efficiently when robust training techniques are used but offer few other advantages unless temporal information is necessary for the prediction of sensor signals. 1.0 Introduction Industries find it difficult or impossible to detect small drifts in sensor instrumentation without taking the system off-line. These drifts can cause incorrect control actions, poor product quality, and decreased process efficiency. The current method used to detect sensor drift is to manually calibrate sensor channels on a periodic basis. These calibrations require that the instrument be taken out of service and be falsely loaded to simulate actual in-service stimuli. This can lead to damaged equipment and incorrect calibration due to adjustments made under non-service conditions. While proper adjustment is vital to maintaining proper plant operation, a less invasive technique is desirable. The application of artificial neural networks (ANNs) for plant-wide monitoring was developed at the University of Tennessee [Wrest et al, 1996]. This work has demonstrated the practicality of this application. Specially, this system will be designed to continuously monitor the condition of process sensors to aid in scheduling maintenance and allow operators to replace faulty sensor values with their best estimates. Continuous sensor calibration monitoring reduces unnecessary maintenance, increases confidence in sensed parameter values and allows for the automatically replacement of faulty sensor values with the system’s best estimate. Similar work using artificial neural networks applied in process systems has also been reported [Hines et al, 1996], [Dong & McAvoy, 1994], [Upadhyaya & Eryurek, 1991]. The autoassociative neural network (AANN), in which the outputs are trained to equal the inputs, is the most frequently used network structure. Two types of AANN architectures are studied in this paper, one is the most widely used multilayer perceptron (MLP) network, in which the input signal propagates through the network in a forward direction, the other is the time-delay neural network (TDNN) in which temporal information is used in the inputs. These two types of neural networks can both give the best estimate of the signals, but work in different situation. The applications will be discussed later in this paper. 2.0 Artificial Neural Networks for Sensor Performance Monitoring Many plant variables that have some degree of coherence with each other constitute the inputs. During training, the interrelationships among the variables are embedded in the neural network connection weights. A robust training procedure is used to force the network to rely on the information inherent in the signals