Sensor Validation and Fault Detection Using Neural Networks

In the power generation industry, periodic sensor calibrations are required to assure sensors are operating correctly. Out-of-calibration sensors can lead to mis-control of electric power output, possibly causing a loss in revenue or damage to equipment. Continuous sensor calibration monitoring is desirable to assure correct power generation and reduce maintenance costs associated with performing unnecessary manual sensor calibrations. An artificial neural network-based sensor calibration monitoring system can provide continuous sensor status information and virtual estimates for faulty sensors. This paper describes the design of an autoassociative neural network (AANN) based sensor validation and fault detection system for a power generation system. INTRODUCTION It is difficult to detect small drifts in sensor instrumentation in complex control systems such as nuclear power plants, fossil fired power plants, and chemical or petroleum processing facilities. These sensor drifts can cause incorrect control actions and decrease system efficiency. Currently, sensors are manually calibrated on a periodic basis assure proper operation. This method requires all sensors to be calibrated, regardless of their condition, leading to possible equipment damage and incorrect calibration due to adjustments made under non-service conditions. With economic factors becoming more important, condition-based maintenance strategies are desired. These strategies require that instruments be manually recalibrated only when their performance is degraded beyond a specific tolerance. Continuous verification of the instrument’s calibration will reduce unnecessary sensor calibrations and provide higher fidelity sensor measurements. Elimination of unnecessary maintenance results in reduced equipment damage and increased plant efficiency. A well-trained neural network is capable of continuously providing the best unfaulted estimate of the measurement, and when the sensor becomes faulty, it acts as a virtual sensor replacing the faulty sensor measurement with the best estimate. Therefore, the utilization of neural networks for sensor calibration monitoring will increase plant reliability, aid in scheduling maintenance, and reduce unnecessary maintenance. FOSSIL POWER SYSTEM Figure 1 is a block diagram of a fossil power generation system. Hundreds of sensors, which measure flows, temperatures, pressures and levels, must operate properly to ensure efficient and safe operation. Any anomalies in the sensors need to be detected. A neural network based sensor validation and fault detection system can fulfill this need. In this paper, a group of 21 signals from a fossil power plant is monitored. Figure 1. Fossil power generation system SIGNAL ESTIMATION AND FAULT DETECTION The use of AANNs for plant wide monitoring has been widely reported in the nuclear industry [B. R. Upadhyaya and E. Eryurek, 1992, R. E. Uhrig, et al, 1996, Nabeshima, et al, 1995]. Similar work using ANNs applied to chemical process systems have also been reported [Dong and McAvoy, 1994, Kramer, 1992]. In this study, an autoassociative neural network with these 21 signals as input is constructed for sensor validation and fault detection purposes. Sensor Monitoring In an autoassociative neural network, the outputs are trained to emulate the inputs over an appropriate dynamic range. 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 redundant information from correlated sensors to estimate that specific sensor’s value. As a result, any specific network output shows virtually no change when the corresponding input has been distorted by noise, faulty data, or missing data. This characteristic allows the AANN to detect sensor drifts or failures by comparing sensor measurements (network inputs) with the corresponding network estimates of the sensor values (network outputs). Because of the correlated information input to the network, ill-conditioning problems will arise. This problem can cause poor network performance and non-unique solutions. Under such circumstances, the estimation of the signals from the network may be unreliable. To deal with these ill-conditioning problems, a network regularization method is applied during network training. The regularized training routine makes the problem better conditioned and tends to stabilize network performance. Figure 2 shows a sensor monitoring module for a group of two sensors whose measurements are correlated to some degree. When a sensor's signal to the autoassociative network is faulty due to a drift failure, the network still gives a valid estimate of the sensor value due to its use of information from other correlated sensors. The difference or residual ( n r ) between the sensor estimate ( n x̂ ) and the actual measurement ( n x ) normally has a zero mean and a variance related to the amount of noise in the sensor's signal. When a sensor is faulty, the mean or variance of its associated residual deviates from normal values and the deviations or drifts can be detected by a fault detection module which is detailed in the next section. A A N N M o d e l 1 x 1̂ x 2 x 2̂ x Σ Σ + + 1 r 2 r Faul t Detec t ion Faul t H ypo thes i s Figure 2. Sensor monitoring module Fault Detection The fault detection module implements a statistical decision algorithm: the sequential probability ratio test (SPRT) which was initially developed by Wald [1945], later used by Upadhyaya [1987, 1989], and more recently by Gross and Singer [Gross 1991, and several others]. This module processes the residuals n r and outputs the condition of the sensor. An output value of zero indicates the sensor is normal, while a value of one indicates the sensor is faulty. Rather than computing a new mean and variance at each sampling time, the SPRT continuously monitors the sensor's performance by processing the residuals at each sampling time to compute a likelihood of failure. This SPRT-based method is optimal in the sense that a minimum number of samples are required to detect a fault existing in the sensor or instrument chain. Faulty Sensor Correction The fault detection module continues monitoring a sensor output even after it has been determined to be faulty. While the sensor is faulty, the best estimate of the sensor value given by the neural network can be used for automatic or manual control, for display to plant operators, or for other critical tasks. Therefore, the plant could continue operating without interruption even with drifting or failed sensors. This estimate also replaces the faulty sensor measurement for input to the network so that the monitoring of other sensors is not degraded. The sensor output is substituted back into the network when the fault has been corrected. This method always gives the operator or control system access to the best estimate of the signal whether it is the unfaulted measured value or the estimated value. This module operates as a virtual sensor that provides the best estimate of the measurement, avoiding possible wrong decisions made by using faulty sensor measurements. System Adaptation The neural network learns the interrelationships among the sensor measurements during training. Although the training set should include samples from all plant operating regions, sometimes the operating state may change to one that was not included in the training set. This can be caused by component wear, cyclical changes, or changes in the plant configuration, among others. These changes would be characterized by several residuals deviating significantly from their normal mean of zero. When this happens, the output of the AANN is not reliable and the network must be retrained to operate under the new conditions. If only one residual deviates from zero, a sensor fault is hypothesized. Although a multi-layer network with at least one non-linear hidden layer is able to model any linear or non-linear processes [Cybenko 1989, Hornick 1989, Funahashi 1989], a three hidden layer neural network (sometimes referred to as a five layer network) has several advantages as an autoassociative neural network for sensor monitoring [Kramer 1992]. The hidden layers act as feature extractors and the linear output layer combines these features to provide a desired mapping. If the features do not change when a plant or process operating condition changes, only the output layer weights need to be adjusted to perform the desired mapping. This assumption holds for fairly small changes in operating conditions. Retraining the entire network may be necessary for major changes in plant operating conditions when adjusting the output weights does not result in satisfactory performance. Because of the linear output layer, adjustments of the output layer weights can be achieved by using a least-squares regression procedure. Several methods of solving for the linear output weights exist including pseudo-inverse methods and the LU or QR decompositions. In this study, it was found that a better method uses a truncated singular value decomposition (TSVD) technique that uses the most relevant information to compute the weight and discards unimportant information due to noise to avoid network overfitting. This help alleviate regression instabilities caused by collinearity. RESULTS This section presents the results of this sensor validation system performance when sensors are corrupted with artificially induced drifts in individual signals. The signal monitoring and fault detection modules work together to determine the sensor status, while the faulty sensor correction module provides correct inputs to the network, and the system adaptation module adapts the network to new operating conditions. Drift Detection A drift error is defined as a slow rate of change i