In a complex control process, instrument calibration is periodically performed to maintain the instruments within the calibration range, which assures proper control and minimizes down time. Instruments are usually calibrated under out-of-service conditions using manual calibration methods, which may cause incorrect calibration or equipment damage. Continuous in-service calibration monitoring of sensors and instruments will reduce unnecessary instrument calibrations, give operators more confidence in instrument measurements, increase plant efficiency or product quality, and minimize the possibility of equipment damage during unnecessary manual calibrations.;In this dissertation, an artificial neural network (ANN)-based instrument calibration verification system is designed to achieve the on-line monitoring and verification goal for scheduling maintenance. Since an ANN is a data-driven model, it can learn the relationships among signals without prior knowledge of the physical model or process, which is usually difficult to establish for the complex non-linear systems. Furthermore, the ANNs provide a noise-reduced estimate of the signal measurement. More importantly, since a neural network learns the relationships among signals, it can give an unfaulted estimate of a faulty signal based on information provided by other unfaulted signals; that is, provide a correct estimate of a faulty signal. This ANN-based instrument verification system is capable of detecting small degradations or drifts occurring in instrumentation, and preclude false control actions or system damage caused by instrument degradation.;In this dissertation, an automated scheme of neural network construction is developed. Previously, the neural network structure design required extensive knowledge of neural networks. An automated design methodology was developed so that a network structure can be created without expert interaction. This validation system was designed to monitor process sensors plant-wide. Due to the large number of sensors to be monitored and the limited computational capability of an artificial neural network model, a variable grouping process was developed for dividing the sensor variables into small correlated groups which the neural networks can handle. A modification of a statistical method, called Beta method, as well as a principal component analysis (PCA)-based method of estimating the number of neural network hidden nodes was developed. Another development in this dissertation is the sensor fault detection method. The commonly used Sequential Probability Ratio Test (SPRT) continuously measures the likelihood ratio to statistically determine if there is any significant calibration change. This method requires normally distributed signals for correct operation. In practice, the signals deviate from the normal distribution causing problems for the SPRT. A modified SPRT (MSPRT) was developed to suppress the possible, intermittent alarms initiated by spurious spikes in network prediction errors.;These methods were applied to data from the Tennessee Valley Authority (TVA) fossil power plant Unit 9 for testing. The results show that the average detectable drift level is about 2.5% for instruments in the boiler system and about 1% in the turbine system of the Unit 9 system. Approximately 74% of the process instruments can be monitored using the methodologies developed in this dissertation.
展开▼
