Automated methods for signal validation and anomaly detection

Instrument fault detection and process diagnostics are necessary to verify the validity of process variables channeled to control systems, protection systems and plant monitoring systems. Three techniques are applied for validating sensor signals and classification of spectral signature functions. The characterization of a data cluster representing a certain process behavior is achieved by steady-state nonlinear modeling of one or more critical signals as a function of one or more process variables in the system. This prediction model is used to detect sensor maloperation or process anomaly by comparing the measurement and prediction of the same variable. Bidirectional associative memory is applied to classification of frequency domain spectral sigantures. Its noise tolerant properties make it suitable for classification of power spectral density (PSD) functions which are difficult to classify due to wide variations possible in spectra derived from the same process. Backpropagation Networks (BPN) are applied to predictive modeling and classification of PSD functions. The BPN allows arbitrary nonlinear mapping of data to data and performs well as a noise tolerant classifier and as a predictive model.