Güç Sistemi Gözetiminde Yapay Nöral Devre İle Stokastik İşaret İşleme

Abstract

The problem of detecting changes in system dynamics has been receiving growing attention in the last decade. Although the reason for this can vary, the case being dependent on the nature of application of concern, in particular, in a power plant the case is closely related to the plant 's operational safety. This is simply because the operational status is determined by the system dynamics and any non-stationary behavior has to be assessed correctly in order to be able to take appropriate action in time. in a steady-state course of a process the process variables fluctuate around their nominal values. This fluctuations are termed as noise and they contain substantial information about the system dynamics. As the noise has its individual statistical characteristics, any change in the system dynamics is reflected in the noise and this additional contribution to noise is relatively easily identified by means of advanced signal processing methods exploiting the modifıcation of the statistical properties. For this, a rather common approach is the signal modelling which concerns the process which is an output of a finite-order discrete time linear system driven by white noise. Such processes are called autoregressive (AR) as a general name and they can be represented in the form of a linear predictor which is actually a finite impulse response (FIR) fılter acting on the data. If a signal which is obtained from a random process is modelled by a time series, the error betvveen real and modelled signals can be presented by ep(n)=yn-y'n. Here, y' is the signal modelled by a time series and given by P y'n = -£«,y«-i 1=1 so that p ep(n) = yn+^aiyn_i i=l in this formulation, the ep is white noise processes. By arranging the last equation we öbtain an AR process of the form ix l The orthogonality of the backward residuals provides a fast convergence required in the training with the stochastic patterns. In the case of failure reflected in the signal subjected to analysis, a stressed outcome at the network 's output is produced as that information is new to the network and therefore is not consistent with the stored statistical information meant for duly estimation in normal operation. The above description of the signal modelling realized by coupling a lattice structure with a neural network indicate the enhanced change detection in system dynamics due to the nonlinearity of the neural network. Namely, the linear signal modelling is forced to be modified to non-linear signal modelling where the correlation between the linear and the non-linear models is greatly reduced by an appropriate choice of the algorithmic model of the neural network. Hence the signal- to-noise ratio where signal is the reflection of the change of system dynamics into the correlation and noise is the minimized correlation as obtained using neural network, is greatly improved as requried particularly in early fault detection and diagnosis applications. The utilization of feedforward neural network with noise signals for failure detection in process systems is demonstrated and verifications with the actual plant data with simulated failures on, are presented. xiv SUMMARY STOCHASTIC SIGNAL PROCESSING WITH NEURAL NETVVORK İN POWER PLANT MONITORING The problem of detecting changes in system dynamics has been receiving growing attention in the last decade. Although the reason for this can vary, the case being dependent on the nature of application of concern, in particular, in a power plant the case is closely related to the plant 's operational safety. This is simply because the operational status is determined by the system dynamics and any non-stationary behavior has to be assessed correctly in order to be able to take appropriate action in time. in a steady-state course of a process the process variables fluctuate around their nominal values. This fluctuations are termed as noise and they contain substantial information about the system dynamics. As the noise has its individual statistical characteristics, any change in the system dynamics is reflected in the noise and this additional contribution to noise is relatively easily identified by means of advanced signal processing methods exploiting the modifıcation of the statistical properties. For this, a rather common approach is the signal modelling which concerns the process which is an output of a finite-order discrete time linear system driven by white noise. Such processes are called autoregressive (AR) as a general name and they can be represented in the form of a linear predictor which is actually a finite impulse response (FIR) fılter acting on the data. If a signal which is obtained from a random process is modelled by a time series, the error betvveen real and modelled signals can be presented by ep(n)=yn-y'n. Here, y' is the signal modelled by a time series and given by P y'n = -£«,y«-i 1=1 so that p ep(n) = yn+^aiyn_i i=l in this formulation, the ep is white noise processes. By arranging the last equation we öbtain an AR process of the form ix l The orthogonality of the backward residuals provides a fast convergence required in the training with the stochastic patterns. In the case of failure reflected in the signal subjected to analysis, a stressed outcome at the network 's output is produced as that information is new to the network and therefore is not consistent with the stored statistical information meant for duly estimation in normal operation. The above description of the signal modelling realized by coupling a lattice structure with a neural network indicate the enhanced change detection in system dynamics due to the nonlinearity of the neural network. Namely, the linear signal modelling is forced to be modified to non-linear signal modelling where the correlation between the linear and the non-linear models is greatly reduced by an appropriate choice of the algorithmic model of the neural network. Hence the signal- to-noise ratio where signal is the reflection of the change of system dynamics into the correlation and noise is the minimized correlation as obtained using neural network, is greatly improved as requried particularly in early fault detection and diagnosis applications. The utilization of feedforward neural network with noise signals for failure detection in process systems is demonstrated and verifications with the actual plant data with simulated failures on, are presented. xiv SUMMARY STOCHASTIC SIGNAL PROCESSING WITH NEURAL NETVVORK İN POWER PLANT MONITORING The problem of detecting changes in system dynamics has been receiving growing attention in the last decade. Although the reason for this can vary, the case being dependent on the nature of application of concern, in particular, in a power plant the case is closely related to the plant 's operational safety. This is simply because the operational status is determined by the system dynamics and any non-stationary behavior has to be assessed correctly in order to be able to take appropriate action in time. in a steady-state course of a process the process variables fluctuate around their nominal values. This fluctuations are termed as noise and they contain substantial information about the system dynamics. As the noise has its individual statistical characteristics, any change in the system dynamics is reflected in the noise and this additional contribution to noise is relatively easily identified by means of advanced signal processing methods exploiting the modifıcation of the statistical properties. For this, a rather common approach is the signal modelling which concerns the process which is an output of a finite-order discrete time linear system driven by white noise. Such processes are called autoregressive (AR) as a general name and they can be represented in the form of a linear predictor which is actually a finite impulse response (FIR) fılter acting on the data. If a signal which is obtained from a random process is modelled by a time series, the error betvveen real and modelled signals can be presented by ep(n)=yn-y'n. Here, y' is the signal modelled by a time series and given by P y'n = -£«,y«-i 1=1 so that p ep(n) = yn+^aiyn_i i=l in this formulation, the ep is white noise processes. By arranging the last equation we öbtain an AR process of the form ix l The orthogonality of the backward residuals provides a fast convergence required in the training with the stochastic patterns. In the case of failure reflected in the signal subjected to analysis, a stressed outcome at the network 's output is produced as that information is new to the network and therefore is not consistent with the stored statistical information meant for duly estimation in normal operation. The above description of the signal modelling realized by coupling a lattice structure with a neural network indicate the enhanced change detection in system dynamics due to the nonlinearity of the neural network. Namely, the linear signal modelling is forced to be modified to non-linear signal modelling where the correlation between the linear and the non-linear models is greatly reduced by an appropriate choice of the algorithmic model of the neural network. Hence the signal- to-noise ratio where signal is the reflection of the change of system dynamics into the correlation and noise is the minimized correlation as obtained using neural network, is greatly improved as requried particularly in early fault detection and diagnosis applications. The utilization of feedforward neural network with noise signals for failure detection in process systems is demonstrated and verifications with the actual plant data with simulated failures on, are presented.