Predrag Tadic

Fault diagnosis for steam separators based on parameter identification and CUSUM classification

A method for diagnosing faults in steam separators is presented. Faults in the water level, water flow and steam flow sensors are analyzed. Precise models of the steam separator system are difficult to obtain, which makes the most common model-based fault detection and isolation approaches unapplicable. An identification-based method is used instead: parameters of the process are identified in real time, and the resulting data samples, which we denote as residuals, are used as inputs to a CUSUM-type classification scheme. It then decides if a fault is present, and if so, which one. In other words, residuals are first generated by parameter identification, and then evaluated by a modification of the CUSUM test. The choice of the CUSUM algorithm was motivated by its optimality with respect to detection delay. The identified parameters are assumed to be normally distributed. This assumption is experimentally verified: the true probability density functions (PDF) are estimated, and the performance of the detector based on these estimated PDFs is compared to that of the previous detector, based on the Gaussian PDF. The proposed method was tested on real-world data, obtained from the TEKO B1 Unit of the Kostolac Thermal Power Plant in Serbia. The results suggest extremely low probabilities of false alarm, missed detection and false isolation. As for detection delay, just one residual sample is needed for proper fault diagnosis in some cases, while 83 samples are needed in the worst-case scenario.

Hand gesture recognition using neural network based techniques

In this paper, two neural network based methods were implemented for recognition of images showing 10 hand gestures. Images were available from 24 subjects and captured on two different backgrounds and with several space orientations. Firstly, Histogram of Oriented Gradients method was applied for feature extraction and training was performed with multilayer feed forward neural network with back propagation algorithm. Within the second method, Sparse autoencoder with 5 hidden layers and decreasing number of neurons was implemented. For both methods it was examined how number of descriptors influences the accuracy of classification and found relationship was used to determine best performing case. Both classification methods achieved accuracy of about 92.5%, by using the similar number of estimated parameters.

On signal-to-noise ratio estimation

A new simple algorithm for estimating signal-to-noise ratio (SNR) for a signal consisting of one sinusoid in white Gaussian noise is proposed in this paper. Algorithm is based on autocorrelation and modified covariance methods for AR (Autoregressive) spectral estimation. The validity of the algorithm is examined by comparing its SNR estimate with SNR estimate obtained by sinusoid magnitude estimation using Pisarenko harmonic decomposition method and noise variance estimation using modified covariance method. By a large number of simulations this algorithm has proven itself as a comparably precise even in case of significantly noise-contaminated sinusoidal signal.

TOWARD ACOUSTIC NOISE TYPE DETECTION BASED ON QQ PLOT STATISTICS

Fault detection and state estimation using acoustic signals is a procedure highly affected by ambient noise. This is particularly pronounced in an industrial environment where noise pollution is especially strong. In this paper a noise detection algorithm is proposed and implemented. This algorithm can identify the times in which the recorded acoustic signal is influenced by different types of noise in the form of unwanted impulse disturbance or speech contamination. The algorithm compares statistical parameters of the recordings by generating a series of QQ plots and then using an appropriate stochastic signal analysis tools like hypothesis testing. The main purpose of this algorithm is to eliminate noisy signals and to collect a set of noise free recordings which can then be used for state estimation. The application of these techniques in a real industrial environment is extremely complex because sound contamination usually tends to be intense and nonstationary. The solution described in this paper has been tested on a specific problem of acoustic signal isolation and noise detection of a coal grinding fan mill in thermal power plant in the presence of intense contaminating sound disturbances, mainly impulse disturbance and speech contamination.

Fault diagnosis in nonlinear stochastic systems via particle filtering

We consider the problem of detecting malfunctions in the actuators or sensors of systems which can be described by nonlinear/non-Gaussian stochastic state-space models. The basic idea is to estimate the state vector of such models using a sequential Monte Carlo technique known as particle filtering. We present several approaches to detecting faults and pinpointing their location within the system, using either one or a bank of particle filters.

To robust identification of water steam-separator process in thermal power plants

One of the primary requirements for a high performance process control is a good quality and reliability of obtained measurements. In order to overcome the problem of sporadic high-intensity irregular measurements (outliers) presence, a robust process identification procedure must be used. The paper presents an application of an adaptive approach to robust parameter estimation of a linear dynamic discrete-time system, based on QQ-plot method together with robustified winsorization technique. The proposed procedure is implemented in a stem separator system in thermal power plants. The comparison to the conventional RLS approach demonstrates the effectiveness of this method in the presence of impulse noise.

A new approach to Doppler filter adaptation in radar systems

This paper presents one way to adaptation of Doppler radar filters. Doppler filters use window function in order to estimate the target velocity. The window function length can be adapted depending on estimated target acceleration and estimated signal to noise ratio in the system. Kalman filter is used for estimation of target acceleration. Signal to noise ratio is estimated using a procedure based on parametric methods of spectral analysis. A new adaptive Doppler-Kalman structure has been presented and efficiency of approach to target velocity estimation is proven by computer simulations.

Thermal Power Plant Fan Drive Load Distribution Control

Abstract This paper presents an approach to pressure/under-pressure and load distribution control of fresh air supply and exhaust fans at thermal power plants, where dual actuators are used. The proposed concept, relative to standard solutions, offers better performance, simplified tuning and start-up, and safer operation in terms of over-current protection. Pressure/under-pressure and load distribution of the two fans by means of flap position relies on two PI pressure controllers, two PI current controllers (when over-current protection is active), and one PD controller for load distribution control between the two drives.

Coal-Shortage Detection in Power Plants by Means of a Fixed Size Sample Strategy *

Abstract An application of a fixed-size sample (FSS) hypothesis testing algorithm to the problem of detecting shortages in the coal supply subsystems of thermal power plants is presented. Blockage of coal supply leads to the occurrence of possibly hazardous operating conditions; it is therefore treated as a fault, and its detection is posed as a fault detection and isolation (FDI) problem. Physical variables of interest are treated as stochastic processes, and the fault modeled as a change in their mean values. Application of the chosen change detection strategy to data taken from a real process is discussed in detail, and the performance of the algorithm is experimentally verified.