Research On Recognition And Diagnosis Of Utility Boiler Combustion State

Combustion optimization is one of the effective ways to realize energy saving and emission reduction in thermal power plants. Judging combustion state quickly and accurately is an important prerequisite for combustion optimizing control. Existing flame monitoring devices are nearly used to judge whether there is flame, and therefore combusiton state cannot be evaluated. Based on correlations of combustion states and measured signals, this dissertation finds out related signals of combustion states from large amounts of measured data, picks up features which can reflect variation of combustion states via analysis of statistical rules, and then identify and diagnose combustion stability by data classification and information fusion techniques respectively. The four questions of signals selection, features extraction of combustion states, recognition and diagnosis on combustion stability are studied through analysis of data stored in Supervisory Information System in Plant Level (SIS), and the main work of this dissertation can be presented as following:1. Selection of signals correlated with combustion states. Firstly, a new correlation analysis method was proposed based on calculating the variance of data in sliding window, which could effectively evaluate signal fluctuation similarity of combustion characteristic frequency range. Then, a multi-scale correlation analysis of thermal signal based on wavelet transform was proposed to study their correlation in different frequency ranges. Finally, the proposed method was applied in a 600 MW thermal unit, and flame monitoring signals, furnace pressure, main steam pressure, drum water level and oxygen content before air pre-heater were chosen as related signals.2. Feature extraction of combustion states. Characteristics in time-domain of selected signals under eight typical conditions were compared. Signal features were extracted via statistics and complexity measures, and mean values, standard deviations, peak-to-peak values and complexity were selected as features of related signals.3. Combustion states pattern recognition was studied by extracted features. Firstly, self-organizing neural network was adopted to cluster combustion features of different working conditions, and the result indicates that the neural network could analyze the nonlinear mapping relation between combustion state and their features. Then, the square support vector machines were used to recognize combustion states. The SVMs model parameters were chosed by grid-searching combined with cross-validating. Finally, the classication accuracy of three types of SVMs was compared and the effectiveness of the method was verified through on-site data calculation.4. According to information fusion theory, combustion stability diagnosis was studied based on evidence theory. With regard to the question of how to obtain basic probability assignment (BPA), a new method based on self-organizing neural network was used to obtain BPA values. Then, related signals were fused via evidence theory, and fused information was adopted to diagnose combustion states. The data analysis results showed that the proposed method was effective for combustion stability diagnosis, and its accuracy and reliability were better than that of judging by a single sensor such as flame monitoring signals or furnace pressure.