Research On Sensors Fault Diagnosis Approach Based On Multi-Source Information Fusion

With the gradual development of the automatic control level in the power plant, the control system is becoming the centrum of control system for safe operation in the power unit. The sensor network is the bridge to connect control system and the controlled devices. The parameters measured by sensors can reflect the operating process of the unit and the working state of devices. The measured value has important influence on the regulation quality of control system and safe operation of the power units. Therefore, the fault diagnosis of sensor is very necessary.Presently the Distributed Control System (DCS),widely used in power plant, can realize the real-time, omniberaing and multi-level supervision on the power unit even the whole power plant. All this lot of information from DCS is the precious resources for the fault diagnosis of the devices and the thermal system. Multi-source information fusion technology can make full use of the information. The more useful information could be derived from it by optimizing and combining them, and then the uncertainty of the information could also be reduced and more accurate and reliable information could be gotten for state supervision and fault diagnosis.Considering the problem of counterintuitive results in the face of significant conflict by D-S evidence combination rule,the principles, algorithm and features of several evidence combination rules are thoroughly studied, it is proposed that these combination rules should be chosen based on the conflict intensity and relevancy and how to deal with the conflict evidence.According to the problem of amount and complexity of calculation exponent increased with the measurement dimension increment when multi-data source related by D-S Fusion arithmetic, a fault diagnosis method combinied PCA and D-S evidence theory is proposed. Then, the characteristic of data dimension reduction of PCA and the completeness of fault diagnosis and the advantage of presenting and reasoning inaccurate information of evidence theory are utilized, and the problem of nonuniqueness of fault isolation of PCA and combination explosion of evidence theory is solved. And calculation amount and complexity are redyced quit, and the capability of fault detection and iaolated is improved significantly.According to the problem of multiple faults diagnosis of sensors, a method of refinement-fusion-coarsing based on evidence theory is proposed. The operators of refinement and coarsing are adopted to solve the problem of evidence fusion in the different but compatible discernment framework, and establish a certain connection between these frameworks.To reduce the uncertainty of believe assignment of sensors state under the incomplete condition, a modified arithmetic as multiplestep refinement-grouping fusion-coarsing is proposed to utilize the redundant and complementary information between known resource more adequately and efficiently. Accuracy and reliability of multiple sensors faults diagnosis couble be improved significantly.According to the features of multiple faults diagnosis, a king of two inputs and one output neural network based on RBF is proposed. Then, the problems of unusable brought by the change of input parameters and number using multiinputs- multioutputs neural network and of net training time long or rather not convergence caused by dimension or sample data troppo are be gotten over. This kind modularization neural network could be satisfied with the need of engineering and real time.