Study Of Power Equipment Fault Diagnosis Based On Artificial Neural Network

Equipment fault diagnosis using neural network origined end of 1980s. In 1989, Venkat Venkatasubramanian and King Chan, American Berdu university, applied neural network in fault diagnosis, and compared it with professional system based on knowledge. They decided 18 kinds of symptoms (input nodes) and 13 kinds of exception classes ( output nodes ), and 5~27 nodes in hidden layer, which could decide 94%~98% of the fault reasons. But the disadvantage was that the training would last long, and the input data when training were not real time, and neural network mapping continuous variables was more difficulty than mapping boolean variables.Even though, they were the first to apply neural network in mode matching and fault diagnosis successfully.With the technical level of modern facility improving continually, the fault probability increases greatly. Power transformer has a very significant influence to power system, enterprise's production and people's life. How to forecast transformer's fault ahead of time and find the fault causes quickly after the fault happens is a good way to increase work efficiency and to lighten the economy losing.In the paper, ANN is applied to fault diagnosis system to overcome shortcomings of traditional fault diagnosis and reduce error brought by single method. Furthermore because fault symptom space and fault space have complicated non-linear relations, the mathematical model of diagnosis system is difficult to get. Artificial neural networks proposes a new way for this problem because of its advantages such as parallel processing, self-adaptation, self-study, association memory, non-linear mapping, etc. Thereby ANN is used as fault-classifying implement in exploitting the system.In the beginning of the paper, basic theories of fault diagnosis and ANN are presented. BP algorithm is performed by computer program. Some improving is done to the algorithm. Adding momentum item while correcting weight and limiting range of input value reduce error and improve diagnosis correctness greatly. While normalizing the input value, a new way is put forward that normalization is performed item by item according to its sort. In this way, error training can avoid going into the flat field that is caused by existing 0 or 1 of the input value. This way can improve the efficiency and correctness greatly. Combining ANN with Dissolved Gas Analysis (DGA), a typical method in transformer diagnosis, transformer fault diagnosis system with friendly interface and convenient capability is finished by GUI system, included in MATLAB toolbox. Besides, the method of selectingnetworks parameter is discussed detailedly, and the influence of different parameters on diagnosis results is analyzed. Finally, the paper summarizes the excellent capability of ANN fault diagnosis system and its shortcomings, and then analyses the outlook and development direction of ANN in fault diagnosis systems in the future.