Fault Diagnosis Of Power Transformer Based On Active Diverse Learning Neural Network Ensemble

Power transformer is one of the most important equipment in power system. Its operational status has greate importance to the system's safety. By monitoring the occurrence of latent failures in the power transformers and their progression, it can help to prevent the happence of major accidents. The gas's content and composition in oil are highly related to the fault type and the fault level when there are accidents in the power transformer. Therefore, it is in common use for transformer's diagnosis technology through analysing the gases and their change. Actually, there are complexity and uncertain mapping relationship between gas content and fault category, so it is a very difficult learning problem to build accurate mathematical model for power transformer fault diagnosis. And it is hard to resolve just relying on a single classifier. As a multi-classifier method, neural network ensemble has a natural advantage and greate applications foreground in transformer fault diagnosis by decomposing the learning task of into a number of subtasks for different individual networks. The diversity ensemble learning can resolve the randomness and blindness of individual networks generation and greatly reduce the networks'redundancy, through strengthening the activeness in designing networks and cooperation in parallel learning. And it is an effective neural network ensemble method for engineering application.The individual networks traning is independent and lack of cooperation in the traditional neural network ensemble method, due to randomness and blindnessm, it is difficult to ensure the diversity of the networks. An ensemble learning algorithm is proposed here by analyzing the error function of neural network ensembles, in which individual neural nerworks are actively guided to learn diversity. By decomposing the ensemble error function, error correlation terms are included in the learning criterion function of individual networks. And all the individual networks in the ensemble are leaded to learn diversity through cooperative training. Based on this idea, Active diverse learing(ADL) method is applied to fault diganosis of power transformer base on Dissolved Gas Analysis. The experimental results show that, under the same conditions, it doesn't require the individual network with high classification accuracy, but it can significantly improve the system's stability and generalization. This algorithm has higher accuracy than IEC method and BP network. And the performance is more stable than conventional ensemble method, i.e., Bagging and Boosting.A new divesity neural network ensemble called output-corrected data(OCD) is proposed in this paper. It differs from all previous ADL methods that instead of modifying every component network's error function to incorporate error correlation information, OCD modified the training data and creates sets of output-corrected data (OCD) as new training data, which induce diverse learing when component networks are trained on them. Previous ADL method developed can only assemble networks using backpropagation training algorithm, and demands prohibitively high communication bandwith(between component networks) that hinders parallel processing speed. In addition, every component network must be reprogrammed to include the error correlation terms in the training objective function, which raises difficulties to use the third party codes. These drawbacks significantly limit the practical application of ADL method for different classifiers. Thus, without the requirement of recoding each component network, new diverse learning method is simple to implement and can be used for assembling heterogenous networks, e.g., multilayer perceptrons and radial basis functions. Another major advantage is that OCD significantly reduces network communication bandwidth, making parallel processing more effective. The analysis indicates that OCD significantly reduces network communication cost. It is an effective neural network ensemble method for engineering application.By introducing activeness and cooperation and based on negative correlation diversity assessment, this paper studies two approachs to generate individual networks in ensemble,. Both of them were applied in fault diagnosis of power transformers, and achieved good results.