Defect distribution prognosis of high voltage circuit breakers with enhanced latent Dirichlet allocation

The degradation of high voltage circuit breakers is the result of the interaction between the equipment's attributes and the environmental stress such as the electrical, thermal or mechanical stress. Traditional methods of degradation analysis often rely on many theoretical assumptions to offer the detailed relationship between the machine's condition and the failure probability. However, this kind of theory approximation may not reflect the real complicated relationship. Besides, one of the challenges of predicting the defect distribution lies in the rich kinds of different defects. The sparsity of the defect kinds on a single circuit breakers makes it hard to establish a statistical significance. This paper introduces an data-driven latent Dirichlet allocation method to automatically build the quantitative relationship between the individual circuit breaker and its defect distribution on many kinds of defects. Firstly, a latent layer is introduced between the circuit breaker and its defect distribution as a matrix factorization method. The defect distribution of the circuit breakers is remodeled as the distribution between the circuit breakers and the latent layers, and the distribution between the latent layers and the defects based on Dirichlet allocation, which greatly reduced the state space to build a reliable discrete distribution on each kind of defect. Second, an Bayesian inference is also introduced as an online extension of the basic method. The method offers a new way to analyze large amounts of log data in the power grid, and has the ability to predict the defect distribution of a single circuit breaker. The experiment shows that the accumulated probability of the enhanced LDA method is 23.1% better than the statistical model and 6.2% better than the LDA model. the proposed method has a good performance of defect distribution prognosis and can offer reasonable operation advice.