Varying Balancing Transfer Learning for BN Parameter Estimation

To solve the problem of Bayesian network (BN) parameter estimation accuracy under small dataset conditions, this paper proposes a parameter Varying Balancing Transfer Learning algorithm (VBTL) based on varying weight transfer learning. Firstly, the MAP method and the MLE method are used to learn the initial parameters of the target domain and the parameters of each source domain. Then, the source weight factors of the source domain are obtained according to the different data source contributions. Based on the sample statistic the data size threshold values, the balance coefficients for the target initial parameters and the source domain parameters are calculated to obtain the final target parameters. The experimental results show that under the condition of the small data set, the learning accuracy of VBTL algorithm is better than MLE algorithm, MAP algorithm or classical transfer learning algorithm (LoLP). Under the condition of sufficient data set, the learning accuracy of VBTL algorithm approaches the classical MLE algorithm, and the correctness of the algorithm is verified. Moreover, we demonstrate the successful application to real-world bearing fault diagnosis case studies. Compared with the LoLP algorithm, the VBTL algorithm achieves about 10% enhancement for the average diagnosis precision.