A Novel Discount-Weighted Average Fusion Method Based on Reinforcement Learning For Conflicting Data

Dempster-Shafer theory (DST) is widely used in multisensor data fusion because of its effectiveness in dealing with uncertain data. However, when the sensor data is highly conflicting, counterintuitive fusion results may be obtained. To implement intelligent fusion of conflicting data, a novel discount-weighted average fusion method based on reinforcement learning (RL) is proposed. First, an adaptive weight adjustment method based on RL is devised, which can make each data have different reliability. Next, the weights are used to discount the data to obtain highly reliable data. Then, considering the uncertainties of data, DST is utilized to achieve the discount-weighted average fusion. In addition, since the prior knowledge is unable to be obtained, the information volume of data is measured to set a reward function to improve the fusion accuracy. Ultimately, a fault diagnosis example of conflicting data is given to illustrate the effectiveness of the proposed method. The results show that our proposed method for fault diagnosis outperforms other methods, where the belief value is 91.29%.