【Purposes】 Bearing is one of the core components in the manufacturing industry. Its health status determines the safety of the host. Short-term failure prediction can effectively ensure the smooth progress of the industrial production process. 【Methods】 In order to solve the end-to-end problem, a temporal convolutional network (TCN) based short-term fault prediction strategy was proposed. The network could directly output the types of failure that would eventually occur in the bearing and the degradation stage that would be in the next moment through the data monitored at the current moment. In addition, soft threshold with attention mechanism is proposed to solve the problem of background noise in the working environment of bearings or noise interference in the process of data acquisition. During the short-term fault prediction process, the attention mechanism adaptively generates a soft threshold according to the prediction target of the TCN network, and the soft threshold acts on the spatiotemporal features extracted by the TCN to achieve the purpose of reducing noise impact. 【Findings】 The experimental results show that the proposed algorithm has high accuracy, which verifies the effectiveness and high practical engineering application value of the proposed algorithm.