Abstract: As an important mechanical transmission device, the health status of the gearbox directly affects the operational reliability of the equipment, and conducting degradation state prediction for it is of great significance. Effective degradation indicators and high-precision degradation trend prediction models are the key elements to ensure the accuracy of degradation prediction. Most traditional methods for constructing degradation indicators are based on the manifestation of the overall health status of the gearbox in data, and it is difficult to achieve precise positioning of the faulty gears inside the gearbox. Based on the knowledge of the degradation mechanism of the gearbox, this paper proposes a method for constructing degradation indicators based on side-band energy. Compared with other data-driven indicators, it can achieve the positioning of the faulty gears inside the gearbox and has better monotonicity and trend characteristics. Subsequently, an attention mechanism and a temporal convolutional neural network (TCN) are integrated to construct a degradation trend prediction model, which comprehensively utilizes the abilities to learn sequential features and improves the accuracy of gearbox degradation prediction. Experimental verification was carried out using the gearbox degradation dataset of Chongqing University, and the results show that the proposed prediction method can achieve higher prediction accuracy compared with the contrast models.