The effective and in-time detection of bolt looseness in the brake disc of high-speed rail is of great significance to ensure the safe operation of the train. The hammer tapping method is one of the most widely used methods for bolt looseness detection. Due to the complex structure of the brake disc, the vibration characteristics associated with bolt looseness are coupled with the transmission path from the head of the bolt through the brake disc to the sensors, which make it difficult to quantitively identify the bolt looseness. In this work, a method based on wavelet packet decomposition and a one-dimensional convolutional neural network (1D CNN) is proposed to quantitively detect the bolt looseness of brake disc. Firstly, the vibration signals collected from the two sensors mounted on the nut side are fused by autocorrelation summation to reduce the influence of transmission path. Secondly, the fused signal is decomposed to sub signals in low frequency and high frequency components by wavelet packet decomposition. The relative difference of wavelet packet energy among sub signals is extracted as the features to enhance the difference among different degrees of looseness. Finally, the 1D CNN model is established and trained by the features of energy relative difference to quantitively identify the bolt looseness. To validate the effectiveness of the proposed method, an experimental platform for bolt looseness detection in brake disc is constructed. Compared with the single-channel 1D CNN and fused signal-1D CNN models, the accuracy of the proposed method is approximately 97%, which confirms the superiority of the proposed method.