In order to solve the problem that it is difficult for a single sensor to accurately characterize the running state of rotating bearings under complex working conditions, this paper proposes a data-level fusion method based on multi-source isomorphic sensors to monitor spindle bearings. First, new vibration signals in the X,Y,Z direction were obtained through the process of decomposing, de-noise, and reconstructing. Second, the PCA algorithm was used to select the time-domain and frequency-domain features of the vibration signals, construct the feature matrix, and perform dimensionality reduction in the feature matrix. Finally, the entropy weight method was introduced to obtain the initial weights of the three directions as the inputs of the adaptive function. The chaotic particle swarm optimization algorithm proposed in this paper helps particles jump out of the local optimum. Chaotic mapping is used to initialize the velocity and position of the particles, which calculates globally optimal weights in three directions. In order to extract bearing signal features more accurately and efficiently, a DenseNet and Transformer (DAT) feature extraction model is proposed to deal with the complex changes and noise interference of bearing signals. Through the open data set of Jiangnan University and the data collected by our own experimental platform, the maximum accuracy of the DAT model was verified to be 100%.