Rotating machinery is a machine with a rotating component to do the energy transformation, which is widely used in vehicle engines, power plants, manufacturing factories, etc. Because of the continuous rotation, the bearing and gear are subject to a defect in the rotating machines, which damages the reliability of the machine. To provide a sustained normal working status, predictive maintenance (PdM) is usually applied to monitor the working health during the machine running by gathering different sensing data. However, highly diverse and massive sensing data increase the challenge to analyze the fault signals in the machine. Besides, the situation with composite faults (i.e., faults that happen in different components in a machine) worsens the difficulty to diagnose the target machine efficiently. To solve this problem, we propose a kind of hierarchy collaborative learning method in this work. Different from conventional centralized learning to analyze heterogeneous sensing data, collaborative learning uses multiple sub-learning units to analyze the local homogeneous sensing data priorly. Then, perform fusion operation for each sub learning unit obtained from homogenous sensing data to predict the final composite fault diagnosis results. In this way, we can not only ensure the quality of the signal fault diagnosis but improve the accuracy of the composite fault's diagnosis significantly. Compared with the centralized learning methods, the proposed collaborative learning method can achieve a 12% improvement in the accuracy of predicting the composite fault diagnosis.