Automatic and accurate identification on health condition of internal combustion engine system is still a major challenge in modern industry because of its complex mechanical system. In this paper，an innovative deep learning model called Random Convolutional Neural Networks（RCNNs）is proposed for intelligent health monitoring of internal combustion engine. This novel network framework is constructed with several individual convolutional neural networks，which can automatically extract the feature of vibration signals by convolutional calculation and pooling operation. An improved optimizer Adabound and the technique of Dropout are applied in this framework of RCNNs. The Adabound optimizer uses adaptive learning rates to adjust the network’s weight.Meanwhile，the Dropout technique makes neurons drop out with a probability in order to preventing complex co-adaptations on training data. The Dempster’s combinational rule is used to obtain the fusion diagnosis results from several individual networks.The proposed health monitoring scheme is used to analyze the experimental vibration signals acquired from engine. Experiments prove that the proposed RCNNs can overcome the limitations of health monitoring method based on traditional data-driven or convolutional neural network，which gets rid of dependence on the manual feature design and delivers state-of-the-art performances.Therefore，the proposed RCNNs method is suitable for machine health monitoring.