Due to the couplings among different vibration sources and other factors， single and compound faults in the gearbox usually show specific shared characteristics， leading to inaccurate diagnosis results. Meanwhile， the traditional convolutional neural network （CNN） based methods often show poor robustness when facing such circumstances. This paper proposes a new intelligent identification method of gearbox fault based on a one-dimensional CNN （1DCNN） to improve the performance of the above methods. In this network， the LeakyRelu activation function is introduced to replace the activation function in the convolution layer of the original network structure to prevent neuron failure during training. The LookAhead optimizer is used to avoid the training result converging to the local extremum during reverse parameter optimization. Also， a similarity loss measurement function is proposed to minimize the sequence spacing of similar samples and maximize the sequence spacing of different samples to strengthen the label recognition ability and classification stability of the network structure. The above network is named sLL-1DCNN. Moreover， the signals collected on the gearbox fault simulation test-bed are used to train the network and identify various types of faults. The results show that the network has better feature extraction and generalization ability when the number of sample sequences in the training set is small. The classification ability and stability are better than the other three CNNs when the number of sample sequences in the training set increases.