Vibration signals are widely applied in mechanical fault diagnosis methods. Strong and nonstationary noise can easily affect the vibration signals, and the vibration signals will differ depending on the load conditions. Additionally, there are not many vibration signals with faults that can be employed because mechanical equipment typically operates as intended. A brand-new data augmentation and composite multiscale network (DACMSN) for mechanical fault identification is proposed as a solution to these issues. The method of data augmentation is to generate samples simply and efficiently through a mathematical permutation and combination method. Each generated sample consists of two original samples, and each generated sample contains all fault information of the original sample. The dense multiscale convolution neural network (DMSCNN) and the multiscale residual network (MSRN) make up the composite multiscale network (CMSN). The CMSN has rich multiscale feature extraction capabilities and coarse-fine feature extraction capabilities. Convolutional kernels of various sizes are used in DMSCNN, which also takes into account the results of previous layers to initially discover fault features. MSRN further performs multiscale feature extraction on each subnetwork of DMSCNN, achieving the effect of fully exploring fault features and effectively accelerating loss backpropagation. The proposed model can demonstrate unique advantages through the use of these mechanisms, such as a wider feature extraction scale and a more complete ability to extract coarse-fine features. This article compares the proposed method with different methods using noise and variable load dataset for tests. The experimental results demonstrate that the proposed method can perform well for fault diagnosis even in noisy environments.