Convolutional neural networks (CNN) have achieved significant advancements in intelligent fault diagnosis of rotary machines. However, the limitations of using a single scale convolution kernel in convolutional layer and the exclusive focus on classification accuracy by the cross-entropy loss function during model training result in suboptimal diagnostic performance and generalization ability of CNNs in environments with strong background noise and imbalanced data. To address these challenges, a fault recognition method for rotary machines utilizing a minimum maximum regularized multiscale CNN (MMRMCNN) is proposed. A multiscale feature extraction module is devised, which uses convolutional layers with diverse scale kernels to capture multiscale features form input data. Additionally, a minimum maximum regularized objective function is introduced to augment the original cross-entropy loss function. This modification enables the model to consider not only recognition accuracy but also the compactness within classes and separation between classes of learning features during network training. The proposed approach effectively narrows the intra class margin of device health status features while widening the inter class margin, thereby mitigating the impact of noise and data imbalance on the mapping of health status relationship. Performance evaluation of the MMRMCNN is conducted using a measured dataset, the PU bearing dataset, and a rotor dataset. We found that the fault recognition accuracy of the proposed method exceeds 97.79 %, and the accuracy of fault recognition under noisy background and unbalanced data conditions is also above 94.81 % and 94.72 %, respectively. This demonstrate the superior fault recognition capabilities of the proposed method in scenarios characterized by strong background noise and data imbalance. Overall, the results attest to the exceptional performances of the developed MMRMCNN in fault recognition under challenging conditions, underscoring its potential in advancing the field of in Telligent fault diagnosis for rotary machines.