Because of continuously updated fault data and categories, traditional artificial intelligence-based fault diagnosis methods frequently need more adaptability. When the fault diagnosis model is trained with data from new categories, it drops when handling old categories. This phenomenon is termed catastrophic forgetting. Although incremental learning is an effective solution, existing methods typically depend on static feature selection and cannot be dynamically adjusted to adapt to new data. Consequently, when dealing with new categories, the model fails to utilize existing information fully. This consequently impacts its ability to recognize categories. For this reason, we propose an adaptive channel module (ANCM) method. Under the control of resource occupation, this module dynamically weights and adjusts network channels. It uses adaptive parameters to highlight important channels and suppress unimportant ones to ensure that the model’s recognition ability for old categories will not be overly damaged when learning new categories. In addition, traditional loss functions typically only concentrate on individual samples and overlook the distribution differences among all samples. This might result in the loss of overall information. Based on this consideration, we design a logit maximum mean discrepancy loss function (LMMD) in transfer learning. It is intended to assist the model in learning the overall distribution of old category samples in the feature space, providing a stable mechanism for maintaining previously learned knowledge. Compared with classical algorithms, our method exhibits superior performance on the CWRU and MFPT datasets while effectively controlling the occupation of resources.