Commit messages provide a natural language description of the changes made to the code, enabling developers to swiftly comprehend the alterations without delving into the implementation complexities. Nevertheless, generating commit messages faces a considerable challenge due to the semantic and structural differences between code and natural language. Several researchers have put forward automated techniques aimed at generating commit messages. However, the full potential of code-related information is not yet fully harnessed. In this paper, we propose a Multi-channel based Code Change Representation Learning for Commit Message Generation(Mucha). We first compare the changed code and the corresponding AST before and after the change. Subsequently, we extract the altered information from various granularities and employ a multi-channel approach to capture the code changes, utilizing the extracted information as the basis for our analysis. In addition, we also use the query mechanism and attention mechanism to assist in learning the final code change representation. We build the experimental dataset, since there is still no publicly sufficient dataset for this task. The release of this dataset would serve as a valuable contribution towards advancing research in this particular field. We conduct a comprehensive experiment to assess the effectiveness of Mucha. The experimental evaluation demonstrates that our model outperforms the baseline model, which has significant improvements of at least 18.2%, 72.2%, and 10.5% against the baselines.