Recent multivariate time series classification methods typically extract features from all variables (channels) but struggle to dynamically analyze and select channels within the model, and usually focus only on the local dependencies between channels. To address the above challenges, we propose a channel scores module based on the self-attention mechanism, which is designed to capture global dependencies between channels and generate channel scores. By adding the channel scores to the time series data, the influence of important channels is effectively enhanced, while the impact of less relevant channels is diminished, thereby boosting the classification performance of the model. We further propose a multivariate time series classification model based on the Transformer architecture, named ChannelFormer, which adds the channel scores module to the model framework to improve the quality and utilization efficiency of time series data. Experiments on 30 benchmark datasets show that our model outperforms other state-of-the-art methods across multiple metrics, demonstrating its significant potential and exhibiting the effectiveness of the proposed channel scores.