Mechanical abnormal sound recognition(MASR) based on deep learning is an important technology for fault diagnosis.Current deep learning frameworks often assign uniform weights to both anomaly-specific acoustic signatures and background noise components, thereby impeding the discriminative extraction of critical fault features embedded in mechanical sound signals. In this paper, an abnormal sound dataset is constructed, and Abnormal Sound Channel Attention(ASCA) and Abnormal Sound Spatial Attention(ASSA) are studied respectively. An abnormal sound recognition method based on ASCA and another based on ASSA are proposed. By introducing the ASCA and ASSA, the methods focus on abnormal sound features from different perspectives and suppress non-fault-related features. The recognition accuracy of networks with ASCA and ASSA improved by 4% and 1% respectively.In addition, the recall rate, precision, and F1-score of the network have also been significantly improved. Experimental results show that the proposed method enables the network to focus on the differences in the low-frequency segment of the line spectrum between normal prototypes and abnormal sound prototypes, improving the recognition accuracy of neural networks for abnormal sounds.