Abstract: Anomalous sound detection (ASD) enables rapid, non-invasive machine condition monitoring. However, complex industrial noise environments and the acoustic similarity among different machine operating states can interfere with the accurate detection of anomalous sounds by the model. This paper proposes a multi-scale time-frequency feature modeling approach for ASD, enhancing the perception of anomalous patterns while maintaining computational efficiency. A densely connected convolutional encoder compresses input signals, while a multi-scale receptive field module captures spectrogram features along temporal and frequency dimensions. To further boost generalization and class separability, a hybrid data augmentation strategy and an additive angular margin loss are introduced. Experiments on the DCASE 2020 and 2021 Task 2 datasets show that the proposed method consistently outperforms state-of-the-art models, demonstrating its effectiveness.