灰尘沉积会影响设备的正常运行。为准确且高效地对设备灰尘进行检测，制定科学的清洁策略，提出一种基于Fast-UNet的轻量化灰尘沉积检测方法。通过对UNet进行合理剪枝，采用最大池化与双线性插值进行下采样与上采样操作，减少了参数冗余，得到了结构紧凑的基础网络；使用轻量化的幽灵卷积模块（Ghost Module）替换基础网络中的普通卷积，进一步降低了网络的复杂度；在编码过程中嵌入融合了通道和空间注意力的CBAM（convolutional block attention module），在引入极少参数量的同时，使网络更加关注于目标区域。在模拟灰尘沉积的数据集上进行实验，结果表明：Fast-UNet与原模型相比，参数量减少了99.6%，计算量减少了98.7%，推理速度可达94.18帧／s，且识别精度高达91.17%。与其他5种主流分割模型相比，Fast-UNet在精度与速度上均具有优势。所提方法能够满足灰尘检测对检测精度与效率的需求，为灰尘定量分析提供技术参考。

Dust deposition can affect the normal operation of equipment. To accurately and efficiently detect dust on equipment and formulate a scientific cleaning strategy, a lightweight dust deposition detection method based on Fast-UNet was proposed. By effectively pruning UNet and adopting max pooling and bilinear interpolation for down-sampling and up-sampling operations, the parameter redundancy was reduced, and a compact basic network was obtained. The lightweight Ghost Module was used to replace the ordinary convolution in the basic network, further reducing the complexity of the network. An convolutional block attention module(CBAM) that integrated channel and spatial attention was embedded in the encoding process, which made the network pay more attention to the target area while introducing minimal parameters. Experiments on a simulated dust deposition dataset show that, compared with the original model, Fast-UNet reduces the number of parameters by 99.6%, decreases computational complexity by 98.7%, achieves an inference speed of 94.18 frames per second, and maintains a recognition accuracy of 91.17%. Compared with five other mainstream segmentation models, Fast-UNet also demonstrates advantages in both accuracy and speed. This method meets the needs of dust detection for both accuracy and efficiency, providing a technical reference for dust quantitative analysis.