为了更准确地识别机械设备中的正常和异常声音，提出了一种融合注意力机制和域泛化技术的异常声音检测算法。具体来说，通过联合训练两个神经网络，结合使用子族群余弦自适应损失函数Sub-Cluster AdaCos（SCAdaCos），并采用高斯混合模型对声音特征进行建模。异常评分使用负对数似然值计算，并通过90%分位数设定检测阈值。实验结果显示，该算法在风扇、轴承等7类机器上，ROC（receiver operating characteristic curve）曲线下的面积AUC（area under curve）和部分曲线下面积pAUC（the partial AUC）的调和平均值分别为 76.69%和87.99%，其中阀门数据表现最佳。相比于两种基线系统，AUC和pAUC指标分别提高了24.08%和20.68%。进一步的消融实验表明，高斯混合模型、注意力机制以及SCAdaCos损失函数均对性能提升有显著贡献。与同一数据集上的其他8种算法相比，新算法在AUC和pAUC的调和平均值上提升了4.16%，展现了其在异常声音检测任务中的显著优势。

A new anomaly sound detection algorithm was studied that combines attention mechanisms and domain generalization techniques to more accurately identify normal and abnormal sounds in mechanical equipment. Specifically, two neural networks were jointly trained using a sub-cluster Adacos loss function, with features modeled by a Gaussian mixture model (GMM). Anomaly scores were calculated using negative log-likelihood values, and a 90th percentile threshold was set for detection. The algorithm demonstrated strong performance across seven types of machines, including fans and bearings, achieving harmonic mean AUC(area under curve) and pAUC values of 76.69% and 87.99%, with the highest performance observed on valve data. Compared to two baseline systems, the algorithm improved AUC and pAUC by 24.08% and 20.68%, respectively. Ablation studies further confirmed the positive impact of the GMM, attention mechanism, and Scadacos loss function. When tested against eight other algorithms on the same dataset, the proposed method showed a 4.16% improvement in the harmonic mean of AUC and pAUC, highlighting its significant advantage in anomaly sound detection tasks.