在工业场景中，因为设备异常现象的罕见性和高度多样化，以及机器的操作条件或环境噪声在训练和测试阶段的不同，会改变训练和测试数据之间的声学特性。为解决上述问题，提出一种基于联合深度学习和变分贝叶斯高斯混合模型的无监督异常声音检测算法。通过两种神经网络联合训练进行信息提取，并利用变分贝叶斯高斯混合模型对其所获得的嵌入进行聚类分析；引入一种新的混合示例数据增强方法，用多种方式相结合的替代方法来生成示例，以对齐不同域之间的分布；应用了一种改进的子集群AdaCos损失函数，以排除潜在的异常值。实验结果表明，该方法在三种工业机器类型的数据集上目标域的平均曲线下面积达到了79.03%，平均F₁分数达到了67.23%；对比基线模型，谐波平均值提升约20%，在工业设备无监督异常声音检测中表现良好。

In industrial scenarios, acoustic properties between training and test data would be changed due to the rarity and high diversity of equipment anomalies, as well as differences in machine operating conditions or ambient noise during the training and test phases. To solve these problems, an unsupervised anomalous sound detection algorithm based on joint deep learning and variable Bayesian Gaussian mixture model was proposed. The information was extracted through the joint training of two kinds of neural networks, and cluster analysis was carried out on the embedding obtained by variable Bayesian Gaussian mixture model. A new hybrid sample data enhancement method was introduced to generate samples with a combination of alternative methods to align the distribution between different domains. An improved sub-cluster AdaCos was applied to exclude potential outliers. The experimental results show that the average AUC (area under curve) of the target domain on the datasets of three industrial machine types reaches 79.03%, and the average F₁ score reaches 67.23%. Compared with the baseline model, the average harmonic value is improved by about 20%, and it performs well in unsupervised anomalous sound detection of industrial equipment.