面向电厂设备异常检测的自组织映射深度自编码高斯混合模型研究

Research on Self-Organized Mapping Deep Autoencoding GMM for Power Plant Equipment Anomaly Detection

工业领域的生产设备异常检测实际上是采用无监督技术准确预测设备早期劣化的异常工况和定位具体的异常参数。深度自编码高斯混合模型DAGMM在生产设备数据集上异常检测性能较优,但仍有提升空间。针对深度自编码存在的高维信息丢失的问题,提出使用自组织映射辅助均匀流形近似与投影改进的模型SOM-UMAP-DAGMM。通过将UMAP算法改造为神经网络,在原来的损失函数上新增一项交叉熵实现与DAGMM联合训练,补充高维数据分布信息;并结合预训练SOM,补充空间拓扑结构信息。在两个公开数据集和3个生产设备数据集的实验结果上显示,SOM-UMAP-DAGMM较DAGMM性能得到了一定的提升。

Anomaly detection of production equipment in industrial field is actually using unsupervised technology to accurately predict the abnormal condition of early deterioration of equipment and locate specific abnormal parameters.The Deep Autoencoding Gaussian Mixture Model(DAGMM)has better anomaly detection performance in production equipment data sets,but there is still room for improvement.Aiming at the problem of high-dimensional information loss in deep autoencoding,this paper proposes a model named Self-organized Mapping-assisted(SOM)Uniform Manifold Approximation and Projection(UMAP)DAGMM.And then,by transforming the UMAP algorithm into a neural network,the paper adds a new cross entropy to the original loss function,so as to realize joint training with DAGMM and supplement the high-dimensional data distribution information.Finally,the paper combines with pre-training SOM to supplement the space topology structure information.The results of the experiment conducted in the two public data sets and three production equipment data sets show that the performance of SOM-UMAP-DAGMM has got a certain improvement compared with that of DAGMM.