With the expansion of modern plants and the growth of process complexities, industrial processes generally have multimodes, which are reflected in the multiple distributions and complex correlations among the process data. On the basis of single mode process monitoring methods, research on multimode process monitoring have attracted much attention. At present, most of it has been based on prior process knowledge, including the number of operating modes and the historical dataset for each mode, which are usually unavailable in many cases. This paper proposes a new multimode process monitoring method based on the hierarchical Dirichlet process (HDP) and a hidden semi-Markov model (HSMM). Firstly, HSMM is used to overcome the limitation of state durations in the traditional HMM. Then, HDP is introduced as a prior of infinite spaces solving the problem of missing mode information. Secondly, based on the HDP-HSMM framework, an automatic mode classification and identification strategy, including unknown mode identification, is established. Finally, a global–local monitoring strategy is put forward based on the Mahalanobis distance and the negative log-likelihood probability to give multimode process monitoring has been achieved. The effectiveness of the proposed method is verified on the Tennessee Eastman process and a real hot strip mill process (HSMP).

随着现代工厂的扩大和过程复杂性的增长，工业过程通常具有多模式，这反映在过程数据之间的多种分布和复杂的相关性中。在单模过程监控方法的基础上，对多模过程监控的研究引起了广泛的关注。目前，大多数方法都是基于先前的过程知识，包括操作模式的数量和每种模式的历史数据集，而在许多情况下通常是不可用的。提出了一种基于分层狄利克雷过程（HDP）和隐马尔可夫模型（HSMM）的多模式过程监测新方法。首先，HSMM用于克服传统HMM中状态持续时间的限制。然后，HDP是无限空间的先验，它解决了模式信息丢失的问题。其次，基于HDP-HSMM框架，建立了一种自动模式分类识别策略，包括未知模式识别。最后，基于马哈拉诺比斯距离提出了一种全球-局部监测策略，并实现了对数似然概率负的多模式过程监测。该方法的有效性在田纳西州伊士曼工艺和真实的热轧机工艺（HSMP）上得到了验证。提出了一种基于马哈拉诺比斯距离的全球局部监测策略，并实现了对数可能性的负对数可能性，从而可以进行多模式过程监测。该方法的有效性在田纳西州伊士曼工艺和真实的热轧机工艺（HSMP）上得到了验证。提出了一种基于马哈拉诺比斯距离的全球局部监测策略，并实现了对数可能性的负对数可能性，从而可以进行多模式过程监测。该方法的有效性在田纳西州伊士曼工艺和真实的热轧机工艺（HSMP）上得到了验证。