Modern manufacturing industries are urgently demanding intelligent process monitoring systems for plant maintenance and accident prevention in the Industry 4.0 era. With the rapid development of deep learning, data-driven process monitoring methods are attracting wide attention and have been applied to many processes. However, most deep learning methods do not model process latent dynamics and are deficient to detect dynamic variations. In this work, a novel dynamic-inner convolutional autoencoder (DiCAE) is proposed. Unlike previous autoencoders that only focus on input reconstruction, DiCAE innovatively integrates a vector autoregressive model into a 1-dimensional convolutional autoencoder to monitor nonlinear processes, as well as capture process dynamics. When applied to a numerical simulation, DiCAE could detect the dynamic variation and distinguish different process data into separate clusters with an intuitive visualization, while other conventional methods cannot. The effectiveness of DiCAE is also demonstrated on the benchmark Tennessee Eastman process.

现代制造业在工业 4.0 时代迫切需要用于工厂维护和事故预防的智能过程监控系统。随着深度学习的快速发展，数据驱动的过程监控方法受到广泛关注，并已应用于许多过程。然而，大多数深度学习方法并没有对过程潜在动态进行建模，并且缺乏检测动态变化的能力。在这项工作中，提出了一种新颖的动态内部卷积自编码器（DiCAE）。与之前只关注输入重构的自编码器不同，DiCAE 创新地将向量自回归模型集成到一维卷积自编码器中，以监控非线性过程，以及捕捉过程动态。当应用于数值模拟时，DiCAE 可以检测动态变化，并通过直观的可视化将不同的过程数据区分为单独的集群，而其他传统方法则不能。DiCAE 的有效性也在基准 Tennessee Eastman 过程中得到证明。