Deep learning-based process monitoring methods utilize the features extracted from deep neural networks to perform fault detection and diagnosis. Traditional deep learning models are constructed in a fully connected manner and stacked by a sequential connection style. However, the feature information of the observed data would be discarded as the information is compressed and propagated layer-by-layer in hidden layers. Motivated by this, an intensified iterative learning (IIL) model which is developed from the stacked autoencoder is proposed in this study. In the process of feature extraction of IIL model, the traditional constraints of the hidden layer connection in deep neural networks are disregarded and the feature information of the current hidden layer comes from the information of all previous hidden layers to avoid the loss of information. In the process of real-time process monitoring, the features which are advantageous to accomplish fault detection would be intensified to obtain the most favorable features for monitoring. Finally, Euclidean distance and reconstruction error are employed to indicate and visualize the process status. The monitoring performance of the proposed IIL model is evaluated on three process tasks, and the results show it outperforms other deep learning methods on fault detection.

基于深度学习的过程监控方法利用从深度神经网络中提取的特征来执行故障检测和诊断。传统的深度学习模型以全连接方式构建，并按顺序连接方式堆叠。然而，观察数据的特征信息将被丢弃，因为信息在隐藏层中被逐层压缩和传播。受此启发，本研究提出了一种从堆叠式自动编码器开发的强化迭代学习 (IIL) 模型。在IIL模型的特征提取过程中，忽略了传统的深度神经网络中隐藏层连接的约束，当前隐藏层的特征信息来自之前所有隐藏层的信息，避免了信息的丢失。在实时过程监控过程中，将强化有利于完成故障检测的特征，以获得对监控最有利的特征。最后，使用欧几里德距离和重建误差来指示和可视化过程状态。在三个过程任务上评估了所提出的 IIL 模型的监控性能，结果表明它在故障检测方面优于其他深度学习方法。