Predictive business process monitoring (PBPM) is particularly valuable in dynamic business environments, and it can help organisations mitigate risks and optimise resource allocation. An interesting task in PBPM is next activity prediction (NAP), which allows the prediction of future activities that will be executed at a certain time based on ongoing business processes. Existing methods typically only utilise the order information of traces when predicting the next activity, without fully leveraging the attribute information present in the logs. Given the usefulness of these for NAP, combining them can help neural networks gain a deeper understanding of the actual business process. In this study, we propose a dual-view deep learning approach to fully extract and fuse the aforementioned two aspects of information. First, we treated traces as sequential texts and extracted the trace order information based on a long short-term memory based self-attention network. Then, we treated traces as unstructured images and captured the implicit attribute fusion information among events using a 12-layer residual network. Finally, two parts of information were fused for NAP. Experiments on 12 real-life event logs prove that the proposed approach is superior to state-of-the-art approaches, exhibiting good performance in accuracy, macro-precision, macro-recall, macro-F1-score, and macro-Gmean.

Convolutional neural networks (CNNs) have been widely used in fault diagnosis due to their superiority in feature extraction. Traditional CNNs are a type of closed-box techniques with little interpretability, and their effectiveness is greatly affected when fault mechanisms and modes are extremely complex. To cope with such issue, this article presents a way to integrate outlier-type prior knowledge into CNNs based on an attention mechanism for fault diagnosis. First, outliers of the image-like data obtained by a sliding window processing from the raw data are formally defined as prior knowledge. Then, the defined outlier-type prior knowledge is integrated into any layer of CNNs by a parameter-free attention mechanism. Compared with existing similar methods, the proposal realizes a novel and flexible definition of prior knowledge and achieves deep fusion of prior knowledge and CNNs with low computational cost. The performance of the proposal was evaluated on the Tennessee Eastman process dataset and the real wind turbine blade icing dataset, which indicates that the proposal could not only realize accurate results but also had good model interpretability in terms of achieving high accuracy. The acquisition of outlier-type prior knowledge was discussed and the results demonstrate the effectiveness of the proposed prior knowledge integration method.