(509d) Deep Transfer Network with Multi-Channel Feature Extraction for Cross-Domain Chemical Process Fault Diagnosis

In recent years, deep learning models have become increasingly popular for intelligent chemical process fault diagnosis. However, the majority of the developed approaches assume by default that the training and testing data is taken from the same feature distribution, which becomes invalid in actual application under multimode working conditions [1]. Additionally, it is widely recognized that the acquisition of fault data from manufacturing is quite limited due to the long time span of normal state [2]. Therefore, it is crucial to learn efficient representation of fault data and transfer the knowledge learnt from data-abundant to data-scarce scenarios for fault diagnosis.

Transfer learning presents a promising solution for the transfer of knowledge from a domain with ample labeled data to one with limited or no labeled data. Domain transfer is often realized by enforcing specific domain distribution metrics on one or multiple network layers within the deep transfer model. But it is worth noting that all input data for training is treated equally for domain adaptation, which may result in the loss of essential information. In spite of the adopted transfer learning approaches, the learning of an intermediate data distribution space is necessary to align the source and target domain data. Unfortunately, there is no assurance that the features of data belonging to the same category from different domains can be transferred to the same cluster in the intermediate feature space [3]. Hence, it is imperative to acknowledge that merely transferring all source samples to a shared feature space and utilizing the transferred cross-domain common features for fault diagnosis does not guarantee high classification accuracy.

In order to facilitate the preservation of domain-specific characteristics and the extraction of common features across both domains simultaneously, a novel domain adaptation deep network with multi-channel feature extraction is proposed in this research. The model structure is displayed in the attached image. The multi-channel deep transfer network (MDTN) employs a dual-channel architecture to achieve the concurrent acquisition of representations for common features across domains and discriminant features specific to individual domains. The domain-specific information carried by samples that are deemed unsuitable for transfer is effectively suppressed to fulfill precise classification. Polynomial kernel-induced maximum mean discrepancy (PK-MMD) is introduced to accomplish distribution adaptation between the source and target domains, while a relation score between the template sample and query sample is used to implement fault classification. The two combined modules including feature extraction and relation adaptation make the model diagnose faults efficiently even with limited training data.

The proposed model has been verified on the benchmark Tennessee Eastman process (TEP) and an industrial case of fluid catalytic cracking. Comprehensive results demonstrate that the model can obtain superior diagnostic performance compared with other transfer learning models. The designed multi-channel deep transfer network provides a new paradigm for cross-domain fault diagnosis and shows excellent capabilities not only in the specific simulation dataset but also in industrial sites. This general transfer learning approach will be further investigated for cross-device tasks, and the complete results will be presented in the long paper.

References:

[1] Bi, X., Qin, R., Wu, D., Zheng, S., Zhao, J., One step forward for smart chemical process fault detection and diagnosis, Computers and Chemical Engineering, 164, 107884 (2022).

[2] Wu, H., Zhao, J., Fault detection and diagnosis based on transfer learning for multimode chemical processes, Computers and Chemical Engineering, 135, 106731 (2020).

[3] Lu, N., Cui, Z., Hu, H., Yin, T., Multi-view and Multi-level network for fault diagnosis accommodating feature transferability, Expert Systems with Applications, 213, 119057 (2023).