Robust and constrained tracking of PSV interface using convolutional neural networks and optimistic moving horizon estimation

This manuscript proposes a novel video-based robust and constrained estimation framework using the convolutional neural network and optimistic moving horizon estimation, with applications in interface estimation of oil sand primary separation vessels (PSV). Although convolutional neural networks have achieved notable success across various computer vision and image analysis tasks, image outliers (such as blocking, blurriness, and lighting variations) would inevitably affect recognition/tracking performance. To address this issue, this manuscript proposes a robust estimation approach by leveraging a convolutional neural network and moving horizon estimation. Along this line, the interface recognition results by the convolutional neural network can be modeled as the measurements corrupted by disturbances and outliers, and the internal states can be modeled through a discrete-time finite-dimensional state space model. More importantly, the ubiquitously present constraints in the estimation task can be explicitly and readily handled by the moving horizon estimation. The stability analysis of the proposed method is provided in the presence of disturbances and model-plant mismatch. The effectiveness of the proposed method is validated through a pilot-scale laboratory study and an industrial primary separation vessel case study. • A combined CNN and MHE approach for interface estimation is proposed. • The proposed method can explicitly handle physical constraints in the estimation. • The proposed method is robust to camera blocking, blurriness and lighting variations. • The stability is proven in the presence of disturbances and model-plant mismatch. • Laboratory and industrial videos are used to verify the proposed method.