Multiblock global orthogonal projections to latent structures for fault diagnosis

Abstract Although decentralized fault diagnosis based on multiblock regression models has achieved remarkable results, block information calculated at super block level by most traditional methods does not accurately address the subblock information. To address this information, multiblock diagnosis at block level was proposed, however, the model is complicated and fails to diagnose the output faults caused by external interference and/or noise. Therefore, in this paper, we propose a new fault diagnosis method based on multiblock global orthogonal projections to latent structures (MBGOPLS) to comprehensively diagnose process faults with a simple diagnosis scheme. This framework 1) obtains the full block/subblock output-related information using multiblock modified orthogonal projections to latent structures at block level, 2) combines all block/subblock output-unrelated information into a whole to establish appropriate output-unrelated statistics, and 3) further analyzes the input-unrelated information to diagnose input-unrelated faults. Thus, the proposed MBGOPLS results in a simpler model while retaining output-related and output-unrelated fault diagnosis ability, and capable of diagnosis in input-unrelated faults. A numerical example and a case study on a Tennessee Eastman process demonstrate the validity and superiority of the proposed method with respect to existing approaches.