Batch process quality prediction based on TGM-Reformer model

Abstract Batch process data often exhibit varying degrees of sparsity due to the issues such as inadequate measurement system performance and random data loss. The data sparsity makes it difficult for quality prediction model to capture useful information effectively, thereby severely limiting their prediction accuracy. To address this critical challenge, this paper proposes a temporal-gated multi-scale reformer (TGM-Reformer) model. First, the original three-dimensional batch process data are expanded into a two-dimensional matrix and standardized through preprocessing. Second, a multi-scale locality-sensitive hashing attention mechanism is designed to achieve efficient global feature extraction with high computational efficiency. Third, an adaptive temporal convolutional module is constructed to further enhance the model’s capability to capture local features. Finally, a memory feedback gating module is developed to strengthen the transmission and utilization of valid information while suppressing the interference of irrelevant information. Experimental results on two typical batch processes, namely penicillin fermentation and strip hot rolling, demonstrate that when batch process data suffer from sparsity, the proposed TGM-Reformer model achieves significantly higher prediction accuracy for key quality indicators than the compared models, strongly validating its effectiveness and superiority.