Multidimensional Attention-Based GAIN for Missing Data Imputation in Multimode Batch Processes

In batch processes, due to sensor failure, unstable signal transmission, and human negligence, the collected process data often contains missing values. This leads to the decline of process data quality and restricts the practical industrial application of data-driven process monitoring and control in batch processes. Existing missing data imputation methods always rely on complete training data and do not adequately consider the complex characteristics of batch process data in terms of temporal, variable, and mode dimensions. To this end, a multidimensional attention-based generative adversarial imputation networks (MDA-GAINs) for missing data imputation in multimode batch processes is proposed. Specifically, the constructed data imputation model can directly train using batch process data containing missing values by the missing mask matrix and hint mechanism of GAIN. Furthermore, the density peak clustering (DPC) algorithm is utilized to extract mode center features from multibatch average data, which are incorporated into the imputation model to enhance its capability of learning from multimode data. Building upon this, a multidimensional attention (MDA) module is integrated into the generator, which adaptively captures the temporal dependencies, variable correlations, and multimode characteristics in batch process data. By providing more comprehensive contextual information, the MDA-GAIN model can accurately impute the missing data. Finally, the proposed method is used to impute missing data in two typical multimode batch processes. The ablation experiment results show that adding the temporal-variable attention and mode attention, respectively, reduced the average RMSE of the imputation model by 14.6% and 19.2%, demonstrating the contribution of each dimensional attention module. Imputation results on two datasets indicate that the MDA-GAIN method reduces the average RMSE by at least 7.8% and 12.8% compared with the comparison methods with different missing rates (MRs), effectively improving the imputation accuracy of missing data.