Abstract In modern manufacturing enterprises, quality-related soft sensors are important in production, especially for batch manufacturing processes. In practice, batch processes frequently produce new process modes due to various factors, such as changes in operating recipes, uncertainty in the external environment, and various product specifications. When this is the case, directly applying the historical model may lead to unexpected results. Moreover, it may be impractical to conduct enough trial runs and wait until sufficient batches are available before modeling for the new mode. Therefore, the problem of modeling for a new process mode with insufficient data brings challenges for the soft-sensor issue, which has rarely been addressed before. To solve the problem as mentioned above, a novel transfer and incremental soft-sensor scheme is developed for batch manufacturing processes with the support of multiple historical process modes, termed mode-cloud here. Using the proposed algorithm, the production data from the cloud of historical modes can be explored and utilized regarding their different phase-based relationships between process variables and product qualities. Thus, a new objective function is constructed to automatically identify and quantify the information buried in the vast ocean of historical data to determine the initial soft-sensor model for the new mode with limited data. Besides, with the constant increase of new samples, the initial soft-sensor model can incrementally update model parameters and release the workload of repetitive modeling. With both transfer modeling and incremental updating abilities, the proposed algorithm, which we call mode-cloud based transfer incremental learning (MTIL), can not only offer high adaptability and flexibility to accommodate a new process mode quickly, but also ensure the prediction accuracy. The MTIL based soft-sensor scheme is applied to the real injection molding process for the illustration purpose.

中文翻译：

---

基于模式-云数据分析的具有增量学习能力的制造业软传感器迁移学习

摘要 在现代制造企业中，与质量相关的软传感器在生产中非常重要，尤其是在批量制造过程中。在实践中，由于各种因素，例如操作配方的变化、外部环境的不确定性以及各种产品规格，批处理过程经常会产生新的工艺模式。在这种情况下，直接应用历史模型可能会导致意想不到的结果。此外，在为新模式建模之前进行足够的试运行并等到有足够的批次可用可能是不切实际的。因此，为数据不足的新工艺模式建模的问题给软传感器问题带来了挑战，这在以前很少被解决。为了解决上面提到的问题，为批量制造过程开发了一种新颖的转移和增量软传感器方案，支持多种历史过程模式，这里称为模式云。使用所提出的算法，可以探索和利用来自历史模式云的生产数据，了解它们在过程变量和产品质量之间基于不同阶段的关系。因此，构建了一个新的目标函数来自动识别和量化隐藏在历史数据海洋中的信息，以确定数据有限的新模式的初始软传感器模型。此外，随着新样本的不断增加，初始软传感器模型可以逐步更新模型参数，释放重复建模的工作量。具有传输建模和增量更新能力，所提出的算法，我们称之为基于模式云的转移增量学习（MTIL），不仅可以提供高适应性和灵活性以快速适应新的过程模式，而且可以保证预测的准确性。出于说明目的，将基于 MTIL 的软传感器方案应用于实际注塑成型过程。