基于对抗样本生成的时空特征融合软测量模型

A Soft Sensor Model Based on Adversarial Sample Generation and Spatio-Temporal Feature Fusion

摘要:

数据驱动的软测量方法在现代制造业中得到了广泛应用，它能够基于易于获取的传感器数据推断难以直接测量的关键过程变量。然而，在实际应用中，传感器故障往往会导致软测量模型的预测性能下降。针对这一问题，本文提出了一种具备传感器故障鲁棒性的软测量建模方法。考虑到传感器故障类型多样且真实故障数据稀缺，本文引入对抗学习方法，通过生成对抗样本参与模型训练。该方法仅依赖正常数据与生成的对抗样本，即可有效提升模型对传感器故障的适应能力，从而在故障发生时仍能保持对关键变量的准确预测。此外，本文在模型中引入专用模块，以充分挖掘过程变量的时空特征，进一步提升了软测量模型的预测性能。本文以均方根误差和平均绝对误差为评价指标，在青霉素发酵过程数据集上进行了实验验证。结果表明，在存在传感器故障的情况下，所提方法在预测性能上优于其他常用数据驱动软测量方法，展现出良好的故障鲁棒性；在无故障输入的情况下，该模型亦能保持较高的预测精度。

Abstract:

Data-driven soft sensor methods have been widely applied in modern manufacturing to infer key process variables that are difficult to measure directly from easily accessible sensor data. However, in practical applications, sensor faults often lead to a significant decline in the prediction performance of soft sensor models. To address this issue, this paper proposes a soft sensor modeling approach with enhanced robustness against sensor faults. Considering the diversity of sensor fault types and the scarcity of real fault data, an adversarial learning strategy is introduced to generate adversarial samples for model training. By leveraging only normal data together with the generated adversarial samples, the proposed method effectively improves the model’s adaptability to sensor faults, thereby maintaining accurate prediction of key variables when faults occur. Furthermore, a dedicated module is incorporated to fully exploit the spatiotemporal features of process variables, further enhancing the prediction performance of the soft sensor. The proposed method is validated on the penicillin fermentation process dataset, with root mean square error and mean absolute error used as evaluation metrics. Experimental results demonstrate that, under sensor fault conditions, the proposed model outperforms other commonly used data-driven soft sensor methods, exhibiting strong fault robustness; while under fault-free conditions, it still achieves high prediction accuracy.