This article proposes a novel adversarial training method called domain-adaptive adversarial training (DAAT) to enhance the adversarial robustness of deep learning-based soft sensors (DLSS). The method consists of two stages: historical gradient-based adversarial attack (HGAA) and domain-adaptive training. HGAA incorporates historical gradient information into the adversarial sample generation process, stabilizing the update direction and improving transfer gradient estimation. Domain-adaptive training then learns common features from adversarial and original samples, enhancing adversarial robustness without overfitting. The effectiveness of DAAT is demonstrated using a DLSS model for crystal quality variables in silicon single-crystal growth manufacturing. The results show that DAAT achieves a balance between adversarial defense and normal sample prediction accuracy, offering an effective approach for enhancing DLSS adversarial robustness. The method addresses challenges in transfer gradient estimation and adversarial robust overfitting, which are common issues in adversarial training for DLSS. The proposed DAAT method is validated through experiments on a complex industrial process, showing improved adversarial robustness and prediction accuracy compared to traditional adversarial training methods. The study highlights the importance of developing robust adversarial training methods to ensure the safety and reliability of DLSS in critical industrial applications.This article proposes a novel adversarial training method called domain-adaptive adversarial training (DAAT) to enhance the adversarial robustness of deep learning-based soft sensors (DLSS). The method consists of two stages: historical gradient-based adversarial attack (HGAA) and domain-adaptive training. HGAA incorporates historical gradient information into the adversarial sample generation process, stabilizing the update direction and improving transfer gradient estimation. Domain-adaptive training then learns common features from adversarial and original samples, enhancing adversarial robustness without overfitting. The effectiveness of DAAT is demonstrated using a DLSS model for crystal quality variables in silicon single-crystal growth manufacturing. The results show that DAAT achieves a balance between adversarial defense and normal sample prediction accuracy, offering an effective approach for enhancing DLSS adversarial robustness. The method addresses challenges in transfer gradient estimation and adversarial robust overfitting, which are common issues in adversarial training for DLSS. The proposed DAAT method is validated through experiments on a complex industrial process, showing improved adversarial robustness and prediction accuracy compared to traditional adversarial training methods. The study highlights the importance of developing robust adversarial training methods to ensure the safety and reliability of DLSS in critical industrial applications.