Research on Anomaly Data Preprocessing Technology for Deep Learning Soft Sensor Models Facing Missing Data and Fault Data

Deep learning-based soft sensor technology plays an important role in industrial process monitoring, however anomalous data such as sensor faults and missing values can severely compromise the predictive performance and reliability of the models. Most existing approaches address only a single type of anomaly, making it difficult to cope with the complex scenarios of multiple coexisting anomalies in real industrial environments. To overcome this limitation, this paper proposes a unified two-stage data preprocessing strategy that integrates anomaly detection and isolation with data reconstruction. In the first stage, a parallel Long Short-Term Memory (LSTM)Residual Network (ResNet) architecture is employed for fault detection and isolation to identify and separate abnormal data. In the second stage, an improved masked autoencoder model is applied to reconstruct the data at the detected anomaly positions, thereby leveraging the complementary strengths of fault detection and data reconstruction across different anomaly magnitudes. Experimental results on the Tennessee Eastman Process dataset demonstrate that the proposed method achieves an R² of 0. 9761, a MAPE of 0. 0636%, and an RMSE of 3.6915 in reconstructing anomalous data caused by both sensor faults and missing values.