Research on nuclear power electronic sensor drift fault detection and data reconstruction based on the Seq2Seq-PCA method

摘要: As critical sensing components in industrial automation and control systems, electronic sensors are prone to
performance degradation and drift faults because of their harsh operating environments. Among these, minor
drift faults are characterized by slow variations and weak early-stage features, making them difficult to detect
promptly using conventional methods. Once such faults accumulate to a detectable level, theypose serious
threats to system safety and operational stability. To address this issue, this study proposes a sensor drift fault
detection and data reconstruction method based on the sequence-to-sequence model and principal component
analysis (Seq2Seq-PCA). The method first selects auxiliary variables through Spearman correlation analysis to
construct an input feature set. A Seq2Seq model with an attention mechanism is then employed for multi-step
rolling prediction to capture the dynamic system characteristics. Principal component analysis is applied to the
prediction residuals to establish a statistical monitoring model that enables sensitive detection of minor drift
faults. Upon fault detection, the multi-step prediction values of the Seq2Seq model are used directly used as
the reconstructed output, achieving seamless integration of fault detection and data reconstruction. The results
of experiments on a nuclear power plant simulator demonstrate the proposed method achieved an F1-score of
99.80%, false positive rate of 0%, and false negative rate of only 0.40% under steady-state operating condi#2;
tions with a drift rate of C1=0.00001. Under the extremely slow drift condition of, the F1-score remained at
96.75% under the extremely slow drift condition of C1=0.000001, outperforming conventional PCA (89.01%)
and LSTM-PCA (96.33%) by 7.74 and 0.42 percentage points, respectively. The reconstruction error was con#2;
sistently maintained within the order of 10⁻⁴ (with a maximum error < 3.5 × 10⁻⁴ ), confirming the capability
of the method for accurate detection and reliable reconstruction across various drift rates.