<p>Digital twin models are typically driven by either sensor feedback signals from physical entities or control command signals, both of which may experience anomalies that affect the accuracy of the digital twin model. To improve the accuracy of the digital twin model’s representation of the physical entity, this paper proposes assessing the credibility of the two driving signals separately and then fusing them into a more reliable driving signal. This paper takes industrial robot joint-angle sensor signals as an example for studying signal credibility assessment, proposing to first perform anomaly detection on the signals and then convert the signal’s anomaly degree into a credibility score. To address the strong temporal continuity and nonlinear characteristics of joint-angle signals, as well as the limited robustness of existing anomaly detection methods based on fixed thresholds, we propose an attention-enhanced Long Short-Term Memory (LSTM) autoencoder and a multi-strategy anomaly detection method. The LSTM autoencoder captures long-term dependencies in joint-angle time series, while a temporal attention mechanism highlights critical time steps. In the detection phase, we design four anomaly detection strategies to improve recall and robustness. Finally, the credibility score is computed based on the anomaly degree to support the reliable fusion of twin model driving signals. Experimental results demonstrate that the proposed method achieves a recall of 0.9965 and an F1-score of 0.9767 on the joint-angle sensor dataset, significantly outperforming other models. The method accurately detects anomalies to assess signal credibility and is further validated on control command signals, demonstrating the model’s strong robustness.</p>