Anomaly identification and correction of multi-dimensional operating conditions for a data-driven vehicle noise and vibration platform based on fusion of deep learning models

With the proliferation of data-driven methods in automotive noise, vibration, and harshness (NVH) analysis, the digital transformation of NVH performance evaluation has become increasingly imperative. However, in the actual testing process, signals are inevitably affected by abnormal interference, resulting in a decrease in the evaluation accuracy of NVH performance. To solve this problem, we propose a new adaptive anomaly detection and correction framework. The key methodological innovation lies in the IResNet–IFCNN collaborative architecture, which introduces an improved residual network (IResNet) for high-precision anomaly identification and an improved fully connected network (IFCNN) for adaptive multi-condition correction. The main contribution is a closed-loop “detection – matching – correction” mechanism, which dynamically selects specific weights based on the type of anomaly. Verified on the real vehicle test data, the correction accuracy reached 93.33%, significantly enhancing the robustness of the intelligent NVH assessment under interference.