Physics-constrained data-driven prediction and multi-objective optimization of road noise in electric vehicles.

The overall interior noise level of pure electric vehicles is relatively low, making road noise a dominant factor affecting in-cabin ride quality. Mid- and low-frequency structure-borne road noise mainly arises from the coupled vibration of the tire-suspension system and exhibits multiple spectral peaks, strong nonlinearity, and high sensitivity to structural parameters, posing challenges for accurate prediction and optimization. Conventional experimental and simulation-based methods are applicable to road noise analysis but often involve complex modeling procedures and limited efficiency. To address these limitations, innovations are introduced in both model architecture and physical consistency. First, a road-noise prediction model based on a multi-head attention mechanism is proposed. Key structural parameters, including tire properties, suspension bushings, and dampers, are used as inputs, while the multi-head attention structure enhances feature extraction and weight allocation across different frequency bands, thereby improving the representation accuracy and generalization capability for multi-peak and nonlinear road noise. Second, a physics-constrained learning framework is established by embedding amplitude-frequency response constraints into the loss function, ensuring that the learning process follows vibro-acoustic transmission laws. This strategy reduces prediction bias at typical peak frequencies and improves model interpretability. Based on these strategies, a physics-constrained multi-head attention long short-term memory model is constructed and integrated with NSGA-III to achieve multi-objective collaborative optimization of road noise. Full-vehicle road tests are conducted for validation.