Jiayi Zhou
The subjective perception of noise quality in the cockpit of construction machinery is jointly determined by structural vibration and the radiated sound field, and the traditional evaluation method based on a single physical quantity finds it difficult to accurately describe its multi‑dimensional coupling characteristics. This paper presents a sound quality evaluation model and optimization design method based on multi‑modal sensing. Firstly, a multi‑source acoustic and vibration synchronous acquisition test was designed, and psychoacoustic parameters, variational mode decomposition energy characteristics, and Gammatone cepstral coefficients were extracted to form acoustic features. The transmission path contribution rates and structural mode parameters were combined to form vibration features, and a multi‑mode description system containing 102‑dimensional original features was constructed. Secondly, a multi‑head attention mechanism is introduced to achieve deep feature fusion, which is reduced to 38 dimensions after screening by the maximum‑correlation minimum‑redundancy (mRMR) criterion. On this basis, an improved Transformer evaluation model combining physical prior embedding of the transmission path, two‑headed attention decoupling, and frequency‑time‑domain joint attention mechanism is proposed to predict the annoyance and preference scores. Furthermore, the feature contributions were quantified based on the SHAP method to locate the short board of acoustic quality, and the Gaussian process regression proxy model was constructed to replace the finite element simulation. The improved NSGA‑III algorithm was used to achieve multi‑objective optimization of plate thickness, suspension stiffness, and reinforcement layout. Taking loaders, rollers, and excavators as validation objects, the results show that the mean absolute error of the evaluation model on 1,440 independent test samples is 0.31-0.40, which is about 50% lower than that of the traditional psychoacoustic model. After optimization, the comprehensive improvement index reaches 0.52-0.91, and the prediction deviation of the proxy model is controlled within 0.30. This study provides data-driven methodology support for forward design and rapid evaluation of sound quality of construction machinery.