A novel approach for evaluating the material of moving objects inside a sealed cavity based on model transfer

The material information is of significant reference value for analysing the fault sources in industrial equipment and for targeted improvements in the production process. When a movable object inside a sealed chamber collides with the inner wall, it generates a large number of weak acoustic signals containing material information. Due to the scarcity of material data, traditional material assessment models often suffer from overfitting and poor generalisation ability. To address this issue, this paper presents a material evaluation method for moving objects within sealed chambers using model transfer. By applying model transfer, the neural network's material assessment knowledge from the sourcedomain data is utilised in the target domain. Initially, a deep learning model is trained on the source-domain dataset, leveraging the network's internal layers to extract feature representations of the material dataset. This trained model is then adapted to the target domain by freezing its layers and substituting the fully connected (FC) layers with XGBoost. Subsequently, target-domain data is employed to fine-tune the model, enabling material assessment for small samples through the optimised model. Experimental results demonstrate that the proposed method enhances the accuracy of material assessment under limited sample conditions and improves the capability of the model to evaluate the material of unfamiliar devices. The proposed algorithm demonstrates better accuracy and generalisation in material evaluation compared to other common machine learning methods, achieving an evaluation accuracy approximately 23% higher than the random forest algorithm.