Research And Embedded Implementation Of Deep LightGBM Algorithm For Sound Classification

Sound classification is an important branch in the field of machine learning,which is generally subdivided into three major categories: ambient sound classification,human voice classification,and music classification.In recent years,with the holding of competitions such as DCASE,more and more scholars have also started to pay attention to the research in this field.This technology is now widely used in scenarios such as medical diagnosis,scene analysis,vocalist recognition,and ecological environment analysis.Traditional sound classification methods are mainly implemented using neural networks,and although their accuracy keeps improving,this approach currently has two problems.The first is in the pre-processing of data.Sound classification tasks using neural networks generally involve extracting sound files into sound spectrograms first,thus converting the sound classification task into picture classification.In this way,when dealing with large sample data sets,the converted image data set is very large,thus requiring huge storage space as well as high computational performance to handle.Second,training with neural networks is prone to overfitting without enough data and requires a lot of parameter adjustment and model setup,which are very tedious steps leading to a very high time cost to obtain the best result model.Based on the above two shortcomings,this paper proposes a new model Deep LightGBM,which is an improved LightGBM deep learning model based on the idea of deep forest,effectively improving the classification accuracy and generalization ability,while ensuring the simplicity of the model and reducing the dependence of the algorithm on parameters,thus significantly reducing the time cost of training the model.And because the vector method is used to extract data features,it not only reduces the required storage space,but also speeds up the training of the model.The algorithm is validated on the publicly available environmental sound classification dataset Urban Sound8 K,and the new model achieves an accuracy of 95.84% when the vector method is used to extract sound features.When the features extracted by CNN were fused with the vector method features and then trained with the new model,the accuracy reached 97.67%.The experimental results show that a suitable sound feature extraction method with Deep LightGBM model parameters is easy to adjust,highly accurate and does not produce overfitting.The generalization performance was verified on several other types of sound classification datasets,and the classification accuracy performed very well.At the same time,the trained model is migrated on the embedded platform ZCU102,and the model is optimized on the PS side,implemented in C++ with targeted code improvements,so that the final prediction algorithm runs in less time than the PC and the classification accuracy remains consistent.On the PL side,we designed the acceleration IP based on Vitis and Vivado software for reading and predicting of LightGBM model,a key step in the algorithm implementation,and also designed a high-speed AXI communication module for PS and PL communication.