Research And Application Of Campus Abnormal Sound Detection Method Based On WNMF-CNN

In recent years,cases of endangering campus safety and infringing on students’ rights and interests have been reported frequently.School bullying and violence have occurred from time to time,causing great harm to the parties and bringing very bad social impact.Although real-time monitoring can be achieved through the arrangement of video surveillance in the campus.However,the location of the time is often hidden or not suitable for the deployment of monitoring equipment,such as toilets and dormitories,and these places are exactly where the incident is high.Using more efficient and cheap audio monitoring methods to achieve a full range of abnormal event detection on campus has better feasibility for building a safe campus and protecting the physical and mental safety of teachers and students.With the development of artificial intelligence technology,abnormal events can be effectively determined by detecting environmental sounds,and audio information is easier to store and analyze,which can effectively find some events without obvious visual features.However,deep learning requires a large number of high-quality datasets as support,the collection cost is high and it depends on manual labeling,and there are complex and diverse noise interference in environmental sounds,which seriously reduce the detection performance of the model.Therefore,this thesis introduces the method of weighted nonnegative matrix factorization to improve the quality of the data set,constructs a multi-convolutional neural network model,and aims to accurately detect sound events.The main research methods are as follows:(1)Aiming at the problem that the existing environmental sound datasets are difficult to meet the training requirements of deep learning models,a method based on Weighted Nonnegative Matrix Factorization(WNMF)to improve the quality of the dataset was proposed.According to the frequency characteristics of noise and event,the weighted NMF of the weight matrix is calculated to highlight the frequency domain of the target event and reduce the interference of background noise.The time label is added to the weak label and unlabeled data set to provide data support for the subsequent use of deep learning to build the model.(2)Aiming at the problem of insufficient detection accuracy of existing abnormal sound detection models in complex environments,a method based on Multi Convolutional Neural Network(MCNN)was proposed for sound event detection.Deep convolutional neural network was used to predict frame labels.Shallow neural networks predict target event labels.The activation function and loss function are selected according to the characteristics of the model,and the adaptive median filter is used to eliminate the abnormal points in the frame label.Finally,the two models are combined to predict the output event prediction results of the label.Based on the above researches and practical application requirements for detecting abnormal sound events in the campus environment,a campus abnormal sound detection system with functions such as user management,device management,abnormal event detection,and abnormal event alarm will be implemented.