Object Recognition Technology Based On Sound Signals And Machine Learning

In order for intelligent robots to better understand and adapt to complex environments and complete complex tasks,the first task is to make them perceive and recognize themselves and their surroundings.Most of the existing object recognition technologies focus on visual perception and tactile perception,ignoring auditory perception.In fact,sound data can be used for object recognition in containers.In machine learning,it is assumed that the training sample set and the test sample set have the same category,that is,object recognition is performed in a closed environment.However,the intelligent robot is constantly in contact with the outside world,the types of objects it senses are increasing dynamically,and the amount of data it is sensing is also increasing.How an intelligent robot recognizes objects in an open environment(the training set and test set categories are different,and the test set categories are more than the training set categories)is crucial.And because the sound data collected at different times and different actions have different distributions,how to solve the data distribution differences,reduce data collection costs,and complete object recognition is equally important.Therefore,this paper uses supervised machine learning to verify the advantages of auditory information(sound information)collected in different actions in object recognition in a container,and proposes an Open Set Kernel k Nearest Neighbor(OSKk NN)in an open environment.Find a transfer learning method suitable for this article,which is used to solve the problem of difference in sound data distribution and complete object recognition.The object recognition process in this paper includes the construction of sound collection system,sound data collection,sound data preprocessing,sound feature extraction,establishment of sound data set,and object recognition based on different machine learning pairs.Taking the UR5 robot as the core,and configures it with an AG-95 manipulator.A condenser microphone is configured on the AG-95 manipulator.A sound acquisition system is set up using a Lenovo G480 laptop computer running a Matlab program as a sound acquisition and storage device.Through the Poly Scope user interface of the UR5 robot teach pendant,the UR5 robot is controlled to automatically complete different actions,collect sound data on the interaction between the manipulator,the container and the objects in the container,preprocess the sound data.The Mel-Frequency Cepstral Coefficients(MFCC)feature is used as a data set for object recognition.Four types of supervised machine learning algorithms: support vector machine(SVM),sparse representation classification(SRC),k-nearest neighbor(k NN),and kernel k-nearest neighbor(Kk NN)are used to perform experiments on sound data collected in different actions,and contrast recognition accuracy;also compared four different movements.It is verified that the action mode of horizontal rotation of 180 degrees is more suitable for UR5 robots to collect sound data.It has also been proven that when the action is small,the longer the contact process is,the more sound information is collected and the easier it is to recognize the object in the container.It is shown that the sound information collected by active action and the kernel k nearest neighbor algorithm can be used to recognize objects in the container,and it is proved that the sound collected by active action has a good advantage in object recognition in the container.Based on the Kk NN algorithm setting threshold T,a framework of Open Set Kernel k Nearest Neighbor(OSKk NN)in open environment is proposed,and the Open Set Sparse Representation Classification(OSSRC)algorithm is compared.It is proved that the use of the OSKk NN algorithm to separate known categories from unknown categories works well,and can classify and recognize known categories well.Only by separating the sample data of unknown categories,can we better collect these data and prepare for continued learning of these unknown categories.The experiments verify that the OSKk NN algorithm framework can well solve the problem of identifying objects in a container using sound in an open environment.After a large number of experiments,comparing 12 transfer learning methods and k NN algorithms,finding a metric transfer learning framework(MTLF-Metric Transfer Learning Framework)can solve the problem of the difference in sound data distribution caused by different action methods and different acquisition times in this article.MTLF combined with k NN algorithm has a better recognition effect on objects inside the container.