Applications Of Deep Neural Networks In Pattern Recognition

Pattern recognition is the basic intelligence of human and animals,and how to give computer the ability of pattern recognition is the key research direction of artificial intelligence.Neural network pattern recognition occupies a dominant position in the field of artificial intelligence.Compared with traditional pattern recognition methods,it shows superior performance in most pattern recognition application tasks.However,on the one hand,with the endless emergence of neural network model structure and the deepening of the number of layers,although the model has high accuracy,the model is also becoming larger and larger,and has higher requirements for computing resources.It is difficult to meet the real-time and lightweight requirements,especially for applications in real-time scenarios and mobile devices.On the other hand,in the process of application,the limitations of single mode data are becoming increasingly prominent.With the increase of data volume,data modes are also increasing.Making full use of multi-mode data can effectively improve the accuracy of identification results.Therefore,it is of great significance to explore the lightweight of network model and the fusion of multimodal data.In this paper,two application studies have been carried out.The main work is as follows:(1)In the field of mechanical fault pattern recognition,lightweight neural network is used for fault classification.Aiming at the disadvantages of large amount of calculation and high resource consumption of most neural network models,the lightweight neural network mobilenetv2 is simplified and applied to the fault pattern recognition of rolling bearings under variable speed conditions.Through the time-frequency analysis technology,the collected bearing original vibration signals are preprocessed to obtain a two-dimensional time-frequency diagram,and then input into the simplified mobilenetv2 network.Compared with the original17 reverse residual blocks,the simplified network only uses 3 residual blocks.While obtaining high recognition accuracy,the amount of calculation and parameters of the model is greatly reduced.The experiments show that compared with the traditional neural network model,the model has good effectiveness,lightweight and reliability.(2)In the field of emotion recognition,face emotion recognition and speech emotion recognition are the most studied.They use single modal data and fail to make full use of multiple modal data.In the algorithm,this paper combines multiple modal data and neural network for emotional pattern recognition.Firstly,aiming at the current situation and shortcomings of multimodal emotion data sets at home and abroad,a multimodal Chinese emotion data set including expression,action and voice is designed and recorded.The data set has a large number of people and a large sample size,which provides strong data support for emotional pattern recognition.Secondly,for each modal data,different neural networks are used for feature extraction and classification.Finally,the feature level fusion strategy is used to fuse the three modal data.The final experimental results show that the multi-modal data fusion method is higher than the recognition results of any modal.