Research On Key Technologies Of Production Line Workers’ Production Operation Identification System Based On Deep Learning

Smart manufacturing is an inevitable stage on the development path of informationization and digitalization of each manufacturing enterprise nowadays.At the same time,smart manufacturing is also a new manufacturing model based on the significant digitalization of manufacturing enterprises’ factories,using Internet of Things technology,artificial intelligence technology and equipment monitoring technology to strengthen the production process information management and services.At this stage of the factory intelligence process,the substantial replacement of manual participation is difficult to be achieved overnight,so the use of video surveillance to control the production operation behavior is gradually becoming the mainstream development trend.At present,manufacturing enterprises with a certain scale have arranged a large number of surveillance cameras in the wave of informationization of their factories,and these surveillance cameras generate tens of thousands of video data of workers’ production operations every day,but it is difficult to use them effectively.Therefore,this paper analyzes and researches worker production operation behavior recognition from the perspective of action recognition,and mainly carries out the following work.Firstly,in terms of video data structuring,the target recognition network structure is optimized and the recognition result data structure is designed.Restricted to the lower arithmetic power of practical application scenarios,this paper modifies the classical YOLO network structure to significantly reduce the computing resource requirements.Combining the actual work scenarios and production operation characteristics of workers,we customize the Alpha Pose human posture recognition network skeletal loci recognition content to reduce the risk of network misidentification and improve the accuracy of subsequent action recognition.Finally,we produce two specific production action datasets based on a large amount of field data collection,and test the above optimized algorithm network on this dataset.Secondly,in terms of action recognition algorithm,the method and process of human skeleton-object graph data construction are proposed,i.e.,the target recognition results out the human body and production tool information,the human target information obtains the skeleton information results by the gesture recognition network,and the production tool information and skeleton information are used as graph data points for graph data construction.After that,ST-GCN neural network is designed to recognize the production operation behavior of workers.Finally,experiments are designed and implemented on the self-produced dataset for worker production operation recognition effects to obtain the performance results of the recognition method proposed in this paper.In this study,through the recognition of production operation behaviors of workers at production sites,on the one hand,it is beneficial to the management of production operations,the statistics of operation information and early warning of operation violations,etc.,to improve product quality;on the other hand,long-term data collection can assist in the design of production operation actions and production process improvement to enhance production efficiency.