| Cloud manufacturing is a new service-oriented networked manufacturing model that has been a hot topic of research within the manufacturing industry since it was proposed.Cloud manufacturing is realized through a networked service platform,which is dedicated to better integrate manufacturing resources and provide them to the demanders.However,with the increase of manufacturing resources,demanders usually need to spend a lot of time and energy to select the desired manufacturing services,which leads to low utilization of manufacturing resources,inactive users and other problems,and hinders the development and promotion of cloud manufacturing.Therefore,in order to help demanders more accurately filter through the massive manufacturing services and tap their potential demand points quickly,this thesis investigates the manufacturing service recommendation method in cloud manufacturing scenario from two perspectives of user behavior and time series,as follows:(1)The composition and service model of the cloud manufacturing platform are described.The missing feature problem of manufacturing service category attribute labels and user rating data of experimental data is analyzed.To address the missing manufacturing service attribute labels,this thesis models the data from three dimensions: the industry to which they belong,the manufacturing material and the manufacturing process,and simulates the user ratings using the method of quantified QoS indicators,which lays the foundation for the subsequent research and implementation of the recommendation method.(2)By studying and comparing the current advanced deep learning recommendation methods in mainstream fields,a DIEN model with time-awareness is selected and built.Based on the analysis of demander behavior characteristics of cloud manufacturing scenarios,two model optimization methods based on temporal weights and oversampling are proposed.Experiments are conducted using the cloud manufacturing user behavior dataset,and the experimental results are analyzed in four index dimensions: AUC value,F1-Score,coverage rate and hit rate.The results show O-DIEN model performs best in the other three indexes except for the hit rate index,which has mediocre performance.The efficacy of the O-DIEN model is verified.(3)To address the problem of the mediocre hit rate index of O-DIEN model,this thesis propose a CNRM model,which first clusters users,then calculates the recent popularity of manufacturing services within the clusters and eliminates the popularity bias,and then uses the bipartite graph method to describe the "degree" passed between the temporal neighbors.The recent popularity is then normalized with the "degree" passed in the bipartite graph to obtain the recommendation result.The experimental results show that the method has significantly improved the performance of coverage and hit rate indexes,and more demanders can be correctly recommended relevant manufacturing services,which proves the effectiveness and practical application value of the method.(4)The model fusion methods are compared,and the cross-fusion method is selected and improved to fuse the O-DIEN model and CNRM model,so that the fusion model can dynamically adjust the recommendation results according to the demander’s click behavior.In the online deployment stage,the fusion model is pre-trained by Spark distributed cluster to generate recommendation results and pre-stored in the online database,which shortens the time for the system to respond to the recommendation method;finally,the results of data stream acquisition and online deployment of the cloud manufacturing platform are shown,and the test results of user triage are tested in the actual application,and the test results show that the recommendation results of the fusion model are significantly better than those of the single model in terms of user click rate.The test results show that the recommendation results of the fusion model have a significant improvement in user click rate compared with the single model. |
PDF Full Text Request