| In recent years,industrial wireless sensor networks and advanced manufacturing equipment have reached a new stage of mutual integration,and successful applications have emerged in various industries,including the electronics manufacturing service(EMS)industry,laying a solid foundation for their wide application and boosting industry confidence.Industrial wireless sensor network(IWSN)for intelligent plant monitoring,the key and fundamental technology of factory digitalization,is the prerequisite and guarantee of the realization of intelligent manufacturing-a trending research topic in academia and industry.The research on industrial wireless sensor networks is still in the early stage,and there is no agreed unified standard protocol.The International Association of Automation Industries divides industrial applications into three levels,from low to high,namely,monitoring,control,and production safety,according to network performance requirements.The current wireless sensor networks are mainly used at the monitoring level with relatively low requirements network performance,while there are still many challenges in high-level applications such as high-real-time,highreliability monitoring,and production safety for intelligent plant monitoring.There are a series of scientific problems that need to be solved.Intelligent plant monitoring can be complex due to software and hardware equipment being developed by different manufacturers at different times and following different standards.They were developed independently without considering mutual interaction between systems or platforms.Although the equipment itself has been digitized,much information still exists in isolation due to its closeness and exclusivity.In order to maintain stability,continuity,and reliability,it is necessary to obtain and transmit the plant monitoring data and to realize the digitization,networking,and visualization of the production line without changing the status quo of the production line.Since the plant has already been deployed with multiple wireless systems,achieving highly reliable realtime transmission for plant monitoring with limited resources and considering the networking,energy supply,and maintenance is a huge challenge for such a dense IWSN.Concerning the cognitive industrial wireless sensor networks for intelligent plant monitoring,based on the analysis of the advancement of cognitive industrial wireless sensor network technology and combined with the existing research results,targeting the emerging challenges regarding innovation on theory considered key technologies in realizing digitalization,networking,and visualization of the production line,research in this dissertation mainly carries out theoretical demonstration,related technology research,and experimental verification from the following aspects:(1)Architecture for cognitive industrial wireless sensor network and interactive fusion of multi-source heterogeneous data;(2)Highly reliable real-time transmission scheme of massive multi-source data;(3)Coverage optimization problem in energy supply network for large-scale dense RFpowered network;(4)Performance evaluation of cognitive industrial wireless sensor networks under the premise of reliable communication;(5)Secondary network throughput analysis under jamming attacks;(6)Data analysis of the industrial wireless sensor network for intelligent plant monitoring.A noise-tolerance visual fault detection framework with learning from crowds.The research mainly touches upon the following aspects:1.A production line-oriented architecture for intelligent plant monitoring with CIWSN is designed.Based on an in-depth analysis of the life cycle characteristics and operating modes of the EMS industry regarding the communication needs of plant management,an augmented Open Platform Communication Unified Architecture(OPC UA)using Message Queuing Telemetry Transport(MQTT)is proposed.Therefore,it can be set as the standard communication protocol for plant monitoring applications.Research of interactive fusion in multi-source heterogeneous data is carried out based on the architecture,and with the building of a novel CIWSN architecture,the interconnection framework is clarified and set between the network layers and device,thus providing data transmission,scheduling,and fusion for the CIWSN in supporting the implementation of real-time plant monitoring and control.2.A novel real-time data transmission and scheduling method for massive multisource heterogeneous data is designed.Concerning the massive data collection problem,the modifications of the Media Access Control(MAC))layer in IEEE802.15.4 protocol have been proposed to give an idea of the Transmit Only(TO)network standardization.With multiple collectors to collect data from the transmitters and forward it to the receiver,the reliability of data delivery comes from the rational deployment of multiple collectors in different spaces that enhance the capture effect.TO transfers the methods of resolving channel conflicts from the sending end to the receiving end,increasing the accessible transmitters,and network throughput,achieving the paramount concerns of large-scale dense networking and network expansion from the underlying network.3.The coverage optimization problem(COP)is studied concerning the coverage of the wireless radio frequency(RF)power supply net.An energy supply method based on RF wireless charging is proposed,which not only simplifies wiring and improves the flexibility of node deployment but also simplifies the need for later maintenance to tackle the energy supply of large-scale network devices.An improved hybrid heuristic algorithm and a hybrid power bank deployment model based on real three-dimensional geographic information are proposed,which integrates the physical obstacles and the interference information of the plant into geographic details to optimize the number of nodes and their deployment location for optimal coverage.4.With stochastic geometry,the network throughput of the secondary network in harsh environments under jamming attacks is formed as a Markov decision process(MDP)while introducing model-free reinforcement learning techniques to evaluate the secondary network,proving the feasibility of deploying a secondary network and the reliability of implementing a secondary network.5.Based on the application of data generated from the cognitive industrial wireless sensor network for intelligent plant monitoring,a visual fault detection framework learned from plant monitoring data is proposed and implemented.A method for learning from crowds with contrastive representation(Crowd Cons)is proposed to obtain reliable sample labels from noisy multi-source sensor data to train robust vision models.Specifically,a pre-trained model is firstly used to extract visual features,high-confidence instances,and positive instance pairs in multi-source crowd-supervised information obtained,and then the supervised contrastive learning loss is extended to obtain a noisetolerant version that supports continuous label consistency.The parameters of each sensor and the overall fault detection model are trained in an end-to-end paradigm.The proposed Crowd Cons is compatible with other existing learning from crowds methods.The method is evaluated on four real datasets,including Label Me,CIFAR10-H,Music,and Ja Val.Experimental results indicate that Crowd Cons can significantly improve performance,and the case study demonstrates the robustness of noisy labeling.And the performance on four datasets verifies the practicability and scalability. |
PDF Full Text Request