| Intelligent manufacturing has become a key lever for industrialized countries to promote high-quality economic development.Intelligent fault detection and production line management are important components of intelligent manufacturing and also hot research topics nowadays.In production practice,abnormal sound detection is an important means of product fault detection.However,due to the lack of data sets,it is difficult to apply artificial intelligent models in specific abnormal sound detection tasks.Moreover,most industrial fault detection is limited to the "product-level",but any operation or equipment failure on the production line will hinder the production progress.Therefore,a suitable information-physical system is needed for comprehensive sensing and analysis of production elements to achieve"line-level" fault detection and intelligent management.In this regard,the following studies are conducted in this thesis.(1)An abnormal sound detection method based on Mel transform and recursive neural network is proposed to replace the traditional manual detection method.This method amplifies the spectral differences between different categories using Mel non-linear transformation,and uses spectral detail features and spectral envelope features as joint features to fully describe the spectral characteristics and realize data dimensionality reduction.Meanwhile,based on the characteristics of fault data,a bidirectional long short-term memory network is selected for feature mapping so as to improve classification efficiency and accuracy.Then experiments are built based on real air conditioner sound data to prove the effectiveness of the classification model.On this basis,the detection process is further complemented,and an abnormal sound detection interface is designed.(2)An Unreal Engine-based building method for digital twin of air conditioner production line is proposed,which introduces the powerful virtual creation ability and realistic physics engine of game development platform into the construction of twin system.The method includes three basic processes:modeling,driving,and feedback.In the process of object modeling and rule injection,the specific building steps and method flow based on Unreal Engine are given.Then in the process of model driving and control feedback,a distributed data transmission scheme and feedback control process is proposed,which improves the timeliness and safety of virtual-real interaction through data sharing and command coding and decoding.Finally,based on the digital twin system,a comprehensive visualization of the production line operation is realized,which improves the "transparency"of the workshop.The construction process and transmission control scheme proposed in this thesis have strong feasibility and significance.(3)A digital twin-based real-time detection method for air conditioner production line fault is proposed.The method uses the infinite virtual modeling space of the digital twin to solve the problem of production line modeling.And the method can help to quickly perceive and locate faults by detecting "virtual-real conflicts" during the parallel simulation,so as to achieve "production line level" fault detection.Moreover,a digital twin-based sample generation method for air conditioner production line failure is proposed to solve the problem of insufficient data support for intelligent production line management due to the scarcity of production line fault scenarios.The method fully utilizes the digital twin’s capabilities of simulating and self-evolving to proactively create fault scenarios,generating sufficient training samples for the network model of intelligent production line management.These methods demonstrate the importance of this digital twin system in air conditioner production line monitoring and intelligent production line management. |
