Language: 简体中文 English
Yunfeng Zhang
Yunfeng Zhang, Ph.D.
Professor,
Department of Mechanical Engineering,
National University of Singapore.
Email: mpezyf@nus.edu.sg

Biography:
Dr. Yunfeng Zhang received his B.Eng. in Mechanical Engineering from Shanghai Jiao Tong University, China in 1985 and Ph.D. from the University of Bath, UK in 1991. He is currently an Associate Professor at the Department of Mechanical Engineering, National University of Singapore. His research interests include (1) operations research, in particular, computational intelligence in design and manufacturing (process planning, scheduling, and their integration, VRP, and multi‐objective optimization for UAV mission planning); (2) hybrid manufacturing (3D printing and 5-axis machining) technology for parts repair. He has authored more than 200 publications and received various international awards including the Kayamori Best Paper Award in ICRA 1999 and the IMechE Thatcher Bros Prize in 2011.

Abstract:
Thermal Analyses for Laser Scanning Pattern Evaluation in Laser Aided Additive Manufacturing using Deep Neural Network

Laser aided additive manufacturing (LAAM) is one of the key metal 3D printing technologies for surface cladding and fabrication of near-net shape parts. Its laser scanning strategies can have significant effects on the temperature distribution for multi-bead, multi-layered additive manufacturing, resulting in distinct residual stress distribution and substrate deformation. However, existing thermal analyses models are time consuming and laser scanning pattern selection still relies on empirical experience. This paper proposed an efficient thermal analyses model for LAAM with Deep Neural Network (DNN), containing following parts:
Firstly, a novel finite element (FE) thermal analyses simulation architecture was designed to continuously predict the temperature field among the whole simulation domain during the entire LAAM process. It laid foundations for the large-scale thermal history data set generation by providing predicting temperature field in each unit simulation step.
Secondly, LAAM thermal analyses data structure was designed to describe the deposition state and corresponding temperature field. For a given simulation time t, a three-dimensional matrix st ∈ S was designed to describe the deposition status, and a three-dimensional matrix tt ∈ T1 was designed to describe the predicting temperature field, as shown in Fig 1. Thereafter, a one-layer LAAM deposition thermal history data set was built, containing over 20,000 deposition status.

Thirdly, a DNN model was designed to describe the connection between laser deposition status and corresponding temperature distribution with a function f: RS→RT1. The RNN model contained 5 fully connected layers, including 3 hidden layers with 200 neurons, as shown in the Fig 2. The rectified linear unit (ReLU) was selected as the activation function in order to speed up the training. The normalized mean square error (nMSE) was selected as the evaluation matrix and the parameters of the DNN was updated using stochastic gradient decent by backpropagating the error information at the output layer.

All training and validation are implemented with Tensorflow. The loss curve during the training phase is shown in Fig 3. It can be observed that the nMSE decrease from over 200,000 to around 7,000 after 1500 epochs, which implies that the DNN model fits the function f well. The comparison of simulation results and DNN model prediction results for a given 2D deposition geometry is shown in Fig 4. It can be found that both simulation results and DNN model prediction results show the comparable temperature concentration at the center part of the plate. The DNN model was validated to be able to predict the temperature field based on the deposition status.

COUNTDOWN
  • DAYS
  • HOURS
  • MINUTES
  • SECONDS
Key Dates
   Deadline for Submission of Abstract:
  October 31, 2018
   Notification of abstract acceptance:
   November 15, 2018