Quality Control of High Pressure Die Casting Using a Neural Network

1
Quality Control of High Pressure Die Casting
Using a Neural Network

• M.I. Khan, Y. Frayman, and S. Nahavandi
• Intelligent Systems Research Laboratory
• Deakin University
• Geelong VIC 3217, Australia
• Email mik, yfraym, nahavand_at_deakin.edu.au

2
Introduction

• Modern manufacturing processes are generally
  complex and require thorough understanding of the
  process behaviour in order to control them
  adequately.
• With an ever decreasing workforce entering shop
  floor, it is likely that the specialized
  knowledge about such complex industrial processes
  may eventually disappear.
• It is important to take measures to overcome this
  likely disappearance of this specialized
  knowledge.
• Computational intelligence techniques can help
  here if we can adequately model manufacturing
  processes and further develop a required control
  strategy.

3
High Pressure Die Casting

• High Pressure Die Casting (HPDC) is a
  manufacturing process that produces a range of
  consumer products from small mobile phones to
  large automotive engine castings.
• The process begins by injecting liquid metal in
  the die by a moving plunger.
• The metal is solidified under high pressure and
  the part is extracted, generally by a robot, when
  the die is opened.

4
Optimal Setting of HPDC

• The optimal settings of process parameters in
  HPDC is a difficult task due to complex
  inter-relationships between process variables.
• Sub-optimal process settings can result in
  producing castings with undesirable defects like
  blistering, cold shuts and porosity.
• In this work we a concerned with the porosity
  defect in castings which is the highest occurring
  defect in the automotive component casting
  industry.
• However, it is possible to apply the same
  framework for other casting defects.

5
Porosity Defect

• The appearance of pores in castings due to
  improper process parameters settings is called
  porosity.
• Two main types of porosity are gas (top figure)
  and shrinkage (bottom figure) porosity.
• Gas porosity is due to the gas being entrapped in
  the castings as result of air being injected in
  the die, the steam produced due to high
  temperatures and the lubricant burnt due to
  elevated temperatures.
• Shrinkage porosity is a result of metal losing
  its volume during shrinking, and hence more metal
  is required to fill the gaps (voids) produced. In
  HPDC, the application of pressure to fill the
  voids when metal is in a solidification state
  attempts to minimize the problem.

6
Modelling the HPDC Process

• The optimal settings of process parameters in
  HPDC are traditionally determined by using
  physical modelling techniques.
• These techniques have often resulted in
  conflicting results being reported by different
  authors.
• Some of these traditional approaches suffer from
  the problem that they tend to concentrate on the
  material properties while ignoring the influence
  of the process parameters of the process itself.
• Our hypothesis is that the HPDC process can be
  modelled more adequately by using inductive
  methods.

7
Neural Network Model

• Out of several available inductive methods, we
  have decided to use a feed-forward neural network
  model, because
• The data is available in input/output format is
  suitable for supervised learning.
• Offline learning is more feasible at the moment,
  but in the future it may be required to learn
  online and adaptively for a real time control
  (RTC). The feed forward neural networks can
  handle adequately both the offline and online
  learning.
• Speed of learning is important for RTC, which
  rules out other supervised learning methods like
  Boltzmann machines.
• The HPDC process is highly nonlinear, which rules
  out linear and other simpler regression methods.
• Feed forward neural networks have been used
  previously with considerable success in die
  casting.

8
Neural Network Model

• Neural network model in this work had four hidden
  nodes with tangent transfer function.
• Inputs consisted of eight process parameters
• 1st stage velocity
• 2nd stage velocity
• Changeover position
• Intensity of tip pressure
• Cavity pressure
• Squeeze tip pressure
• Squeeze cavity pressure
• Biscuit thickness.
• The outputs consisted of a four-level quality
  measure of the casting with 1 representing the
  best quality (minimum porosity) and 4
  representing the worst quality (maximum
  porosity).

9
Impact of Change-Over Position on the Level of
Porosity
10
Impact of Cavity Pressure on the Level of Porosity
11
Impact of Tip Pressure on the Level of Porosity
12
Results Summary
13
Discussion

• The obtained results show that a neural network
  model is able to predict the level of porosity
  adequately as the results are generally in
  agreement with that available in the literature.
• The conflict with the literature regarding the
  influence of the changeover position on the level
  of porosity is of no surprise, since the HPDC
  process is a very complex one and further
  research in needed to find out if the results
  obtained are correct.

14
Control Strategy

• The proposed control strategy utilizes the
  obtained neural network model of the influence of
  process parameters on the level of porosity.
• In the control framework the obtained neural
  network model is used to select between the
  different HPDC process settings to the one that
  is likely to result in castings with acceptable
  levels of porosity.

15
Control Strategy

• The process starts with the initial state and
  supplies its state (values of process parameters)
  to the neural network model.
• The neural network model then verifies the
  optimality of the process state by predicting the
  likely level of porosity in the castings.
• If the porosity level is predicted as acceptable,
  the neural network produces a signal to the
  hardware controller to maintain the current state
  of the process.
• Otherwise it notifies the controller that the
  process has to be switched to another one of the
  available settings which is then evaluated again
  and if found adequate is used.
• After every cycle of the casting production, the
  neural network can re-evaluate the likely effect
  of the particular process settings on the quality
  of the castings and generate required advise to
  the controller.