32 Prediction of Width and Thickness of Injection Molded …
459
Liau, Lee (2018), on the other hand, explain that simulation should not be based
only on the historical data, but needs to include current process data to make a meaningful decision. Their solution is a framework for the development of a digital twin
for injection molding that models the entire process and allows bidirectional control
of the physical process. In (Guo 2012) fractional factorial design of experiments was
used to screen some of the injection molding process parameters obtained from the
process simulation in the MoldFlow™ software. Later the parameters were used to
create a mathematical model with the help of central composite design and finite
element simulation to predict warpage during the plastic injection molding.
Hardware solutions
Some other researchers use hardware development to enhance control and prediction
of the outcomes of the IM process. For example, in (Panchal and Kazmer 2010) the
authors propose a button cell type in-mold shrinkage sensor. The sensor’s performance was validated and compared to traditional shrinkage prediction methods.
“The sensor signals acquired during each molding cycle were analyzed to validate
the sensor performance in a design of experiments as a function of packing pressure,
melt temperature, cooling time, and coolant temperature” (Panchal and Kazmer
2010). In (Johnston 2015) an auxiliary process controller for online multivariate
optimization of the injection molding process is proposed. The objective function
includes terms related to process variation and energy control.
32.3 Experimental Setup
The research described in this paper includes the following steps:
Step 1: Process data collection. To collect the data design of experiment has
been used, namely Latin Hypercube sampling technique (Seaholm et al. 1988). 32
combinations of holding pressure, holding pressure time, backpressure, cooling time,
injection speed, screw speed, barrel temperature, and mold temperature process
parameters were included in the DOE. Each of these combinations has been launched
five times, resulting in 160 machine runs in total and 320 HDPE dogbone specimens,
since during each run two of them are produced. This should be sufficient for the
application of machine learning algorithms, as the minimum recommended number
of samples according to (Scikit-Learn 2018) is 50.
The use of DOE resulted in the following ranges for width and thickness [mm]
for the dogbone specimens 1 and 2: thickness1 ∈ [3.3, 3.84]; thickness2 ∈ [3.32,
3.84]; width1 ∈ [9.02, 9.95]; width2 ∈ [8.84, 9.96]. According to the CAD model
of the target parts, the width is 10 and the thickness is 4 mm. However, 3–3.5%
shrinkage rate is considered a regular shrinkage rate for the HDPE, thus resulting
in a width of 9.65–9.7 mm and thickness of 3.86–3.88. It is possible to say that the
use of the design of experiments allowed to vary the width and thickness of the parts
quite significantly covering values close to the best possible ones, as well as those
significantly lower.
459
Liau, Lee (2018), on the other hand, explain that simulation should not be based
only on the historical data, but needs to include current process data to make a meaningful decision. Their solution is a framework for the development of a digital twin
for injection molding that models the entire process and allows bidirectional control
of the physical process. In (Guo 2012) fractional factorial design of experiments was
used to screen some of the injection molding process parameters obtained from the
process simulation in the MoldFlow™ software. Later the parameters were used to
create a mathematical model with the help of central composite design and finite
element simulation to predict warpage during the plastic injection molding.
Hardware solutions
Some other researchers use hardware development to enhance control and prediction
of the outcomes of the IM process. For example, in (Panchal and Kazmer 2010) the
authors propose a button cell type in-mold shrinkage sensor. The sensor’s performance was validated and compared to traditional shrinkage prediction methods.
“The sensor signals acquired during each molding cycle were analyzed to validate
the sensor performance in a design of experiments as a function of packing pressure,
melt temperature, cooling time, and coolant temperature” (Panchal and Kazmer
2010). In (Johnston 2015) an auxiliary process controller for online multivariate
optimization of the injection molding process is proposed. The objective function
includes terms related to process variation and energy control.
32.3 Experimental Setup
The research described in this paper includes the following steps:
Step 1: Process data collection. To collect the data design of experiment has
been used, namely Latin Hypercube sampling technique (Seaholm et al. 1988). 32
combinations of holding pressure, holding pressure time, backpressure, cooling time,
injection speed, screw speed, barrel temperature, and mold temperature process
parameters were included in the DOE. Each of these combinations has been launched
five times, resulting in 160 machine runs in total and 320 HDPE dogbone specimens,
since during each run two of them are produced. This should be sufficient for the
application of machine learning algorithms, as the minimum recommended number
of samples according to (Scikit-Learn 2018) is 50.
The use of DOE resulted in the following ranges for width and thickness [mm]
for the dogbone specimens 1 and 2: thickness1 ∈ [3.3, 3.84]; thickness2 ∈ [3.32,
3.84]; width1 ∈ [9.02, 9.95]; width2 ∈ [8.84, 9.96]. According to the CAD model
of the target parts, the width is 10 and the thickness is 4 mm. However, 3–3.5%
shrinkage rate is considered a regular shrinkage rate for the HDPE, thus resulting
in a width of 9.65–9.7 mm and thickness of 3.86–3.88. It is possible to say that the
use of the design of experiments allowed to vary the width and thickness of the parts
quite significantly covering values close to the best possible ones, as well as those
significantly lower.