466
O. Ogorodnyk et al.
All these algorithms require certain tuning to receive a model of the highest quality.
For kNN it is important to find the best matching number of nearest neighbors, multilayer perceptron requires choosing the proper number of layers, neurons, selecting
the learning rate, etc. In the case of decision trees, the minimum number of instances
in a leaf is an important parameter, while for the random forest choosing the correct
number of trees is crucial. However, out of the methods tested in this study MLP is
probably the hardest to tune.
When it comes to the interpretability of the chosen methods, it is only the decision
trees algorithm that is easily interpretable for a human. Models obtained through the
use of the other algorithms are not as easy to interpret, especially the multilayered
perceptron and the random forest.
The reliability of the proposed methodology can depend on the quality of data,
computational tools used to induce the methods, as well as the correct interpretation
of the results. The data used in this study was obtained through the above-mentioned
laboratory experiment. The experiment was designed and planned in a way that would
exclude the possibility of getting missing or erroneous data. Unfortunately, that might
not always be the case, especially when the data is obtained from manufacturing
companies. As mentioned before, the use of DOE helped to vary dimensions of the
produced parts quite significantly. This allowed obtaining a dataset that includes
dimension values close to the ideal ones, as well as those remarkably lower.
Use of different knowledge discovery platform or application of variation of
any machine learning algorithm presented might lead to receiving results that are
slightly different from the ones presented in this study. The ability to correctly interpret obtained results is also crucial, as a misunderstanding of certain model quality
measures might lead to missing important aspects related to the model’s performance
capabilities.
The models presented in this study are capable of predicting dimensions of only the
dogbone part since the dataset they are trained on doesn’t include samples related
to any other products. To create models for the prediction of dimensions of other
products, the corresponding dataset needs to be obtained at first. At the same time, it is
hard to say how accurate will the models’ predictions be if the mold for production of
the dogbones is changed. However, it is possible to assume that if mold has properties
similar to the one used in the experiment, the models’ predictions should be of similar
quality. Such an approach could be a good starting point for manufacturing industries,
including thermoplastics injection molding, to start analyzing and interpreting their
data.
Amounts of data logged through manufacturing execution systems, enterprise
resource planning systems, etc. continue to grow. There might be interesting and
unrecognized at the moment patterns, that could potentially improve the quality of
manufactured products, the overall flow of the production process, decrease amounts
of produced scrap and energy consumption.
Future work includes obtaining more data, possibly for products of different geometry and material and testing machine learning methods on larger amounts of data.
In addition to the real production or experimental data, simulated data can be used
to pre-train the prediction models for further use of transfer -learning, as shown in
O. Ogorodnyk et al.
All these algorithms require certain tuning to receive a model of the highest quality.
For kNN it is important to find the best matching number of nearest neighbors, multilayer perceptron requires choosing the proper number of layers, neurons, selecting
the learning rate, etc. In the case of decision trees, the minimum number of instances
in a leaf is an important parameter, while for the random forest choosing the correct
number of trees is crucial. However, out of the methods tested in this study MLP is
probably the hardest to tune.
When it comes to the interpretability of the chosen methods, it is only the decision
trees algorithm that is easily interpretable for a human. Models obtained through the
use of the other algorithms are not as easy to interpret, especially the multilayered
perceptron and the random forest.
The reliability of the proposed methodology can depend on the quality of data,
computational tools used to induce the methods, as well as the correct interpretation
of the results. The data used in this study was obtained through the above-mentioned
laboratory experiment. The experiment was designed and planned in a way that would
exclude the possibility of getting missing or erroneous data. Unfortunately, that might
not always be the case, especially when the data is obtained from manufacturing
companies. As mentioned before, the use of DOE helped to vary dimensions of the
produced parts quite significantly. This allowed obtaining a dataset that includes
dimension values close to the ideal ones, as well as those remarkably lower.
Use of different knowledge discovery platform or application of variation of
any machine learning algorithm presented might lead to receiving results that are
slightly different from the ones presented in this study. The ability to correctly interpret obtained results is also crucial, as a misunderstanding of certain model quality
measures might lead to missing important aspects related to the model’s performance
capabilities.
The models presented in this study are capable of predicting dimensions of only the
dogbone part since the dataset they are trained on doesn’t include samples related
to any other products. To create models for the prediction of dimensions of other
products, the corresponding dataset needs to be obtained at first. At the same time, it is
hard to say how accurate will the models’ predictions be if the mold for production of
the dogbones is changed. However, it is possible to assume that if mold has properties
similar to the one used in the experiment, the models’ predictions should be of similar
quality. Such an approach could be a good starting point for manufacturing industries,
including thermoplastics injection molding, to start analyzing and interpreting their
data.
Amounts of data logged through manufacturing execution systems, enterprise
resource planning systems, etc. continue to grow. There might be interesting and
unrecognized at the moment patterns, that could potentially improve the quality of
manufactured products, the overall flow of the production process, decrease amounts
of produced scrap and energy consumption.
Future work includes obtaining more data, possibly for products of different geometry and material and testing machine learning methods on larger amounts of data.
In addition to the real production or experimental data, simulated data can be used
to pre-train the prediction models for further use of transfer -learning, as shown in