A Review on Machine Learning Models in Injection Molding Machines

Table 4

Supervised learning cases in injection molding.

Algorithm used	AI model taxonomy	Depth layer sizes, training time, and testing time	Dataset	Framework, core language, and interface	Advantages	Disadvantages	Ref.
Polynomial regression techniques	Supervised learning	NA	Simulation data, actual data	CADMOULD 3DF of SIMCON kunststofftechnische Software	The simulation predicts the direction of influence correctly	All factors’ interactions are not modeled properly	[64]
The artificial intelligence algorithm used is based on random forest	Supervised learning	Time of the molding press (approximately 55 seconds).	Data were collected from a worker for hours which were later tested by other datasets	Collaborative human-robot framework light interface	Cobot’s repeatability and precision, mixed with the operator’s flexibility, gives the capability to manage unpredictable conditions effortlessly	Cobot does not have bendy behavior; for that reason, an employee is forced to adapt to a cobot’s advent, coping along with its pace	[65]
Artificial neural network gradient descent optimization approach, which iteratively tries to reduce the mistake	Supervised learning	Cooling time, holding pressure level, injection time, mold temperature, holding pressure time	Large image datasets	The underlying neural networks adjust themselves from the simulation to the real data between the transfer learning phases	Covers the space between simulation and real statistics in the manufacturing process plan. Accurately predicts the quality criteria from the process parameter.	Requires continuous improvement in learning version by automatic triggering of a new experiment	[66]
Each dataset is trained on 7 classification models: medium tree, medium KNN, fine KNN, logistic regressions, and coarse KNN	Supervised	The model training utilizes the 10-fold cross-validation	Sensor data are extracted through an injection molding machine	Classification models are used. Python language is used to find the accuracy of the classification models.	Predicts the time when machine should be changed exactly which makes damage cost minimum	There is a lack of fault relating to the data	[63]
Naive Bayes classifier and classification by decision trees and SVM	Supervised, unsupervised	Ten-fold cross-validation method was used, with a complete database of two hundred and thirty-seven samples	A signal dataset was used with 152 correct and 85 faulty pieces	Classification models were used, and Python language was used	SVM predicted and detected the fabrication with good accuracy of the plastics part	Samples from the databases did not adapt to the distribution of Gaussian. The same good and lousy part examples are not present.	[67]
Support vector machine (SVM)	Supervised learning	NA	36 test series are used, with 80% used for training and 20% used for testing	NA	The quality components are produced from the first shot onwards	The recommendation and classifier are not optimized	[68]
The linear regression technique was used	Supervised learning	Cycle time was 45 seconds, including 15 seconds for packing time and 20 s for cooling time	Dataset consisted of various points in a part and a part of the weight	ML-based framework	Minimized the number of physical experiments. Predicted the quality characteristics of the injection molding processes about various process parameters’ values by utilizing the intelligence combination for the simulation data and measurement.	A considerable amount of data is required for training these models reasonably	[69]
6 various machine learning methodologies are utilized. The first three were based upon the decision tree: AdaBoost, ID3, and random forests. The other 3 were K-nearest neighbors (KNN), artificial neural network, and Naive Bayes	Supervised learning	The six classifiers are tested by using five-fold ten-time cross-validation.	Sensor dataset like pressure transducers, thermocouples, velocity, and flow sensors. Position sensors.	A human-machine graphical interface creation on Python by utilizing library PyQT graph and PyDaqmx along with standard Python libraries like NumPy	Improves overall performance according to the target productivity goal, for example, extending manufacturing equipment life and reducing area rejection, extending manufacturing equipment life, and improving energy management	Overall processing performance is not optimized	[70]
ARIMA model	Supervised learning	NA	Data consist of 100000 measurements for all machines and 53767 for machine 2605; Data are split into training and development sets using an 80:20 ratio	Used Python for graphical and predictive analysis	Models have relatively low MAPEs and hence can be considered as good	Verification of results was not present since there were no ground truths.	[71]