Table 4.

A summary of ML models in idlers detection using vibration and acoustic signals.

Models	Strengths	Limitations
Random Forest (RF)	RF algorithm is capable of handling large data sets with a high degree of dimensionality. RF produces good predictions that are easy to understand.	RF is primarily limited by the fact that the algorithm becomes too slow and inefficient for real-time predictions if there are a large number of trees.
K-nearest neighbors algorithm (KNN)	During the process of making real-time predictions, it does not have a training period and only makes use of the training dataset while making predictions in the future. KNN is easy to implement since only two parameters are required, namely K and the distance function.	It does not perform well when dealing with large datasets. It is not suitable for large dimensions.
Support Vector Machine (SVM)	It is more effective to use SVM in high-dimensional spaces. There is a relatively low memory requirement for SVM.	It is not recommended to use SVM algorithms for large datasets. This is due to the high training complexity of SVMs.
Gradient boost decision tree (GBDT)	There is a great deal of flexibility—it can be optimized on various loss functions and has several options for tuning hyperparameters to make the function fit as flexibly as possible. It is able to predict faults with a high degree of accuracy.	The computational cost of GBMs is high as they often require many trees (more than 1000), which takes a great deal of time and memory.
Convolutional Neural Network (CNN)	It requires little pre-processing, reducing the human effort required to develop its functionalities. A good performance was achieved when extracting local features from images.	There is a need for a large amount of training data. Training is computationally intensive.
Isolation Forest (IForest)	A small sample size is more effective. A low memory requirement and minimal computational effort. Scalable to handle extremely large data sets and multidimensional problems with a large number of irrelevant attributes.	Has difficulty finding anomalous items that are closely surrounded by ordinary items. Does not perform well when an ordinary item is in close proximity to an anomalous item.
Autoencoder (AE)	A great tool for extracting features. Suitable for representing highly complex and nonlinear patterns.	Rather than capturing as much relevant information as possible, learns to capture as much information as possible.
Multilayer perceptron (MLP)	It is applicable to complex nonlinear problems. Ability to handle large amounts of data. Predict after training on time.	Model performance depends on the quality of the training data.