Table 13.

Acoustic FD methods applied to belt conveyor idlers via ML Models.

Authors	Detection Methods	Accuracy	Main Findings	Advantages	Disadvantages
Ericeira et al. [50]	The ultrasound sensing combined with RF and MLP machine learning techniques and signal pattern recognition.	Four experiments using different numbers of samples. The best result was achieved in the fourth experiment by extracting from the FFT applied in every 5 s of the 20 recordings divided into 40 parts the MLP10, which in one case attained 89.47% of correctly classified instances.	The results show that the detection performance depends on the features they use to input the classifier.	When frequency domain features are used with more data, the proposed methods demonstrate the best accuracy.	To achieve the highest degree of accuracy, it is necessary to tune the number of trees in RF, the number of neuron layers in MLP, etc.
Liu et al. [8]	MFCCs as features and GBDT algorithm for classification.	Detection accuracy of 94.53%.	The proposed MFCCs and GBDT approach is a viable method for detecting idler roll failures based on sound signals.	A Gradient Boost algorithm with self-learning automatically determines which MFCC feature to apply at which step and what threshold value to determine for this feature.	Window size for extracting MFCC significantly impacts the accuracy performance of GBDT models.
Rocha et al. [59]	For the detection of roller failures, a fast Fourier transform and means of the magnitude of the sound signal are used along with a random forest algorithm.	The trained model has an accuracy of 95% in identifying damaged bearings noise correctly.	Test results demonstrated that ROSI could stand up to harsh operating conditions while carrying out all necessary inspection tasks in a mining site, establishing it as a disruptive solution for belt conveyor.	The RF is simple to use and shows better results.	An unblanched dataset was used to train the method.
Yang et al. [40]	DNN, DCNN, SVM and KNN.	The accuracy is 90% KNN. The detection accuracy of SVM is 91.9% DNN: The average accuracy is 94.4% DCNN: The classification accuracy is 98%.	Based on the results, the fault detection system works very well for roller fault detection, with an accuracy rate of more than 90.0%.	Models based on deep learning produce better results than traditional models.	Developing a deep learning model requires a large amount of data to train and build a robust model.
Peng et al. [24]	Wavelet packet transformation and CNN have been used.	The classification accuracy rate of the mean as a feature is 86%, and the classification accuracy rate of the standard deviation as a feature is 93%.	According to the experiment results, using the standard deviation as the data feature is more effective than the mean in detecting roller faults.	A CNN can handle a large amount of input data and take into account the location information between the data.	After extracting wavelet packet transformation, data from the lowest frequency band can significantly affect CNN performance.
Xiao-ping Jiang and Guan qiang Cao [41]	Wavelet transform and Neural network	The accuracy rate can reach more than 96%.	In belt conveyors, fault characteristics are contained in fault sounds and can be obtained by adding the energy of each band after the wavelet transform has been applied.	The neural network can easily recognize and classify faults.	The accuracy of the method is affected by environmental noise and the sound of the belt conveyor.