Anomalous sound event detection: A survey of machine learning based methods and applications

Mnasri, Zied; Rovetta, Stefano; Masulli, Francesco

doi:10.1007/s11042-021-11817-9

Cited by 32 publications

(16 citation statements)

References 116 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…Although the SELD(t) problem has been discussed at length in the literature and solutions have been proposed for both indoor (e.g. speaker identification [6], anomalous sound detection [7]) and outdoor applications (e.g. acoustic scene classification [8], traffic monitoring [7]), existing solutions do not cover these specific challenges of the automotive domain.…”

Section: Project Goalsmentioning

confidence: 99%

Real-Time Acoustic Perception for Automotive Applications

Yin¹,

Damiano²,

Verhelst³

et al. 2023

Preprint

View full text Add to dashboard Cite

Section: Project Goalsmentioning

confidence: 99%

Real-Time Acoustic Perception for Automotive Applications

Yin¹,

Damiano²,

Verhelst³

et al. 2023

Preprint

View full text Add to dashboard Cite

“…As defined by V. Chandola, A. Banerjee, and V. Kumar in 2009 [11], anomalies refer to the behaviour patterns that do not conform to a welldefined notion of normal behaviour. Based on the definition, the characteristics of the anomalies to differ them from the normal ones includes (1) the scarcity as the anomaly event occurs quite less frequently than normal ones; (2) the features which are able to be identified or extracted from the normal ones;(3) the meaning to be represented by the anomality [12]. Because of the characteristics of the anomalies, many algorithms, mainly unsupervised and semi-supervised, are created for the fault diagnose and anomaly prediction.…”

Section: Limitation Of Existing Solutionsmentioning

confidence: 99%

Untitled

2022

IJSC

View full text Add to dashboard Cite

This paper is to introduce a novel semi-supervised methodology, the enhanced incremental principal component analysis ("IPCA") based deep convolutional neural network autoencoder ("DCNN-AE) for Anomalous Sound Detection ("ASD") with high accuracy and computing efficiency. This hybrid methodology is to adopt Enhanced IPCA to reduce the dimensionality and then to use DCNN-AE to extract the features of the sample sound and detect the anomality. In this project, 228 sets of normal sounds and 100 sets of anomaly sounds of same machine are used for the experiments. And the sound files of machines (stepper motors) for the experiments are collected from a plant site. 50 random test cases are executed to evaluate the performance of the algorithm with AUC, PAUC, F measure and Accuracy Score. IPCA Based DCNN-AE shows high accuracy with the average AUC of 0.815793282, comparing with that of Kmeans++ of 0.499545351, of Incremental PCA based DBSCAN clustering of 0.636348073, of Incremental based PCA based One-class SVM of 0.506749433 and of DCGAN of 0.716528104. From the perspective of computing efficiency, because of the dimensions-reduction by the IPCA layer, the average execution time of the new methodology is 15 minutes in the CPU computing module of 2.3 GHz quad-core processors, comparing with that of DCGAN with 90 minutes in GPU computing module of 4 to 8 kernels.

show abstract

“…F1 is the evaluation index for the comprehensive evaluation of the accuracy and completeness, which is the summed average of the accuracy and completeness, and the formula of calculation method is as follows (10) where TP denotes the number of correctly classified positive class samples, TN denotes the number of correctly classified negative class samples, FP denotes the number of misclassified positive class samples, and FN denotes the number of misclassified negative class samples [35]. R2 is used to measure the ratio between the test result and the true value of the dependent variable, and its calculation formula is as follows…”

Section: Evaluation Metricsmentioning

confidence: 99%

A pipeline defect estimation method based on supervised learning fusion model

Li²,

Wang

et al. 2022

Preprint

View full text Add to dashboard Cite

Pipeline health assessment is an important work in industry, and information on the type and size of defects is an essential basis for assessing the health of a pipeline. Therefore, a pipeline defect estimation method based on supervised learning ensemble model is proposed in this paper. Firstly, several typical feature factors are calculated using feature formulas in the field of acoustics, capable of distinguishing the defect signal variability. Thereafter, Pearson correlation coefficient analysis and Random Forest importance ranking feature analysis methods are utilized to filter out the more valuable features. To improve the performance of defect estimation, a fusion model combining qualitative and quantitative analysis based on Random Forest and XGBoost is constructed to preferentially identify the type of defect signal qualitatively and then predict the size quantitatively. Finally, experimental results and comprehensive analysis with other mainstream supervised learning methods indicate that the prediction error of this method is basically below 1.5%, which addresses the issue of the low estimation accuracy of traditional methods.

show abstract

Anomalous sound event detection: A survey of machine learning based methods and applications

Cited by 32 publications

References 116 publications

Real-Time Acoustic Perception for Automotive Applications

Real-Time Acoustic Perception for Automotive Applications

Untitled

A pipeline defect estimation method based on supervised learning fusion model

Contact Info

Product

Resources

About