Review of anomalous sound event detection approaches

Abstract

<p>This paper presents a review of anomalous sound event detection(SED) approaches. SED is becoming more applicable for real-world appliactaions such as security, fire determination or olther emergency alarms. Despite many research outcome previously, further research is required to reduce false positives and improve accurracy. SED approaches are comprehensively organized by methods covering system pipeline components of acoustic descriptors, classification engine, and decision finalization method. The review compares multiple approaches that is applied on a specific dataset. Security relies on anomalous events in order to prevent it one must find these anomalous events. Audio surveillance has become more efficient as that artificial intelligence has stepped up the game. Autonomous SED could be used for early detection and prevention. It is found that the state of the art method viable used in SED using features of log-mel energies in convolutional recurrent neural network(CRNN) with long short term memory(LSTM) with a verification step of thresholding has obtained 93.1% F1 score and 0.1307 ER. It is found that feature extraction of log mel energies are highly reliable method showing promising results on multiple experiments.</p>

A survey on predicting oil spills by studying its causes using deep learning techniques

“…In [51] the analysis matches previously observed responses to irregular sound event detection (SED) on the structure of the characteristic SED pipeline ingredients. The system's performance on a lone and multi-class base was estimated.…”

An intelligent oil accident predicting and classifying system using deep learning techniques

“…The number of epochs has played a big role in the result, as with this data, 2000 epochs were needed to obtain that result. Also, 10000 batch-sized was used in this model as demonstrated in Figures 5,6,7,8,9,10,11,12,13,and 14 show different results for our comparison of deep learning techniques. The SoftMax activation function was necessary to use in the last layer as our model predicts and classifies nine classes, and it can't predict multi-classification without that activation function.…”

“…The investigation in [14] compares the structure of the typical sound event detection (SED) pipeline constituents to previously documented responses to irregular SED. On an isolated and multi-class basis, the system's performance was estimated.…”

A sound event detection based on hybrid convolution neural network and random forest

“…The researchers used CNN-long short-term memory (CRNN-LSTM) to obtain 93% F1 score [3]. CNN obtained 91% [4], multilayer layer perceptron (MLP-CNN) with 84% [5] and ensemble with 78% [2]. Although these experiments used artificial datasets, they have demonstrated how SED is feasible for industrial applications.…”