Deep autoencoders for acoustic anomaly detection: experiments with working machine and in-vehicle audio

Abstract

The growing usage of digital microphones has generated an increased interest in the topic of Acoustic Anomaly Detection (AAD). Indeed, there are several real-world AAD application domains, including working machines and in-vehicle intelligence (the main target of this research project). This paper introduces three deep AutoEncoders (AE) for unsupervised AAD tasks, namely a Dense AE, a Convolutional Neural Network (CNN) AE and Long Short-Term Memory Autoencoder (LSTM) AE. To tune the deep learning architectures, development data was adopted from public domain audio datasets related with working machines. A large set of computational experiments was held, showing that the three proposed deep autoencoders, when combined with a melspectrogram sound preprocessing, are quite competitive and outperform a recently proposed AE baseline. Next, on a second experimental stage, aiming to address the final in-vehicle passenger safety goal, the three AEs were adapted to learn from in-vehicle normal audio, assuming three realistic scenarios that were generated by a Deep Autoencoders for Acoustic Anomaly Detection synthetic audio mixture tool. In general, a high quality AAD discrimination was obtained: working machine data -72% to 91%; and in-vehicle audio -78% to 81%. In conjunction with an automotive company, an in-vehicle AAD intelligent system prototype was further developed, aiming to test a selected model (LSTM AE) during a pilot demonstration event that targeted the cough anomaly. Interesting results were obtained, with the AAD system presenting a high cough classification accuracy (e.g., 100% for front seat locations).

AutoOC: Automated multi-objective design of deep autoencoders and one-class classifiers using grammatical evolution

“…where l denotes the length of the predicted data (e.g., l = m for a validation set) and T h ∈ [0, 1] is a threshold decision value, allowing to interpret the predicted anomaly class as positive if S i > T h. The ROC curve plots the False Positive Rate (FPR) versus the True Positive Rate (TPR) for all threshold values. AUC is a popular binary classification measure of performance, providing two main advantages [39]. Firstly, quality values are not influenced by the presence of unbalanced data, which occurs in OCC tasks.…”

“…AEs can be used in a variety of tasks, including dimensionality reduction, anomaly detection, and data generation [50]. Following the success of Deep Learning, there has been a growing usage of AEs to perform OCC [39]. Within this context, AEs can be trained on normal data and attempt to produce outputs that are similar to the inputs.…”

“…Besides the ANN structure, deep learning architectures include a large number of additional hyperparameters. In order to reduce the search space, using modeling knowledge from previous OCC works (e.g., [39,13]) we fixed some choices, such as the usage of the MAE measure as the loss function for both AE and VAE and usage of Batch Normalization layers for AE. We also restricted the search space for some hyperparameters.…”