Attention Augmented Convolutional Neural Network for acoustics based machine state estimation

Abstract

The rapid development of technology is leading to the emergence of smart factories where the Artificial Intelligence paradigm of deep learning plays a significant role in processing data streams from machines. This paper presents the application of Augmented Attention Blocks embedded in a deep convolutional neural network for the purposes of estimating the state of remote machines using remotely collected acoustic data. An Android application was developed for the purposes of transferring audio data from a remote machine to a base station. At the base station, we propose and developed a deep convolutional neural network called MAABL (MobileNetv2 with Augmented Attention Block). The structure of the neural network is constructed by combining an inverted residual block of MobileNetv2 with an augmented attention mechanism block. Attention Mechanism is an attempt to selectively concentrate on a few relevant things, while ignoring others in deep neural networks. Due to the presence of audio frames containing silent features not relevant to the task at hand, an Attention Mechanism is particularly important when processing audio data. The MAABL network proposed in this paper obtains the state of the art results on the accuracy and parameters of three different acoustic data sets. On a relatively large-scale acoustic dataset regarding machine faults, the method proposed in this paper achieves 98% accuracy on the test set. Moreover, after using transfer learning, the model achieved the state of the art accuracy with less training time and fewer training samples.