UDT 2020 UDT Extended Abstract Template Presentation/Panel Heterogeneous Classification System for Underwater Acoustic Recognition F. CHAILLAN 1 , S. MEUNIER 2 1 Research Director, NAVAL GROUP, Ollioules, FRANCE 2 Marketing Director, NAVAL GROUP, Paris, FRANCE Abstract — Automatic underwater acoustic recognition is a capability used on board to provide operator assistance. The principle is to associate a class to each sound recorded by the passive sonar. IA algorithms based on supervised learning are answers to this problem. Despite of their efficiency, these techniques are limited by the lack of contextual information that would make the automatic decision process more robust. As a first answer, this study attempts to introduce the concept of heterogeneous classification, as an extension of the classical automatic classification task. A first simple experimentation on the case of “unclassification task” is presented and discussed. 1 Introduction As a first answer we propose in this work an original architecture, spatially dedicated to submarines, In the underwater world, submarines use the data which allows to take into account heterogeneous provided by their passive sonars to automatically information and to merge it with the original recognize underwater noises [1]. This classification task classification in order to make the decision robust and provides the operator with an automatic identification statistically exhaustive. proposed to help him to make its own decision. The last decades have gradually shown that AI-based classifiers Finally, we try to test our architecture for the are replacing classical rule-based classifiers, and have heterogeneous classification of underwater acoustic become the most performant and usual approach to noises by an experiment on the mimicry of the “un- designing automatic classification systems [2]. classification” task performed by the operator when he tries to first eliminate safe noises before to focusing on However, designing an AI-based classification potential menaces. tool (AI) that is powerful enough to provide a robust support to the operator is a difficult task. The main issue 2 Automatic Recognition of Underwater is linked with the passive sonar context,: as the Acoustic Noises from Passive Sonar propagation channel is highly dependent on time and Data space, the noise emitted by the same physical object can be received by the sensor in a slightly different way, To introduce the concept of heterogeneous depending on where and when the recording is made. In classification for underwater acoustic recognition, we practice, even well-known noises can become difficult to first briefly describe the classical principle of automatic recognize in the case of unfavourable propagation. classification and its limitation in terms of performance induced by the lack of fresh information available. Consequently, any AI to be designed must be both performant and robust enough to face this limitation. In As a consequence, we then try to describe the the case of a great amount of available data, performance inventory of all information that can be found and used can be enhanced by trying different AI techniques from on board, potentially useful to enhance the strength of the Machine Learning (ML) [3] or Deep Learning (DL) [4], classification task. This allows us to provide a with fine tuning for each algorithm assessed. methodology for designing a heterogeneous classification Performance is also enhanced by refining data selection, system capable of recognizing underwater acoustic noises and by optimizing output classes with split & merge by coupling different sources of available data to operations. eventually make its own decision. Thus, the final decision is ideally an exhaustive synthesis of all available Robustness also needs to be improved, but how information. It is finally brought to the operator to can we enhance classification robustness when the provide him with assistance. considered equipment, in this case passive sonar, has already given all the information it can provide? Our insight is to couple a classification provided by the 2.1 Automatic Recognition of Underwater passive sonar with other data available on-board and at Acoustic Noises: the Classical Approach shore, yielding to the concept of heterogeneous The problem of automatically recognizing classification. underwater acoustic noises from passive sonar data is a difficult one, because the sounds to be recognized may
UDT 2020 Presentation/Panel UDT Extended Abstract Template have similarities whereas they do not belong to the same accurately extracting the signal of interest among the sea class. These physical similarities are such that it is noises, then computing and selecting the features [6]. possible to confuse the biological or mechanical origin of Features are scalars grouped in a vector which represents the noise sources to be recognized. Furthermore, as the the “DNA” of the signal. The IA takes the features as passive sonar device does not provide the range, the same inputs to provide the class associated to the signal. noise source caught at different ranges can have different Conversely, the Deep Learning approach is for the signature; moreover, the propagation channel depends on IA to take the sound samples directly for input. In space and time, so that the same noise source can have a practice, the architecture of the network is divided in two slightly different signature depending on the date and main blocks of neural layers: the self-encoding block and area of the recording. Finally, we recall the sampling the classification block, so that the features are directly frequency of passive sonar is so that all frequencies of the computed by the network. Nevertheless, Deep Learning noise source lying above the half sampling frequency are is not explainable. not observed by the sensor. Consequently, this is particularly damaging when trying to recognize Then, when learning is OK, fine tuning techniques biological activity, as some biological noises may have such as cross-validation are applied to the network in signatures with high frequencies components, at least order to provide optimal values for the hyper-parameters. significantly higher than the half sampling frequency. The next step is to assess the performance of the IA. In this context, the problem is to associate a class The performance is calculated based on the confusion with a sound emitted by the acoustic landscape. Each matrix, the ROC curve and the AUC. Finally, the trained class provides knowledge about the physical origin of the and tuned IA is deployed so that a class can be associated underwater acoustic sources. Consequently, this class with a sound according to a given performance level. serves as a decision-making support for operators. IA algorithms can provide a solution to this complex This proposed methodology, synthetized by Fig.1 problem through supervised learning. has been used in [7] to design from scratch the proposed IA algorithm. The proposed IA algorithm for underwater The purpose is to build the unknown function that acoustic recognition is used in our study as a reference. perfectly associates each acoustic noise to its class. This As an example, Fig.2 shows an output of this classifier. function is approximated by the considered learning algorithm, from a set of sound data labelled by a human expert. Fig.1 describes the steps to be taken to design such a system. The first step consists in building a database, containing selected sounds in order to be sufficiently representative of the classes of sounds to be recognized. Fig. 2. Output of the IA algorithm for underwater acoustic noises recognition based on DL approach The results show that despite high performance, we have no explanation for the output, especially in terms of confidence interval. In particular, we want to make the difference between a cautious decision and a coercive one. Furthermore, the interpretation of the output of the Fig. 1. Principle of automatic recognition of underwater algorithm must be carefully considered. Indeed, as the acoustic noises. The classical approach output is here a probability density function, the most likely way of doing is to attribute the class to the mode of Each selected sound, called an example, must be labelled by a human expert. This is a time-consuming task that the distribution. But it is possible to go further, by considering the pattern of the density instead of the mode. requires ergonomic annotation tools and the full attention of the operator to avoid labelling errors that lead to Because a in real situation on board, the operator reduced performance. Then, the learning step is not only listens to acoustic noises but also reads conducted according to the IA algorithm family documents, consults databases, takes into account the considered. The two main families are Machine Learning location and environmental conditions, we propose to do and Deep Learning. the same for our IA: in order to make the decision more In our context, Machine Learning techniques such robust we associate it with all available and useful information. as Support Vector Machines [5] have the advantage of being explainable, in the sense that it is possible to rigorously justify why the IA provided an output with a known input. However, these techniques require first
Recommend
More recommend
