{"title":"A Neural Network Based Framework For Classification Of Oceanic Acoustic Signals","authors":"A. Maccato, R.J.P. de Figueii-edo","doi":"10.1109/OCEANS.1989.587491","DOIUrl":null,"url":null,"abstract":"This paper presents a new framework for intelligent acoustic signal processing by artificial neural networks. Problems addressed are the detection, classification, and estimation of signal parameters. The methodology consists of decomposing the above tasks into two stages. First, a highly structured, hierarchical/symbolic representation of the data is created using scale space algorithms. This calculation overcomes moderate noise and warping distortion present in the acoustic recording, and at the same time reduces the data to be processed. Second, neural network architectures are applied to the resulting symbolic structures to obtain the desired signal parameters. The use of neural network techniques allows training to be used in cases where the signals of interest are not easily characterized. Illustrations using simulated and real data will be presented. INTRODUCTION Artificial neural networks provide a new computational paradigm for solving a large class of signal recognition problems. In this paradigm, collections of elementary units called neurons work in parallel to perform a desired computational task. Each neuron performs a component of the overall calculation, and communicates its result to the other units in the network via neurosynaptic interconnections. The distinguishing characteristic of artificial neural networks, with respect to classical methods of computing, is that they can learn to perform a required calculation through training. Hence, rather than designing a procedure for computing the solution to a problem, a selection is made of a training set of exemplary input/output pairs. This is a very powerful property in problem instances for which the application data space is not characterized sufficiently well to allow explicit programming. This is often the case with oceanic acoustic signals, such as short duration transient signals resulting from spurious mechanical events in a vessel. Signal distortion caused by the water medium (amplitude and time warping effects), noise contamination, imprecise or unknown time of occurrence, and high nonstationarity are all difficulties to be confronted when trying to detect a signal, classify it, or estimate meaningful parameters of its source. Additionally, any viable algorithm for recognizing acoustic signals in real-time must possess a fast implementation and, in the case of neural nets, a relatively short training time. To this end, we propose to decompose the recognition task into two stages. The acoustic signal is first preprocessed by a numerical transformation that partially removes noise and distortion effects present in the raw data. In the second stage the resulting information is fed to a neural network for final recognition. The overall method benefits from the filtering characteristics (and possibly data reduction properties) of the first stage, and also from the learning ability of the second stage. OUTLINE O F T H E ALGORITHM To illustrate this concept, we have selected a scale space transformation [l] to preprocess the incoming acoustic signals, and a feedforward, graded response neural net, with the error backpropagation training algorithm [2], to complete the recognition task. The system operates by first obtaining the waveform of the recorded acoustic signal. It then applies to it a scale space transformation, which, as described in the following section, is a time domain algorithm which maps waveforms into surfaces. Elements of the surface topology are consequently captured by a hierarchic/symbolic tree data structure. This step reduces the amount of data to be processed by extracting and ordering the relevant surface features. Finally, the tree is interpreted as a pattern and given as input to the neural network. The neural net, in turn, responds with a code identifying the class to which the tree was assigned, hence identifying the original acoustic signal. Figure 1 is a diagram of the overall process. In the following sections we describe in greater detail the three stages of processing and provide some example applications. SCALE SPACE SIGNAL TRANSFORMATION In general, scale space is a linear transformation of R\" into R\" x R+. For n = 1, it maps waveforms z ( t ) defined on the real line into surfaces (b(t,u) defined on the semi-infinite plane. The mapping takes the form of a convolution integral. The scale parameter, U , determines the 'size' of the smallest allowed intensity features in the filtered signal. That is, if we interpret U as a fixed parameter, say uo, then the filtered waveform zgO(t ) = (b(t,a,) is a smoothed version of the original signal containing only those intensity features larger in size than U,, (in the sense that smaller scale features are greatly attenuated). Figure 2 is an example surface generated by the scale space computation. As an aside, we note that the gaussian kernel in the integral can be replaced by any of its derivatives, allowing 'higher order' surfaces to be generated. The natural hierarchy of inclusion of intensity features, according to size and temporal location, is made explicit by the topology of the level curves of the surface 4, as illustrated in Figure 3. For any fixed ao, these curves are defined by the restriction +(t,a) = a,. The resulting contours partition the semi-infinite surface into well defined regions, amenable to representation by symbolic tree structures. This is ensured by the nature of the transformation and the use of gaussian kernels [3]. Several attributes can be associated with these regions to help characterize the waveform features they represent. Among these are the maximum height (or depth) of the surface within each region, the maximum extent of the region in the time or scale dimensions, the region’s area, and so forth. The scale space transformation is such that gradual signal distortions due to both noise and time warping result in gradual distortions of the surface. Consequently, the contour topology can absorb moderate signal degradation, and provide the next stage of processing with a data structure representative of the underlying signal.","PeriodicalId":331017,"journal":{"name":"Proceedings OCEANS","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"1989-09-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"10","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Proceedings OCEANS","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/OCEANS.1989.587491","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 10
Abstract
This paper presents a new framework for intelligent acoustic signal processing by artificial neural networks. Problems addressed are the detection, classification, and estimation of signal parameters. The methodology consists of decomposing the above tasks into two stages. First, a highly structured, hierarchical/symbolic representation of the data is created using scale space algorithms. This calculation overcomes moderate noise and warping distortion present in the acoustic recording, and at the same time reduces the data to be processed. Second, neural network architectures are applied to the resulting symbolic structures to obtain the desired signal parameters. The use of neural network techniques allows training to be used in cases where the signals of interest are not easily characterized. Illustrations using simulated and real data will be presented. INTRODUCTION Artificial neural networks provide a new computational paradigm for solving a large class of signal recognition problems. In this paradigm, collections of elementary units called neurons work in parallel to perform a desired computational task. Each neuron performs a component of the overall calculation, and communicates its result to the other units in the network via neurosynaptic interconnections. The distinguishing characteristic of artificial neural networks, with respect to classical methods of computing, is that they can learn to perform a required calculation through training. Hence, rather than designing a procedure for computing the solution to a problem, a selection is made of a training set of exemplary input/output pairs. This is a very powerful property in problem instances for which the application data space is not characterized sufficiently well to allow explicit programming. This is often the case with oceanic acoustic signals, such as short duration transient signals resulting from spurious mechanical events in a vessel. Signal distortion caused by the water medium (amplitude and time warping effects), noise contamination, imprecise or unknown time of occurrence, and high nonstationarity are all difficulties to be confronted when trying to detect a signal, classify it, or estimate meaningful parameters of its source. Additionally, any viable algorithm for recognizing acoustic signals in real-time must possess a fast implementation and, in the case of neural nets, a relatively short training time. To this end, we propose to decompose the recognition task into two stages. The acoustic signal is first preprocessed by a numerical transformation that partially removes noise and distortion effects present in the raw data. In the second stage the resulting information is fed to a neural network for final recognition. The overall method benefits from the filtering characteristics (and possibly data reduction properties) of the first stage, and also from the learning ability of the second stage. OUTLINE O F T H E ALGORITHM To illustrate this concept, we have selected a scale space transformation [l] to preprocess the incoming acoustic signals, and a feedforward, graded response neural net, with the error backpropagation training algorithm [2], to complete the recognition task. The system operates by first obtaining the waveform of the recorded acoustic signal. It then applies to it a scale space transformation, which, as described in the following section, is a time domain algorithm which maps waveforms into surfaces. Elements of the surface topology are consequently captured by a hierarchic/symbolic tree data structure. This step reduces the amount of data to be processed by extracting and ordering the relevant surface features. Finally, the tree is interpreted as a pattern and given as input to the neural network. The neural net, in turn, responds with a code identifying the class to which the tree was assigned, hence identifying the original acoustic signal. Figure 1 is a diagram of the overall process. In the following sections we describe in greater detail the three stages of processing and provide some example applications. SCALE SPACE SIGNAL TRANSFORMATION In general, scale space is a linear transformation of R" into R" x R+. For n = 1, it maps waveforms z ( t ) defined on the real line into surfaces (b(t,u) defined on the semi-infinite plane. The mapping takes the form of a convolution integral. The scale parameter, U , determines the 'size' of the smallest allowed intensity features in the filtered signal. That is, if we interpret U as a fixed parameter, say uo, then the filtered waveform zgO(t ) = (b(t,a,) is a smoothed version of the original signal containing only those intensity features larger in size than U,, (in the sense that smaller scale features are greatly attenuated). Figure 2 is an example surface generated by the scale space computation. As an aside, we note that the gaussian kernel in the integral can be replaced by any of its derivatives, allowing 'higher order' surfaces to be generated. The natural hierarchy of inclusion of intensity features, according to size and temporal location, is made explicit by the topology of the level curves of the surface 4, as illustrated in Figure 3. For any fixed ao, these curves are defined by the restriction +(t,a) = a,. The resulting contours partition the semi-infinite surface into well defined regions, amenable to representation by symbolic tree structures. This is ensured by the nature of the transformation and the use of gaussian kernels [3]. Several attributes can be associated with these regions to help characterize the waveform features they represent. Among these are the maximum height (or depth) of the surface within each region, the maximum extent of the region in the time or scale dimensions, the region’s area, and so forth. The scale space transformation is such that gradual signal distortions due to both noise and time warping result in gradual distortions of the surface. Consequently, the contour topology can absorb moderate signal degradation, and provide the next stage of processing with a data structure representative of the underlying signal.