Thesis (MSc)--Stellenbosch University, 2015. / ENGLISH ABSTRACT: Rock falls are the cause of the majority of mining-related injuries and fatalities in deep tabular South African mines. The standard process of entry examination is performed before working shifts and after blasting to detect structurally loose rocks. This process is performed by a miner using a pinch bar to ‘sound’ a rock by striking it and making a judgement based on the frequency response of the resultant sound. The Electronic Sounding Device (ESD) developed by the CSIR aims to assist in this process by performing a concurrent prediction of the structural state of the rock based on the acoustic waveform generated in the sounding process. This project aimed to identify, develop and deploy an effective classification model to be used on the ESD to perform this assessment. The project was undertaken in three main stages: the collection of labelled acoustic samples from working areas; the extraction of descriptive features from the waveforms; and the competitive evaluation of suitable classification models. Acoustic samples of the sounding process were recorded at the Driefontein mine operation by teams of Gold Fields employees. The samples were recorded in working areas on each of the four reefs that were covered by the shafts of the mine complex. Samples were labelled as ‘safe’ or ‘unsafe’ to indicate an expert’s judgement of the rock’s structural state. A laboratory-controlled environment was also created to provide a platform from which to collect acoustic samples with objective labelling. Three sets of features were extracted from the acoustic waveforms to form a descriptive feature dataset: four statistical moments of the frequency distribution of the waveform formed; the average energy contained in 16 discrete frequency bands in the data; and 12 Mel Frequency Cepstral Coefficients (MFCCs). Classification models from four model families were competitively evaluated for best accuracy in predicting structural states. The models evaluated were k-nearest neighbours, self-organising maps, decision trees, random forests, logistic regression, neural networks, and support vector machines with radial basis function and polynomial kernels. The sensitivity of the models, i.e. their ability to avoid predicting a ‘safe’ status when the rock mass was actually loose, was used as the critical performance measure. A single-hidden-layer feed-forward neural network with 15 nodes in the hidden layer and a sigmoid activation function was found to best suited for acoustic classification on the ESD. Additional feature selection was performed to identify the optimised form of the model. The final model was successfully implemented on the ESD platform. / AFRIKAANSE OPSOMMING: Rotsstortings is die oorsaak van die meerderheid van mynbouverwante ongelukke en ongevalle in diep tabulêre Suid-Afrikaanse myne. Die standaard proses van pretoegang ondersoeke om strukturele los rotse te erken, word uitgevoer voor enige werkskof en na skietwerk. Dit word gedoen deur ‘n myner wat ‘n breekyster teen die rots kap en ‘n oordeel vel op die frekwensie weergawe van die gevolglike klank. Die ‘Elektroniese Klinking Toestel’ (Electronic Sounding Device, ESD) is ontwikkel deur die WNNR met die doel om die proses te ondersteun. Dit word gedoen deur ‘n gelyktydige voorspelling van die strukturele toestand gebaseer op die akoestiese golfvorm gegenereer in die proses van klinking. Die projek stel ten doel om ’n effektiewe klassifikasie-model te identifiseer, te ontwikkel en toe te pas in die ESD om hierdie assessering uit te voer. Die projek vind in drie stadiums plaas: die insameling van geëtiketteerde akoestiese monsters van die werkareas; die ekstraksie van beskrywende kenmerke van die golfvorms en die mededingende evaluering van geskikte klassifiseringsmodelle. Klinking akoestiese monsters is opgeneem by Driefontein mynbouoperasie deur spanne van Gold Fields se werknemers. Die akoestiese monsters is opgeneem in werkareas van elk van die vier goudriwwe wat deur die skagte van die mynkompleks gedek word. Monsters is as ‘veilig’ of ‘onveilig’ geëtiketteer as aanduiding van die ekspert se oordeel van die rots se strukturele toestand. ‘n Laboratorium gekontroleerde omgewing is ook geskep om ’n platform te skep vanwaar akoestiese monsters met objektiewe etikettering waargeneem word. Drie stelle van kenmerke is onttrek van die akoestiese golfvorms om ‘n beskrywende datastel van kenmerke te vorm: vier statistiese momente van die frekwensie verspreiding van die gevormde golfvorm; gemiddelde energie ingesluit in sestien diskrete frekwensiebande in die data; en twaalf ‘Mel Frequency Cepstrum Coefficients’ (MFCCs). Klassifikasie modelle van die vier modelsamestellings was kompeterend geëvalueer vir die beste akkuraatheid in voorspellings van strukturele toestande. Klassifikasie modelle het k-naaste bure, selforganiserende kaarte, besluitnemingsbome, lukrake woude, logistieke regressie, neurale netwerke en steun-vektor masjiene met radiale basisfunksie en polinominale kerne. Die meting van die sensitiwiteit van die modelle, met betrekking tot die vermoë van die modelle om veilige voorspellings te beperk wanneer die rotsmassa los is, was gebruik as ’n kritiese werksverrigtingsmeting. ‘n Enkel-verskuilde-laag neurale netwerk met 15 nodes in die verskuilde laag en ’n sigmoïde aktiveringsfunksie is gevind as die mees geskikte vir die ESD. Addisionele keuse van kenmerke is uitgevoer deur die geoptimiseerde vorm van die model te identifiseer. Die model was suksesvol geïmplementeer op die ESD platform.

Finding lost persons, collecting information in disturbed communities, efficiently traversing urban areas after a blast or similar catastrophic events have motivated researchers to develop intelligent sensor frameworks to aid law enforcement, first responders, and military personnel with situational awareness. This dissertation consists of a two-part framework for providing situational awareness using both acoustic ground sensors and aerial sensing modalities. Ground sensors in the field of data-driven detection and classification approaches typically rely on computationally expensive inputs such as image or video-based methods [6, 91]. However, the information given by an acoustic signal offers several advantages, such as low computational needs and possible classification of occluded events including gunshots or explosions. Once an event is identified, responding to real-time events in urban areas is difficult using an Unmanned Aerial Vehicle (UAV) especially when GPS is unreliable due to coverage blackouts and/or GPS degradation [10]. Furthermore, if it is possible to deploy multiple in-situ static intelligent acoustic autonomous sensors that can identify anomalous sounds given context, then the sensors can communicate with an autonomous UAV that can navigate in a GPS-denied urban environment for investigation of the event; this could offer several advantages for time-critical and precise, localized response information necessary for life-saving decision-making. Thus, in order to implement a complete intelligent sensor framework, the need for both an intelligent static ground acoustic autonomous unattended sensors (AAUS) and improvements to GPS-degraded localization has become apparent for applications such as anomaly detection, public safety, as well as intelligence surveillance and reconnaissance (ISR) operations. Distributed AAUS networks could provide end-users with near real-time actionable information for large urban environments with limited resources. Complete ISR mission profiles require a UAV to fly in GPS challenging or denied environments such as natural or urban canyons, at least in a part of a mission. This dissertation addresses, 1) the development of intelligent sensor framework through the development of a static ground AAUS capable of machine learning for audio feature classification and 2) GPS impaired localization through a formal framework for trajectory-based flight navigation for unmanned aircraft systems (UAS) operating BVLOS in low-altitude urban airspace. Our AAUS sensor method utilizes monophonic sound event detection in which the sensor detects, records, and classifies each event utilizing supervised machine learning techniques [90]. We propose a simulated framework to enhance the performance of localization in GPS-denied environments. We do this by using a new representation of 3D geospatial data using planar features that efficiently capture the amount of information required for sensor-based GPS navigation in obstacle-rich environments. The results from this dissertation would impact both military and civilian areas of research with the ability to react to events and navigate in an urban environment. / Doctor of Philosophy / Emergency scenarios such as missing persons or catastrophic events in urban areas require first responders to gain situational awareness motivating researchers to investigate intelligent sensor frameworks that utilize drones for observation prompting questions such as: How can responders detect and classify acoustic anomalies using unattended sensors? and How do they remotely navigate in GPS-denied urban environments using drones to potentially investigate such an event? This dissertation addresses the first question through the development of intelligent WSN systems that can provide time-critical and precise, localized environmental information necessary for decision-making. At Virginia Tech, we have developed a static ground Acoustic Autonomous Unattended Sensor (AAUS) capable of machine learning for audio feature classification. The prior arts of intelligent AAUS and network architectures do not account for network failure, jamming capabilities, or remote scenarios in which cellular data wifi coverage are unavailable [78, 90]. Lacking a framework for such scenarios illuminates vulnerability in operational integrity for proposed solutions in homeland security applications. We address this through data ferrying, a communication method in which a mobile node, such as a drone, physically carries data as it moves through the environment to communicate with other sensor nodes on the ground. When examining the second question of navigation/investigation, concerns of safety arise in urban areas regarding drones due to GPS signal loss which is one of the first problems that can occur when a drone flies into a city (such as New York City). If this happens, potential crashes, injury and damage to property are imminent because the drone does not know where it is in space. In these GPS-denied situations traditional methods use point clouds (a set of data points in space (X,Y,Z) representing a 3D object [107]) constructed from laser radar scanners (often seen in a Microsoft Xbox Kinect sensor) to find itself. The main drawback from using methods such as these is the accumulation of error and computational complexity of large data-sets such as New York City. An advantage of cities is that they are largely flat; thus, if you can represent a building with a plane instead of 10,000 points, you can greatly reduce your data and improve algorithm performance. This dissertation addresses both the needs of an intelligent sensor framework through the development of a static ground AAUS capable of machine learning for audio feature classification as well as GPS-impaired localization through a formal framework for trajectory-based flight navigation for UAS operating BVLOS in low altitude urban and suburban environments.

A growing recognition of the need for effective marine environmental management as a result of the increasing exploitation of marine biological resources has highlighted the need for high speed ecological seabed mapping. The practice of mapping making extensive use of satellite remote sensing and airborne platforms is well established for terrestrial management. Marine biological resource mapping however is not readily available except in part from that derived for surface waters from satellite based ocean colour mapping. Perhaps the most fundamental reason is that of sampling difficulty, which involves broad areas of seabed coverage, irregularities of seabed surface and depth. Conventional grab sample techniques are widely accepted as a standard seabed mapping methodology that has been in use long before the advent of acoustic techniques and continue to be employed. However. they are both slow and labour intensive, factors which severely limit the spatial coverage available from practical grab sampling programs. While acoustic techniques have been used for some time in pelagic biomass assessment, only recently have acoustic techniques been applied to marine biological resource mapping of benthic communities. Two commercial bottom classifiers available in the market that use normal incidence echosounders are the RoxAnn and QTC View systems. Users and practitioners should be cautious however when using black box implementations of the two commercial systems without a proper quality control over raw acoustic data since some researchers in their studies have indicated problems with these two bottom classifiers such as, among others, a depth dependence. In this thesis, an alternative approach was adopted to the use of echosounder returns for bottom classification. / The approach used in this study is similar to,~ used in the commercial RoxAnn system. In grouping bottom types however, Multivariate analysis (Principal Component Analysis and Cluster Analysis) was adopted instead of the allocation system normally used in the RoxAnn system, called RoxAnn squares. In addition, the adopted approach allowed for quality control over acoustic data before further analysis was undertaken. As a working hypothesis, it was assumed that on average 0 and aE2 = 0 where E1 and E2 are the roughness and hardness indices, respectively, and RO is the depth. For roughness index (E1), this was achieved by introducing a constant angular integration interval to the tail of the first OM returns whereas for hardness index (E2), this was achieved by introducing a constant depth integration interval. Since three different frequencies, i.e. 12, 38 and kHz, were operated, Principal Component Analysis was used here to reduce the dimensionality of roughness and hardness indices, formed from the three operated qu frequencies separately. The k-means technique was applied to the first principal component of roughness index and the first principal comp component of hardness index to produce separable seabed types. This produced four separable seabed types, namely soft-smooth, soft-rough, hard-smooth and hard-rough seabeds. / Principal Component Analysis was also used to reduce the dimensionality of the area backscattering coefficient sA, a relative measure of biomass of benthic mobile biota. The bottom classification results reported here appear to be robust in that, where independent ground truthing was available, acoustic classification was generally congruent with ground truth results. When investigating the relationship between derived bottom type and acoustically assessed total biomass of benthic mobile biota, no trend linking the two parameters, however, appears. Nevertheless, using the hierarchical agglomerative technique applied to a set of variables containing average first principal component of the area backscattering coefficient sA, the average first principal component of roughness and hardness indices, the centroids of first principal component of roughness and hardness indices associated with the four seabed types and species composition of fish group of the common species in trawl stations available, two main groups of quasi acoustic population are observed in the North West Shelf (NWS) study area and three groups are observed in the South East Fisheries (SEF) study area. The two main groups of quasi acoustic population in the NWS study area and the three main groups of quasi acoustic population in the study area are associated with the derived seabed types and fish groups of the common species.