We present an anomaly-based drone classification scheme. High dimensional spectrum data is encoded using a convolutional neural network autoencoder. This is trained on data generated from a generic mathematical drone model. Once encoded, we use quadratic discriminant analysis for non-drone classes and define anomalies in terms of the log likelihood and prior knowledge from the drone model. When integrating ten samples, we can discriminate drones from non-drone samples such as birds, with an average accuracy of 98% at 20 dB signal to noise ratio. This corresponds to an effective observation time of 90 ms.

We present two neural network solutions for data driven track before detect applications. The detected tracks may be used to estimate good initial states for traditional trackers such as Kalman filters. We evaluate the method on different scenarios with multiple targets, non-linear trajectories, and different signal to noise ratio (SNR) values. Depending on scenario, the presented method achieves 99% detection probability on Swerling 3 and 4 targets at 5 - 13 dB SNR, with 0.04 - 0.001 false tracks per frame. The presented method is compared to a theoretically optimal detector.

Multipath interference while tracking sea-skimming targets can significantly distort the estimated height of the target. If accounted for however, this interference can be used to obtain more accurate estimates. In this study, we accomplish this with a convolutional neural network (CNN) used as a parameter estimator. The performance of this network is compared with maximum likelihood and least-squares methods. We found that the CNN performs well in comparison to these methods with only a fraction of the computations.

This thesis combines radar signal processing, with data driven artificial neuralnetwork (ANN) methods. Signal processing algorithms are often based on modelingassumptions of how the data was formed. In some cases, such models are sufficientfor designing good, or even optimal, solutions.In many cases however, these models may be too complicated to form analyticalsolutions; be too simplified, such that practical results may differ significantly fromwhat was theoretically indicated; be unknown in the sense that one of several knownmodels or parameter values may fit the data, but we do not know which; or be toocomplex such that the solution will be too heavy to compute.Data driven ANN methods provide a simple way of bridging these gaps. Wedemonstrate this in four different studies, where we make use of radar data modelsto formulate data driven solutions that are both accurate and computationallyefficient.We compare ANN based results to computationally demanding least squares,and exhaustive matched filtering approaches. We show that the performance of theANNs are comparable to these, but come at a fraction of the computational load.We train ANNs on data sampled from models using a wide range of parametervalues. This naturally handles drifts and unknown parameter values, which maychange the data, but not the desired prediction. We show that an ANN classifiertrained on data from simple models may in practice perform significantly worsethan what was expected from theory. We improve this by combining a limitedamount of real data with synthetic model data. In all cases, we make use of modelsthat are simple to evaluate. They are however not simple to analyze for the purposeof creating analytical solutions.In particular we present a method for achieving non-coherent pulse compressionthat resolves targets within a single pulse width. We present a method fordetecting weak target trajectories that does not require prior assumptions ontarget acceleration, the signal-to-noise ratio, etc. We present different methodsof incorporating imperfect model data in the training of classifiers of drone andnon-drone targets. Finally we present a method for estimating the path differencein sea surface multipath propagation, for the purpose of target tracking.

Denna avhandling kombinerar radarsignalbehandling med datadrivna artificiellaneuronnät (ANN) metoder. Signalbehandlings algoritmer är ofta baserade påmodellantaganden om hur data kommit till. I vissa fall är sådana modeller tillräckligaför att designa bra, eller till och med optimala, lösningar.I många fall kan dock dessa modeller vara alltför komplicerade för att ta framanalytiska lösningar; vara alltför förenklade, så att resultat från praktik skiljer sigmarkant från vad som var teoretiskt väntat; vara okända i den bemärkelsen att enav många kända modeller eller parameter värden kan passa data, men vi vet intevilket; eller vara alltför komplexa så att lösningen blir för tung att beräkna.Datadrivna ANN metoder kan användas för att brygga dessa gap. Vi demonstrerar detta i fyra olika studier, där vi använder oss av modeller för radardata föratt formulera datadrivna lösningar som är både noggranna och beräkningsmässigteffektiva.Vi jämför resultat från ANN lösningar med resultat från beräkningsmässigt merkrävande minsta kvadrat och matchade filter lösningar. Vi visar att prestandan avANN är jämförbar med dessa, men kräver endast en bråkdel av beräkningarna. Vitränar ANN på data från modeller med ett brett span av parameter värden. Dettaför att hantera okända parameter värden och förändringar i dessa, som kan ändradata men inte den önskade prediktionen. Vi visar att en ANN klassificerare som ärtränad på data från enkla modeller kan prestera anmärkningsvärt sämre i praktikenfrån vad som vad var teoretiskt väntat. Vi förbättrar detta genom att kombinera enbegränsad mängd riktigt data med syntetiskt modelldata. I samtliga fall användervi oss av modeller som är enkla att exekvera. Dom är dock inte enkla att analyserai avsikt att skapa analytiska lösningar.Mer konkret presenterar vi en metod för icke-koherent pulskompression somlöser upp mål inom en pulsvidd. Vi presenterar en metod för att detektera svagamålspår som inte kräver antaganden om målets acceleration, signal-till-brus förhållande etc. Vi presenterar olika metoder för att inkorporera bristande modelldata iträningen av klassificerare för drönare och icke-drönare mål. Till sist presenterar vien metod för estimering av skillnaden i sträcka under flervägsutbredning över sjö,i avsikt av målföljning.

Classifiers using convolutional neural networks (CNNs) often yield high accuracies on samples that come from the same distribution as the training data. In this study we evaluate a CNN classifier's ability to discriminate drones from non-drone targets, such as birds, when they are not represented in the training data. We found that the mean accuracy on such out-of-distribution drones was 78%. By introducing a synthetic drone class, generated from a mathematical model, the out-of-distribution drone accuracy was improved to 86%. When trained on all drone types the mean accuracy over all classes was 90%. The data was collected with a 77 GHz mechanically scanning radar with only 9 ms dwell time.

A deep neural network (DNN) is used for achieving subpulse resolution in non-coherent stepped frequency waveform radar. The trade-off between high resolution and long range in radar systems is often addressed using pulse compression, allowing both long pulses and high resolution by increasing the pulse bandwidth. This typically requires a coherent radar. In this study we present a deep learning based solution for achieving subpulse resolution with a non-coherent radar. Our results for such a system are comparable to an equivalent coherent system for SNRs greater than 10 dB. All results are based on simulated data.

Energy efficiency is of profound importance for prospective 5G wireless networks, especially in sparsely populated rural areas where broadband mobile services should be provided at a reasonable cost. In this paper the impact of beamforming (BF) and cell discontinuous transmission (cell DTX) technologies on the average area power consumption is studied. The required density of base stations for a 5G cellular system in a rural environment is also investigated. For this purpose, we propose a simple rural area model that captures a non-uniform distribution of users and employ the generalized Lloyd algorithm to determine the deployment of base stations. We assume a 5G system operating in mmWave band centered at 28 GHz with the bandwidth of 100 MHz, compared with existing LTE networks at 0.8 GHz with a 20 MHz bandwidth. Simulation results show that for the 5G network the density of base stations needed to provide 50 Mbps for 95% of users at the busy hour will be reduced by 8-9 times with the implementation of BF. It is also observed that BF has a greater effect on the energy saving of 5G networks in rural areas in comparison to the cell DTX.