structure of a neural network
Abstract
Recently, as the interest in autonomous driving increases, research on sensors used for autonomous driving has been actively conducted. Sensors used for autonomous driving include cameras, lidars, radars, and ultrasonics. Among them, the radar has an extended maximum detectable range and is robust to harsh environments such as rain or no light. Thus, radar is an essential sensor for autonomous driving. Radars can be used for a variety of purposes, mainly for adaptive cruise control (ACC) and automatic emergency braking (AEB). The radars can be used to detect the distance to the target, relative velocity, angle amongst other things, and can also detect the type and size of the target.
In this dissertation, I proposed the structure of a neural network to classify the road environments. Autonomous driving faces various road environments, and applying an appropriate target detection algorithm according to the road environment is necessary. For example, in the case of an iron tunnel composed of several iron structures, the reflection signal is powerful, causing the target to be undetected.
Therefore, to detect the desired target, applying a clutter removal algorithm is required. To classify the road environment and recognize the road environment from a distance in advance, I proposed the structure of a neural network. As a result, the accuracy of classifying road environments is improved by about 14%p. Don't use plagiarised sources.Get your custom essay just from $11/page
Besides, I proposed a method to improve the angular resolution using an automotive radar. To know the position of the target, direction-of-arrival (DOA), as well as distance and velocity, is essential. This can be solved by increasing the antenna aperture size, but increasing the aperture size reduces the field of view and takes up much physical space. To solve this problem, I proposed a method of generating virtual received signals using the linearly predicted array expansion. By applying the generated virtual signals and the actual signals to the DOA estimation algorithm, a method of improving the angular resolution while using a small antenna aperture size was proposed. As a result, using the proposed method, the angular resolution is improved by about 3 degrees.
Finally, I proposed a technique to distinguish the transmission signal in the multi-input and multi-output (MIMO) radar to improve the DOA estimation performance. If multiple transmitting antennas are used, targets can be efficiently detected using a small number of antennas. However, if the signals radiated from each transmit antenna element cannot be distinguished, the degradation of DOA estimation performance occurs. To solve this problem, I proposed a method to distinguish the transmission signals using a maximum likelihood estimation method. Thus, I improved to double the maximum detectable velocity. Also, DOA estimation performance is enhanced by about 3 degrees in terms of root mean square error (RMSE).
Introduction
Recently, with the rise of autonomous driving as a leader of the fourth industrial revolution, research on the radar, which is one of the sensors used for autonomous driving, is being actively conducted. Among the various sensor technologies, radar sensors have become essential because radars are robust to harsh weather conditions compared to cameras and lidar sensors, have a more extended detection range than ultrasonic sensors, as well as being cost-efficient compared to lidar sensors. For these reasons, active research on automotive radars is conducted for safe autonomous driving such as target detection, target classification, road clutter recognition, and suppression.
Automotive radar mainly performs functions such as adaptive cruise control (ACC), autonomous emergency braking (AEB), blind-spot detection (BSD), cross-traffic alert (CTA), rear cross-traffic alert (RCTA) and is classified into long-range radar (LRR), mid-range radar (MRR), and short-range radar (SRR) according to the observable distance.
A modulation method used in automotive radar is mainly frequency modulated continuous wave (FMCW) and is divided into slow-chirp FMCW and fast-chirp FMCW depending on the sweep time length and the use of down-chirp. The slow-chirp method extracts up-chirp beat frequency and down-chirp beat frequency using two types of chirp signals, up-chirp, and down-chirp. The sum and difference of these two beat frequencies are used to estimate the range and relative velocity of the target. However, when the slow-chirp signal is used, there is a problem that a pairing is required. Thus, recently, the fast-chirp signal proposed to solve this problem tends to be used more.
The rest of this dissertation is organized as follows.
First, in Chapter 2, I propose an artificial neural network that can classify road environments. When autonomous cars drive on the road, they face a variety of road conditions. In particular, in a road environment where many iron structures exist, the reflected signal strength due to the iron structure is too strong to act as a clutter element, and the target cannot be appropriately detected. Therefore, it is necessary to apply a clutter suppression algorithm or a target detection algorithm according to the road environment. In this chapter, the road environment is classified into sound barriers, iron tunnels, typical underground tunnels, and open roads, and a structure of a feedforward neural network that can classify the road environment is proposed.
In Chapter 3, I propose a technique to improve the angular resolution using linearly predicted array expansion. Angular resolution is affected by the antenna aperture size. The larger the aperture size, the better the angular resolution. However, if the antenna aperture size is large, it takes up sizeable physical space and narrows the field of view (FOV). In this chapter, transformation vectors are extracted using linearity between received signals in a uniform linear array (ULA) to improve angular resolution without increasing the aperture size of the antenna. Virtually received signals are generated using the extracted transformation vector, and the angular resolution is improved by applying the direction-of-arrival (DOA) estimation algorithm to the actual received signal and the virtual received signal.
Finally, in Chapter 4, I propose a method of distinguishing transmission signals using deterministic maximum likelihood (DML) in multi-input multi-output (MIMO) radar. Since MIMO radar uses multiple transmit antennas, it is necessary to radiate orthogonal signals to distinguish the transmit signals. However, if the signals cannot be distinguished from the receiving antenna, a phenomenon similar to a phase distortion occurs, and thus, the DOA of the target cannot be adequately estimated. To solve this problem, in this chapter, DML is applied to received signals to classify the transmitted signals and applied to the angle estimation algorithm to improve the DOA estimation.
Conclusion
In this dissertation, I proposed various signal processing techniques for automotive radar. The research was conducted using 77 GHz FMCW radar, which is mainly used in vehicle radar. I suggested various problems that can occur in a road environment when autonomous driving using radar sensors and the proposed method to solve the problems.
First, I proposed a method for recognizing and classifying various road environments. By recognizing and classifying the road environment in advance, the radar system can determine the performance degradation section in advance and apply a proper target detection algorithm according to the road environment. The road environments are classified by the fact that each road environment has different frequency magnitude response characteristics. The feedforward network structure suitable for road environment classification is proposed to improve the road environment classification performance.
Next, I proposed a new extrapolation technique to solve the angular resolution problem that can occur in a sophisticated road environment with many adjacent targets. Generally, the ranges and velocities of the targets can be used to separate adjacent targets. However, when the ranges and velocities are similar, the high-resolution DOA estimation method is essential. Therefore, I generated virtually received signals using linearity in the ULA and achieved an improvement in angular resolution without increasing the physical aperture size of antennas.
Finally, I proposed a technique for distinguishing transmission signals in MIMO radar. As MIMO radar uses multiple transmit antennas, they must radiate orthogonal transmit signals with each other. In addition, different transmission signals can be distinguished from the receiving antenna to estimate the accurate positions of targets. In this dissertation, I improved the DOA estimation performance in the MIMO radar by distinguishing the transmission signals using the DML method.
Introduction
There are several types of automotive radar systems. Among them, FMCW is widely used due to the low complexity of the received signal processing procedure. The FMCW modulation uses ramp signals, and the duration of the entire ramp signal determines the velocity resolution. In recent years, a modified version of the FMCW modulation, a fast-ramp signal which uses multiple linear frequency modulated signals consecutively in a short period, has become of interest. In the fast-ramp modulation, to estimate velocity and range of a target, two-dimensional fast Fourier transform (2D-FFT) is used.
To enhance the angle resolution of automotive radar systems, the need for MIMO radar systems has also been increasing. MIMO radar systems improve target detection, recognition, and tracking performance. In MIMO radar systems, estimating the DOAs of targets using a small number of antennas is possible because virtual antennas are generated by the relative position of the transmitting and receiving antennas. However, if every transmit antenna transmits the same signal simultaneously, the signals cannot be distinguished at the receiver. For solving the problem, various techniques for distinguishing signals have been proposed.
One method is a time-division multiplexing (TDM). In the TDM method, each transmitter radiates signals at predetermined time intervals. The signals are radiated sequentially from each transmit antenna element. However, in this method, using more transmit antennas decreases the power which can be transmitted in each time slot. In addition, the period of the entire chirp is increased and the maximum detectable velocity is reduced due to the velocity ambiguity. This negates the advantages of a MIMO radar system. To overcome the disadvantage of the TDM MIMO radar system, binary phase modulation (BPM) MIMO, where each transmitter simultaneously radiates signals which are coded differently from each other, can be used. Using this modulation has the advantage that more power can be used in one-time slot than the TDM because it can radiate signals simultaneously from different transmit antennas. However, this modulation method also needs to be divided into odd chirps and even chirps in the decoding process, the maximum detectable velocity is reduced by half like the TDM. Besides, computational complexity is high because the time domain signal is used in the signal distinction process.
In this chapter, I propose an advanced signal distinction method for automotive MIMO fast-ramp radar systems. First, I use space-time block codes to distinguish transmitted signals. Space-time block codes are used to generate different transmit signals which have orthogonal properties, and the phases of the transmit signals are shifted at regular intervals depending on the number of transmit antennas. Then, to find the proper array for DOA estimation, I match each of the distinguished transmitted signals to their transmit antenna by applying the DML algorithm. By identifying what the signal is radiated from the first transmit antenna, the maximum detectable velocity is not reduced, and DOA estimation accuracy is enhanced. If the transmitted signals are mismatched, the correlation between the received signal and the steering matrix is reduced, and DOA estimation performance is degraded. Since DML estimates the method by projecting the received signal vectors onto the null space of the steering matrix, it can be used to distinguish the transmitted signals coming from different transmit antennas. Compared to the conventional BPM MIMO method, different signals can be distinguished through simple signal processing.
The remainder of this chapter is organized as follows. First, the fundamentals of a fast-ramp FMCW radar system, coded MIMO radar signal, and deterministic maximum likelihood is introduced in Section 4.2. In Section 4.3, a method of classifying different transmitted signals using DML is presented. Then, the DOA estimation performances using simulations and experimental results are analyzed in Section 4.4. Finally, conclusions are provided in Section 4.5.
Conclusion
In this chapter, I proposed a method to enhance the DOA estimation performance in the fast-ramp modulation MIMO radar system. Different transmitted signals are radiated using the space-time block code to distinguish the signals radiated from each transmit antenna. To solve the array mismatching problem which occurs when multiple transmit antennas are used, I applied the DML algorithm to the 2D-FFT peaks. Then, by comparing the minimum values of the normalized error functions, I found the exact array combination. For the verification of its application in the real world environment, the proposed method was verified by simulations and experiments with an actual automotive radar system. As a result, I can distinguish several different signals radiated from each transmit antenna element, and the DOA estimation performance was enhanced in the fast-ramp MIMO radar system. The proposed method can be applied to various fields requiring the distinction of different transmitted signals.
Abstract
In this dissertation, I studied an efficient receiver design for wireless communication systems using emerging communication systems standards.
First, I studied the method of mitigating phase noise in single-carrier (SC) frequency domain equalization (FDE) systems based on IEEE 802.11ad, which is the first wireless local area network (WLAN) specification that uses mmWave bands. Specifically, I proposed a technique to mitigate residual phase noise after applying the conventional phase noise mitigation method. The proposed method provides precise estimated signal-to-noise ratio (SNR) under severe phase noise level and reflects the power of residual phase noise, which is difficult to compensate, in the SNR estimation. Then, residual phase noise is suppressed during the data-field SC-FDE process using the estimated SNR. The proposed method can effectively mitigate residual phase noise even in non-line-of-sight (NLOS) environments. In the severe phase noise environment with -82 dBc/Hz low-frequency component power (PSD0) of one-pole one-zero phase noise model, mean SNR estimation error using the proposed method is less than 3 dB while the mean SNR estimation error using the conventional SNR estimation method is over 10 dB. Moreover, the standard deviation of the estimated SNR using the proposed method is less than 0.5 dB, while that of the conventional method is over 2 dB. Link-level simulation results show that the packet error rate (PER) performance using the proposed method can achieve 1\% PER at 36 dB SNR with 64QAM modulation, 13/16 code-rate, and -88 dBc/Hz PSD0 while PER performance using the conventional method in the same simulation environment cannot achieve 1\% PER even at 40 dB SNR.
Next, I proposed an integrated receiver design that can apply the proposed phase noise mitigation method of the first research topic to both transmission schemes, SC, and orthogonal frequency-division multiplexing (OFDM), which are supported by the IEEE 802.11ad specification. The proposed integrated receiver can effectively share the architecture of the SC receiver, which includes modules for efficient phase noise mitigation, and the phase noise mitigation method for SC transmissions can be applied to OFDM transmissions. Link-level simulation results show that the required SNR to achieve 1\% PER is reduced by 1.2 dB and 0.9 dB for 16QAM and 64QAM OFDM transmissions, respectively, by applying the proposed method in the first research topic to OFDM transmissions under severe phase noise level.
Finally, I studied how to increase the implementation feasibility of fifth-generation (5G) mobile communications systems. Specifically, I focused on the newly introduced forward error correction (FEC) technology, polar codes that were not used in past communication systems. For the decoding of polar codes, a list successive cancellation decoder, which guarantees promising performance, is most widely used. However, L number of smallest path metric values are extracted from a 2L number of extended path metric values for every sequential information bit decoding, i.e., extended decoding latency is induced by the path metric extraction process. This path metric extraction process is performed based on the sorting of 2L number of extended path metric values and requires a long sequential process. To reduce this latency, I proposed a method that extracts the L number of path metric values without any sorting process. The proposed method adopts the pipelined heuristic extraction method, which does not sort all the path metric values but extracts path metric values smaller than a threshold value. To determine the threshold value, I analyzed the distribution of the path metric values according to every information bit index. This extraction procedure is performed repeatedly until the extraction of L path metric values is complete. To verify the effectiveness of the proposed extraction method, I performed block-error-rate (BLER) simulations for 5G control and broadcast channels. Simulation results show that the mean logical latency (sequential stage) of the proposed extraction method is less than 1.1. The required SNR to achieve 0.1% BLER is degraded by less than 0.06 dB compared to that of using an exact sorter, regardless of the list size L.
By applying the techniques proposed in this dissertation, the baseband reception performance and the implementation feasibility of beyond 5G wireless communication systems can be improved.
Introduction
Recently, there has been a growing interest in beyond 5G (B5G) wireless communication systems that use the millimeter wave (mmWave) band. These systems aim to use wide frequency bands of high carrier frequency bands that have not been used previously. Interest in the mmWave band does not depend on the communication standard used. For example, the IEEE 802.11 working group (WG) began standardizing on the IEEE 802.11ad standard in 2009, assuming the use of 60 GHz unlicensed bands and was first released in 2013. This specification was the starting standard using 60GHz and also attracted attention as the transitional standard for the next-generation mmWave wireless local area network (WLAN) specification, the IEEE 802.11ay, scheduled for the first release in 2020.
Meanwhile, in the field of mobile communication systems, using the 28 GHz band is introduced as the primary use band of fifth-generation mobile communications (5G) systems. In this emerging communications standard, it is necessary to look back on the problems that have not been significant in the conventional systems. In addition to looking back on the existing degradation factors, I also need to look at the performance improvements of the newly added elements, such as polar codes in 5G communications systems, that the emerging communications systems adopt.
As the first factor to revisit, the influence by phase noise can be considered — imperfections of the local oscillator cause phase noise. When systems use the conventional local oscillator, the intensity of phase noise is several times stronger in mmWave communications systems compared to that in conventional low carrier frequency systems. Especially in single-carrier (SC) based communications systems with mmWave carrier frequency, the strength of the signal distortion can be greater. Therefore, I analyzed the effects of phase noise in the wireless communication system based on the IEEE 802.11ad standard and proposed a technique to mitigate them. Besides, for the design of a transceiver based on IEEE 802.11ad that requires consideration of transmitting and receiving signals using OFDM, a receiver design technique that can apply the same technique to OFDM is studied.
For the improvement of newly added elements, I investigated the implementation feasibility of polar codes, which are introduced for the forward error correction of 5G control and broadcast channels. Polar code is widely known as an algorithm that is less burdensome to implement based on the sequential decoding manner and gives flexibility, such as using the software in the implementation. However, compared with other forward error correction codes, such as low-density parity-check codes, it has the disadvantage of the low performance at short codeword length. For overcoming this drawback, a list decoding method is introduced, but this list decoding method has an extended processing delay due to having to determine the priority of each path metric. Therefore, I have studied techniques to reduce this processing delay time. The proposed technique significantly reduces processing latency and shows competitive performance based on the methodology for fully determining the priority of path metrics.
The rest of the dissertation is organized as follows. In Chapter 2, I analyze the effect of the phase noise and introduce the phase noise mitigation method for SC frequency domain equalization systems based on IEEE 802.11ad. Integrated SC and OFDM receiver design for efficient phase noise mitigation using the same standard is introduced in Chapter 3. The complexity and latency reduction method for list successive cancellation decoder of polar codes will be followed in Chapter 4. This dissertation will be ended with concluding remarks in Chapter 5.
Conclusion
In this dissertation, I analyzed the effects of phase noise and introduced an efficient phase noise mitigation method for mmWave communication systems based on IEEE 802.11ad. The proposed algorithm was derived based on the analysis and effectively mitigated the residual phase noise regardless of the phase noise level in NLOS environments. In addition to the phase noise mitigation method, the integrated receiver design to support the phase noise mitigation for both transmission schemes, SC and OFDM, is proposed. The proposed phase noise mitigation can be efficiently applied to OFDM transmission by using the proposed integrated receiver design. For enhancing the error correction module feasibility of emerging B5G communications systems, a low-latency path metric extractor for list successive cancellation decoder of polar codes was proposed. By using the proposed path metric extractor, the LSCD decoder achieved near one, logical latency regardless of the list size. Furthermore, the proposed path metric extractor showed competitive performance compared to the conventional full path metric extractor. By applying proposed algorithms, implementations of emerging B5G communications systems can be simplified and the performance can be improved.