Institute for Critical Technology and Applied Science. Machine Learning for Radar State Determination. Status report 2017/11/09

Transcription

1 Institute for Critical Technology and Applied Science Machine Learning for Radar State Determination Status report 2017/11/09

2 Background/Goals Understand machine learning and its various flavors Demonstrate the efficacy of machine learning for Wideband (RF) radar waveform classification Prediction of radar state changes and determination of approach policy Investigate improvement of combined wideband and state estimation in policy understanding 2017/11/08 2

3 The problem Aircraft attempting to bypass a multi-function tracking radar Consider from the perspective of the aircraft with a radar warning receiver Assume maneuvering and jamming capability Assume limited power available (minimize use of jammer) Level of threat posed by radar depends on its current mode of operation Desire to Discover the states of the radar Uncover the policy behind the changing of states Determine an optimized approach methodology that minimizes detection/tracking Required steps Classify observed radar waveforms for each state Estimate the current mode of operation of the observed radar Learn best approach policy 2017/11/08 3

4 Radar Review Pulse-Doppler radar determines the range to target using pulse-timing techniques Measures range via time between sending and receiving Pulses travel at known velocity so distance is computed Ability to identify two targets at different ranges is range resolution Return signal is delayed copy of transmitted signal Detection is optimally accomplished via matched filtering Cross-correlation of the received signal with the transmitted signal 2017/11/08 4

5 Matched Filter & Pulse Compression Matched filter serves two purposes here Allows for the detection and localization of returns in time Allows for the compression of the pulse in time Pulse compression allows the radar system to transmit a pulse of long duration Low peak power Resolution and detection performance of a short-pulse Range resolution determined by the bandwidth of the TX signal Accomplished by coding the RF carrier to increase the bandwidth of the transmitted waveform Different coding schemes have different tradeoffs (processing requirements, compression ratio, sidelobe levels, etc.) ρ r = c 2B 2017/11/08 5

6 Pulse Coding Many coding techniques that can be used to achieve the bandwidth and/or pulse compression Linear frequency modulation (LFM or chirp) Hyperbolic frequency modulation (HFM) Binary phase codes Polyphase Codes Linear Frequency Modulation Coding of radar waveform analogous to coding in communication waveforms to convey data Binary Phase Coded 2017/11/08 6

7 Another complication Pulse coding allows any given pulse to identify aircraft location (speed, direction, range) to some level of accuracy Multifunction radars use pulse streams optimized for a particular purpose Waveform (pulse coding and/or time characteristics) change over time based on what radar is doing or observing Very large space of possible waveforms Potential to understand what radar is thinking based on observations of waveforms Arasaratnam, I.; Haykin, Simon; Kirubarajan, T.; Dilkes, Fred A. "Tracking the mode of operation of multi-function radars", Radar, 2006 IEEE Conference on, p /11/08 7

8 Let s play a game Attempt to fly aircraft against multi-function radar simulator Radar has policy for which waveforms to use based on SNR, target location, etc. Not modeled on any specific radar, but reasonable policies/waveforms are used Aircraft can maneuver and activate jammer Reasonable maneuver rates (speed, turn rate, etc.) Limited jammer use and power available Physical limits on range (no going out of bounds) Goal is to figure out maneuver and/or jammer policy to Prevent radar from entering/remain track maintenance mode Goodness functions are successful passing, minimum time, minimum jammer usage 2017/11/08 8

9 Machine learning for pulse recognition Ongoing research into using machine learning for waveform recognition Feature based extract features from wideband and pass feature set to ML algorithms J. Lunden, L. Terho and V. Koivunen, "Waveform Recognition in Pulse Compression Radar Systems," 2005 IEEE Workshop on Machine Learning for Signal Processing, Mystic, CT, 2005, pp J. Lunden and V. Koivunen, "Automatic Radar Waveform Recognition," in IEEE Journal of Selected Topics in Signal Processing, vol. 1, no. 1, pp , June Wideband use raw RF samples and convolutional neural networks We ll focus on wideband RF classification 2017/11/08 9

10 Modulation type Neural Network for Waveform Identification Wideband pulse locations identified in I/Q data Mixed and filtered to baseband Baseband I/Q analyzed by Convolutional Neural Network to identify modulation Hauser, Headley, and Michaels, Signal Detection Effects on Deep Neural Networks Utilizing Raw IQ for Modulation Classification 2017/11/08 10

11 Classification Investigate multiple classification algorithms and ANNs Train on wide variety of synthetic pulses Identify/classify between types LFM HFM Costas Frank Estimate relevant parameters not all from ML algorithms Pulse width Bandwidth Chirp type Chirp rate P1 P2 P3 P4 Code Arity Bit rate Etc Feed state determination engine 2017/11/08 11

12 Machine learning for state recognition Receive observations of current waveform Desire to estimate the modes of operation of the observed radar Determine policy for when states change Optimize aircraft resource utilization to Traverse past radar Ensure track maintenance mode not achieved Arasaratnam, I.; Haykin, Simon; Kirubarajan, T.; Dilkes, Fred A. "Tracking the mode of operation of multi-function radars", Radar, 2006 IEEE Conference on, p /11/08 12

13 Reinforcement learning Reinforcement Learning Learning what to do and how to map situations to actions End result is to maximize a numerical reward signal Learner must discover which action will yield the maximum reward Here the Agent is the aircraft, with actions of maneuvers, jammers, etc. Environment is the radar simulation in which the model is run Reward is given for successful traversal, penalties for jammer use, etc. State represents what the aircraft is doing (jammer on/off, position, speed, heading, etc.) 2017/11/08 13

14 Reinforcement learning approach Agent placed in an environment and must learn to behave optimally in it Assume that the world behaves like an MDP except: Agent can act but does not know the transition model Agent observes its current state its reward but doesn t know the reward function Goal: learn an optimal policy Two approaches to reinforcement learning Model based approach Learn the MDP model, or an approximation of it Use it to find the optimal policy Model free approach Derive the optimal policy without explicitly learning the model 2017/11/08 14

15 Current status Project officially started October 1 Team formed over next couple weeks Kickoff meeting October 20 Have completed literature review Have waveform generation code Have investigated languages and ML packages for use Standardizing on Python Looked at Pandas, SciKit-learn, Keras, TensorFlow Decided on TensorFlow possibly augment w/ PySPACE Just beginning ML classification investigation Rubina Adhikari Undergraduate Researchers Vanessa Arndorfer Alexander Giffen PI Tom Krauss PM Ehren Hill 2017/11/08 15

16 Notional Schedule You are here 2017/11/08 16

17 Institute for Critical Technology and Applied Science Questions 2017/11/08 17