Identification of Communication Signals Using Image Processing Techniques, 16-9290Printer Friendly Version
Inclusive Dates: 01/01/02 10/01/05
Background - Signal recognition and identification is a key technology area for the Signal Exploitation and Geolocation Division. This project investigated two techniques commonly used in image processing to determine their efficacy for a priori signal recognition task. A wavelet-based algorithm was evaluated for symbol-rate estimation, and Karhunen-Loève (KL) decomposition was applied to signal classification.
Approach - Evaluation techniques previously used with classical signal-processing algorithms does not give adequate insight into the performance of these new algorithms. Overcoming this limitation required extensive work prior to the actual wavelet and KL analysis. Therefore, the project approach consisted of five principal tasks: 1) Create a large number of simulated test signals with known characteristics that adequately represent the target environment, 2) Develop an automated framework for running these large test sets through an algorithm and collecting the test results, 3) Develop analysis tools to evaluate the efficacy of the algorithm, 4) Test the signals and framework against one or more known algorithms, 5) Develop and test algorithms based on the wavelet transform and KL decomposition.
Accomplishments - The first three tasks required extensive MATLAB® and Simulink® programming to construct a set of simulated signal generators, the test framework, and the analysis tools. Developing a uniform method for recording both test signal data and algorithm test results became an integral part of this effort. With no adopted standard in this area for data interchange, the Extensible Markup Language (XML) was used to encode both the test signal library information and the algorithm test data. For the fourth task, several symbol-rate algorithms were evaluated against a common set of test signals. The resulting analysis yielded several unique insights into their performance. We then improved a published algorithm for symbol-rate estimation based on the continuous wavelet transform. While its overall accuracy was not as high as existing algorithms, it performed well, and we were able to calculate a quality metric that strongly indicated whether the estimated value was correct. Our KL-based classification algorithm was highly selective in identifying PSK signals even with low signal-to-noise ratios. Its specificity for other modulation types diminished sharply in the presence of noise.