Blind Recognition of Punctured Convolutional Codes from an Unknown Bitstream
Mankinen, Arttu (2025)
2025
Sähkötekniikan DI-ohjelma - Master's Programme in Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-01-14
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202501131366
https://urn.fi/URN:NBN:fi:tuni-202501131366
Tiivistelmä
Error correction is an important part of digital communication systems. The basic idea is to add extra bits to the data bitstream. These extra bits are then used to identify and correct errors caused by the channel. Decoding of error correction codes is an essential step in signal analysis. The goal is to reveal the constant patterns and regularities from the bitstream in order to classify and further investigate the signal samples. The problem, however, is that these regularities lie within the error correction. The challenge is that in the case of an unknown signal sample, the error correction algorithms are not known.
In this Master’s thesis done for the Finnish Defence Research Agency, blind recognition of punctured convolutional codes from a bitstream without any prior knowledge was investigated. The goal of the thesis was to build a program for this task. The research began with a literature review, which investigated which parameters of convolutional codes should be considered, and what algorithms already exist for blind recognition tasks. As a result of the review, generator polynomials and puncturing matrices were identified as the most important parameters.
Based on the literature review, a deep learning model was chosen as the methodology for the work. The training and testing data contained four classes, which contained different generator polynomials and puncturing matrices. The model was first trained, after which its performance was evaluated on test data. The performance of the model was also evaluated by comparing its recognition accuracy to other models in the field.
The model built for the Master’s thesis recognizes the punctured convolutional codes included in the training data with moderately good accuracy. The model also performs well in comparison, especially in low signal-to-noise ratio tests. This was due to data preprocessing method of using synchronization patterns implemented in the training data. The main conclusion of the thesis is that deep learning models can perform the tasks of blind recognition of error correction codes at least as well as traditional algebraic methods. Building the architecture and training data of the deep learning model is a meticulous work of tweaking and is not fully completed for this master’s thesis. In the future, the diversity of the training data must be increased so that the model can also recognize punctured convolutional codes more widely.
In this Master’s thesis done for the Finnish Defence Research Agency, blind recognition of punctured convolutional codes from a bitstream without any prior knowledge was investigated. The goal of the thesis was to build a program for this task. The research began with a literature review, which investigated which parameters of convolutional codes should be considered, and what algorithms already exist for blind recognition tasks. As a result of the review, generator polynomials and puncturing matrices were identified as the most important parameters.
Based on the literature review, a deep learning model was chosen as the methodology for the work. The training and testing data contained four classes, which contained different generator polynomials and puncturing matrices. The model was first trained, after which its performance was evaluated on test data. The performance of the model was also evaluated by comparing its recognition accuracy to other models in the field.
The model built for the Master’s thesis recognizes the punctured convolutional codes included in the training data with moderately good accuracy. The model also performs well in comparison, especially in low signal-to-noise ratio tests. This was due to data preprocessing method of using synchronization patterns implemented in the training data. The main conclusion of the thesis is that deep learning models can perform the tasks of blind recognition of error correction codes at least as well as traditional algebraic methods. Building the architecture and training data of the deep learning model is a meticulous work of tweaking and is not fully completed for this master’s thesis. In the future, the diversity of the training data must be increased so that the model can also recognize punctured convolutional codes more widely.