Hello, if you have any need, please feel free to consult us, this is my wechat: wx91due

EEEME30002 APPLIED ML FOR ENGINEERING SYSTEMS

Coursework Brief

ML-based Automatic Modulation Classification

Release date: 1.00 pm, Monday 17th February 2025

Submission deadline: 1.00 pm, Thursday 1st May 2025

Abstract

You have been hired by APPLIED ML SOLUTIONS (AMLS), an AI-focused consultancy that provides closed solutions to open-ended problems. A client (a wireless communication company) has approached AMLS to produce an automatic modulation classification (AMC) tool that is quick and accurate over a range of conditions. You will have to determine the best solution for the task given to you and write a report in an accessible format for someone educated but not expert in ML.

1 BACKGROUND AND MOTIVATION

AMC represents a fascinating intersection of classical signal processing and modern ML approaches. Traditionally, commu nications engineers relied upon carefully crafted algorithms and manual feature extraction to identify modulation schemes. However, these conventional methods often struggle with the complexity and dynamism of modern wireless environments, particularly in the presence of noise, interference, and channel impairments.

This coursework presents an opportunity to delve into this fascinating confluence of communications theory and AI. You will gain practical experience creating and developing Deep Learning architectures and learn more about the fundamental principles of digital communications. The coursework represents a great introduction to the challenges faced in modern wireless systems, where the ability to automatically adapt to different signal types and communication standards is becoming increasingly important.

Moreover, this work holds particular relevance as we move towards more autonomous and adaptive communication systems. The skills and knowledge gained through this coursework will prove invaluable for those interested in pursuing careers in wireless communications, signal processing, or AI. As we continue to push the boundaries of wireless technology with 5G and beyond, the ability to automatically classify and adapt to different modulation schemes will become ever more essential in creating robust and efficient communication systems.

2 DATASET DESCRIPTION

The client has provided you with a reference to download the dataset – RadioML 2016.10a in Ref. [1]. This synthetic dataset generated with GNU Radio consists of 11 modulations (8 digital and 3 analog) at varying signal-to-noise ratios. This file is formatted as a pickle file, which you can open in Python using the appropriate libraries (i.e. library pickle).

In the dataset, there are a total of 220,000 data instances, each one having two arrays of 128 samples. These two arrays correspond to the I (in-phase)/Q (quadrature) components of the signals. An example of these signals and its corresponding modulation is provided in Fig. 1.

Fig. 1. Example of an arbitrary signal in the dataset corresponding to the modulation class 8PSK and SNR = 0.

The Deep Learning model will have to classify any input data into the corresponding modulation class. A key part of the work will comprise processing and sorting the pickle file into a usable format for TF/Keras or PyTorch to make a classification problem. You can open and load your pickle dataset as follows:

import pickle

with open(’RML2016.10a_dict.pkl’, ’rb’) as f:

Xd = pickle._Unpickler(f)

Xd.encoding = ’latin1’

Xd = Xd.load()

Amplitude [-]

Amplitude [-]EEEME30002 APPLIED ML FOR ENGINEERING SYSTEMS – UNIVERSITY OF BRISTOL – 2024/25 2

3 TASK DESCRIPTION

The client wants to have a deep learning tool that solves this complex AMC problem along with a brief report showing and justifying the development process. The client is asking you to address the following areas:

• Data processing

– Process the data so you have the signals, the modulation schemes and SNR ratios providing a summary of what could be used for.

– Split the data appropriately to enable training and testing.

• Model development and implementation

– Implement a deep neural network justifying your choice for a multiclass classification problem

– Use appropriate optimiser, layers, activation functions and loss functions for this type of problem

• Evaluation and visualisation

– Show the overall classification accuracy metrics

– Create a confusion matrix for the different classes

– Show performance for different SNR levels

• Optimisation and/or comparison

– Show the results and metrics for different iterations to optimise the chosen architecture (including at least number and type of layers)

– (Optional) Compare the performance against other types of neural networks that we have covered/are covering throughout the unit

∗ Hint: plot the classification performance for different SNR levels for this comparison exercise

• Critically discuss the results justifying the choices made throughout the coursework. Use literature references to support your choices as much as you can.

4 NOTES & EXPECTATIONS

¬ You have to produce a brief report with 4 pages maximum. Note that the process will be more important than the results. In particular, the following parts will carry a very important weight in your mark:

– Your explanations to justify your choices;

– Your interpretation of the results;

– The discussion of the challenges you have faced throughout the process; and

– Your recommendations for potential future improvements.

The marking rubric is also published along with this coursework brief.

¬ The report is 4 pages maximum. 5 marks will be deducted for any work that exceeds the page limit. The guidance for the formatting are as follows:

– Two templates (1 in MS Word and 1 in LATEX) are provided.

– You are strongly encouraged to use these templates to avoid formatting issues.

– Nonetheless, the general formatting rules are:

∗ 2 columns;

∗ Main text: font size 9, Times New Roman

∗ Section Titles: font size 10, Calibri

∗ Margins: 2.5cm top and bottom; 1.7cm left and right margins;

¬ You have to produce a well-documented Jupyter notebook containing: clear markdown explanations of each step and visualisation of the results.

¬ In the report, you have to include a disclosure statement (in the Acknowledgments section) providing details about (1) the use of Generative AI (if any), (2) what you used it for, and (3) which model (e.g. CoPilot, ChatGPT, etc.) you have used.

¬ The following elements will have to be uploaded to Blackboard:

– The brief report (4 pages max) in PDF format combined with

– A PDF version of the Jupyter Notebook as an appendix of your report (you can use free online conversion tools such as Vertopal [2]); and

– The Jupyter Notebook file (.ipynb)

¬ A potentially useful reference for this work is [3] where visualisation examples of the AMC problem are provided.

Do not try to simply copy what is in the paper as this will be penalised. We need to see your thought process and justifications as well as your unique solutions to this open-ended problem.

¬ To use the dataset in Google Colab, you can download the file from Kaggle and upload it manually into Colab.

Alternatively, you can directly download and unzip it in Google Colab using the following code:

!kaggle datasets download -d nolasthitnotomorrow/radioml2016-deepsigcom

!unzip "/content/radioml2016-deepsigcom.zip" -d "/content/"

REFERENCES

[1] M. Borawski, “Deepsig dataset: Radioml 2016.10a,” Jul 2019. [Online]. Available: https://www.kaggle.com/datasets/ nolasthitnotomorrow/radioml2016-deepsigcom/data

[2] “Online IPYNB (Jupyter Notebook) to PDF Converter — vertopal.com,” https://www.vertopal.com/en/convert/ipynb-to-pdf.

[3] T. J. O’shea and N. West, “Radio machine learning dataset generation with gnu radio,” in Proceedings of the GNU radio conference, vol. 1, no. 1, 2016.

文章