Scholarlink

The aim of this project is to develop a deep learning model with improved accuracy and efficiency for estimating the frequency of noisy signals, specifically modulated signals that are corrupted by Gaussian noise. While frequency estimation methods exist, many have limitations like being computationally intensive, inefficient or time consuming. In our model, we created training and testing data, followed by a neural network using Keras’ Sequential API. The neural network is then trained based on the data generated, allowing it to predict the frequency of a corrupted signal. Afterwards, we varied some parameters in our code, namely the number of data samples and number of epochs in order to determine the accuracy at each. We graphed our results, based on which we determined the number of epochs to be 8 and the number of data samples to be 5000. We then carried out a qualitative comparison of our method with two existing methods: the Fourier Transform Based Method and the Maximum Likelihood Estimator (MLE) based on various criteria. Our research can be used to improve speech recognition models by identifying the frequency components of the speech signal more accurately and using this information to reduce or remove the white noise.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Algorithm for Frequency Estimation of Noisy Signals

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Sanvi Gupta,
  • Iniyaa Srinivasan,
  • Teh Kah Chan

摘要

The aim of this project is to develop a deep learning model with improved accuracy and efficiency for estimating the frequency of noisy signals, specifically modulated signals that are corrupted by Gaussian noise. While frequency estimation methods exist, many have limitations like being computationally intensive, inefficient or time consuming. In our model, we created training and testing data, followed by a neural network using Keras’ Sequential API. The neural network is then trained based on the data generated, allowing it to predict the frequency of a corrupted signal. Afterwards, we varied some parameters in our code, namely the number of data samples and number of epochs in order to determine the accuracy at each. We graphed our results, based on which we determined the number of epochs to be 8 and the number of data samples to be 5000. We then carried out a qualitative comparison of our method with two existing methods: the Fourier Transform Based Method and the Maximum Likelihood Estimator (MLE) based on various criteria. Our research can be used to improve speech recognition models by identifying the frequency components of the speech signal more accurately and using this information to reduce or remove the white noise.