High-Concentration Time-Frequency Representation and Instantaneous Frequency Estimation of Frequency-Crossing Signals

Hui Li, Xiangxiang Zhu, Yingfei Wang, Xuanfeng Cai, Zhuosheng ZhangSensors 2025 · DOI

TL;DR

This paper introduces a novel deep learning framework for analyzing frequency-crossing signals, which are common in many natural and engineered systems. The approach combines a U-Net-based network for high-resolution time-frequency (TF) representation with a separate network for accurately estimating the instantaneous frequency (IF) of individual signal components, outperforming existing methods in concentration and noise robustness.

What Problem Does This Paper Address?

Many real-world signals, such as those found in sonar, radar, and biomedical applications, are non-stationary and composed of multiple oscillatory components. A particular challenge arises when these components' instantaneous frequencies (IFs) intersect or "cross" over time. Accurately representing these signals in the time-frequency (TF) domain and precisely estimating the IF of each component is crucial for understanding the underlying system and predicting signal behavior. Traditional TF analysis methods (like STFT and WVD) struggle with low resolution and cross-term interference, especially for frequency-crossing signals. While techniques like synchrosqueezing improve concentration, they often require distinct component separation and can be sensitive to noise. Existing deep learning approaches have focused on improving TF representation but often neglect the critical aspect of IF separation and estimation for these complex signals.

Methods

The proposed solution consists of two main deep learning networks:

  1. High-Concentration TF Representation Network: This network is built upon the U-Net architecture, known for its effectiveness in image segmentation. It takes a TF map as input and uses a series of convolutional, pooling, upsampling, and deconvolutional layers to generate a high-resolution, concentrated TF representation. A key innovation is the use of feature fusion (concatenation) between encoder and decoder layers, which helps preserve low-resolution details while enhancing high-resolution features, leading to better noise immunity and energy concentration. This approach avoids the cross-term interference issues associated with methods like WVD.
  2. IF Separation and Estimation Network: This network is designed to address the challenge of separating and estimating the IFs of individual components within a frequency-crossing signal. It comprises two stages:
  3. Component Number Discrimination Network: A CNN is used to automatically determine the number of components in the signal based on its TF representation.
  4. P-Branch Separation Network: Once the number of components (P) is known, a P-branch network is constructed. Each branch is responsible for separating and estimating the IF of one component. This network then outputs an IF estimation matrix.

Datasets

The paper utilizes synthetically generated datasets for training and evaluation.

  • Training Data: 10,000 two-component and three-component signals were generated using specific mathematical models. These signals were then processed using STFT with a Hanning window to create 128x256 TF images. 9000 images were used for training, and 1000 for validation.
  • Ideal TF Representations: For training the TF representation network, ideal TF representations were used as ground truth.
  • Real IF Information: For training the IF separation and estimation network, real IF information was used as ground truth.
  • Noise Levels: The synthetic signals were generated with 10 different Signal-to-Noise Ratios (SNRs), ranging from -10 dB to 35 dB, to test noise robustness.
  • Real-World Data: A bat echolocation signal from Rice University was used for practical application testing.

Key Insights and Ideas

  • Feature Fusion is Key for Robustness: The paper highlights that concatenating features from different levels of the U-Net architecture (encoder to decoder) is crucial. This "feature fusion" helps retain important low-resolution information while enhancing high-resolution details, leading to more concentrated TF representations and better noise immunity compared to models without this fusion.
  • Separating TF Representation and IF Estimation: The authors propose a two-stage deep learning approach, recognizing that high-resolution TF representation and accurate IF estimation are distinct but complementary tasks. This modular design allows for specialized optimization of each stage.
  • Adaptive Component Number Determination: The inclusion of a component number discrimination network makes the IF separation process adaptive. This is a significant advantage over methods that require prior knowledge of the number of signal components.
  • U-Net's Versatility Beyond Image Segmentation: The paper demonstrates the successful application of the U-Net architecture, traditionally used for image segmentation, to the domain of time-frequency signal analysis, showcasing its adaptability to different data modalities.

Key Results

  • Improved TF Concentration: The proposed high-concentration TF representation network achieved significantly more concentrated TF representations compared to STFT, Synchrosqueezing Transform (SST), and a standard CNN-based approach. This was quantified by a lower Rényi entropy value (10.1 vs. 11.4 for SST, 13.5 for STFT).
  • Reduced Mean Absolute Error (MAE): The proposed method also showed a lower MAE (0.372 vs. 0.409 for CNN, 0.617 for SST), indicating a closer match to the ideal TF representation.
  • Noise Robustness: Experimental results using MSE showed that the proposed approach achieved smaller MSEs than STFT, SST, and CNN across a wide SNR range (-4 dB to 20 dB), demonstrating its superior ability to characterize signals under noisy conditions.
  • Accurate IF Estimation: The IF separation and estimation network successfully determined the number of components and extracted the IFs of frequency-crossing signals, even in low SNR environments (e.g., -5 dB).
  • Faster Processing for TF Representation: While deep learning networks generally increase computational complexity, the proposed TF representation network required less processing time per batch (361/step) compared to the CNN approach (413/step).

Limitations and Future Work

  • Deeper Networks and Information Loss: The authors acknowledge that while deeper network architectures might yield even more concentrated TF representations, they risk losing signal information, as demonstrated by an example from another study. Optimizing network depth for both concentration and information preservation remains an open challenge.
  • Generalization Capability: While the proposed model shows less interference than CNN on frequency-uncrossed signals, both networks exhibited some representation bias. Improving the generalization capability of these networks to a wider range of signal types is a significant area for future research.
  • Strong Noise Interference: The paper notes that, similar to other methods, the proposed approach still encounters substantial errors in IF estimation when strong noise is present. Further research is needed to address this limitation.
  • Adaptive Parameter Tuning: The paper implies that current methods may require adaptive parameter tuning. While the proposed deep learning approach automates much of this, further exploration into adaptive learning strategies could be beneficial.

Summary updated

Abstract

Frequency-crossing signals are widely found in nature and various engineering systems. Currently, achieving high-resolution time-frequency (TF) representation and accurate instantaneous frequency (IF) estimation for these signals presents a challenge and is a significant area of research. This paper proposes a solution that includes a high-concentration TF representation network and an IF separation and estimation network, designed specifically for analyzing frequency-crossing signals using classical TF analysis and U-net techniques. Through TF data generation, the construction of a U-net, and training, the high-concentration TF representation network achieves high-resolution TF characterization of different frequency-crossing signals. The IF separation and estimation network, with its discriminant model, offers flexibility in determining the number of components within multi-component signals. Following this, the separation network model, with an equal number of components, is utilized for signal separation and IF estimation. Finally, a comparison is performed against the short-time Fourier transform, synchrosqueezing transform, and convolutional neural network. Experimental validation shows that our proposed approach achieves high TF concentration, exhibiting robust noise immunity and enabling precise characterization of the time-varying law of frequency-crossing signals.
BibTeX
@inproceedings{li2025high,
  author = {Hui Li and Xiangxiang Zhu and Yingfei Wang and Xuanfeng Cai and Zhuosheng Zhang},
  title = {{High-Concentration Time-Frequency Representation and Instantaneous Frequency Estimation of Frequency-Crossing Signals}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25072030},
}
Figure