Friday, November 20, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Deep Convolutional Network Accurately Estimates Harmonic Frequencies Across Full Angular Range in Complex Noise

November 20, 2026
in Science News
Cassandra Pierce
By Cassandra Pierce Scienmag Editorial Profile - Systems Neuroscience
Reading Time: 7 mins read
0
Deep Convolutional Network Accurately Estimates Harmonic Frequencies Across Full Angular Range in Complex Noise

Deep Convolutional Network Accurately Estimates Harmonic Frequencies Across Full Angular Range in Complex Noise

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

A research team comprising Shiyong Yang, Bo Zhang, and Jinlong Jiang has developed a supervised deep learning framework designed to resolve discrete harmonic frequencies from complex sample data contaminated simultaneously by multiplicative and additive noise. Published in the Journal on Wireless Communications and Networking, the study formulates frequency estimation across discrete-time multi-component signals as a multi-label classification problem solved by a specialized one-dimensional convolutional neural network. In traditional signal processing, estimating harmonic parameters under harsh interference conditions is fundamental to operating domains such as radar, sonar, radio astronomy, wireless communications, and nuclear magnetic resonance. While classical estimators rely on statistical formulations that degrade under non-stationary interference or require prior knowledge of noise characteristics, the researchers demonstrated that their neural architecture accurately extracts multi-frequency parameters directly from raw measurements across the complete frequency range from negative pi to positive pi without numerical parameter tuning.

Harmonic signals observed in real-world propagation environments frequently encounter combined noise mechanisms rather than simple additive disturbances. In underwater acoustic environments, for example, fluctuations caused by propagation media, changing target orientation, and scatterer interference introduce severe multiplicative noise that modulates the underlying waveforms. Because multiplicative noise destroys signal stationarity, the resulting composite waveforms can no longer be treated with standard stationary processing methods. Instead, harmonics immersed in joint multiplicative and additive noise exhibit cyclostationary characteristics. Although previous theoretical treatments developed cyclic statistics to exploit this underlying cyclostationary nature, conventional cyclic estimators present operational drawbacks. Specifically, cyclic mean estimators require nonzero-mean multiplicative noise, while third-order cyclic statistics restrict the multiplicative disturbance to non-Gaussian distributions. Furthermore, traditional cyclic correlation algorithms are mathematically restricted to estimating frequencies within the narrow half-spectrum from negative pi over two to positive pi over two when handling zero-mean multiplicative noise, while suffering from Rayleigh resolution limits and spurious pseudo-peaks caused by Fast Fourier Transform peak searching.

To overcome these mathematical and computational bottlenecks, the authors designed a learning-based formulation that bypasses iterative optimization routines during inference. Deep learning frameworks present key operational strengths over classical optimization techniques, notably because a converged neural network computes solutions via forward-pass multiplications and additions without requiring parameter adjustments for specific input instances. The authors’ formulation models a discrete-time signal containing a known number of distinct harmonic components, designated as K, observed over T sample snapshots. In their formal problem formulation, the discrete-time K-component harmonic process comprises constant distinct frequencies lying in the interval excluding zero between negative pi and positive pi, accompanied by deterministic constant phases within the half-open interval from negative pi to positive pi. Both the multiplicative noise modulating each component and the background additive disturbance are modeled as zero-mean, mutually independent, stationary, and mixing random processes whose higher-order cumulants are absolutely summable.

A core architectural innovation of the proposed framework is a three-channel data formatting strategy tailored to overcome the inability of standard neural networks to process complex-valued numbers directly. Given raw complex observation sequences of length T, the researchers construct a real-valued input matrix of size T by three. Rather than discarding phase relationships or relying solely on Euclidean coordinate projections, the network splits each complex sample into three explicit parallel channels: the real component, the imaginary component, and the instantaneous phase angle. This multi-channel configuration maps the T temporal observations into a structured three-dimensional tensor that exposes multi-dimensional signal attributes directly to the feature extraction stages of the learning pipeline.

The frequency estimation problem is mapped onto a discrete frequency grid to establish a multi-label classification objective. For a selected frequency resolution step, denoted as mu, the continuous angular frequency domain between negative pi and positive pi is partitioned into a discrete set containing two Q plus one distinct grid points, where Q represents the integer floor of pi divided by mu. In their primary experimental implementation, the authors selected a resolution step of 0.05, yielding an evaluation grid of 125 discrete points spanning from negative 3.10 to positive 3.10 radians. Each training instance containing K active frequencies is converted into a binary label vector of length 125, configured as a K-hot vector where entries corresponding to active grid coordinates are set to one while all unselected positions remain zero. This mapping enables the network to treat the identification of multiple harmonic components as concurrent binary decisions across the discrete frequency coordinates.

The deep neural network architecture consists of seven functional layers, comprising three standard one-dimensional convolutional layers followed by four fully connected layers. Each convolutional layer contains 64 filters featuring a kernel dimension of five by one, operated with a unit stride of one and without padding. Every convolutional operation is followed immediately by an element-wise rectified linear unit activation function to introduce nonlinear feature transformations. Following feature extraction through the convolutional stages, the representations pass through three hidden fully connected layers, each containing 512 neurons activated by rectified linear units. The final output layer consists of two Q plus one neurons, which evaluates to 125 output nodes for the 0.05 resolution grid. This final layer is coupled to an element-wise sigmoid activation function that maps each output entry to an independent probability value between zero and one. A sigmoid activation was specifically chosen over normalized softmax operators because K distinct frequency labels must simultaneously register values approaching unity without mutually suppressing each other.

Training the seven-layer network was conducted offline in a supervised learning scheme using synthetic datasets generated across varied operational conditions. Network parameters were optimized by minimizing a binary cross-entropy reconstruction loss through back-propagation, treating each frequency bin as an independent classification decision. For a model configured with two harmonic components, all possible pairwise combinations across the 125-point grid were evaluated, generating 7,750 unique frequency pairs. To provide comprehensive noise exposure, the researchers synthesized 300 independent realization trials for each frequency pair across four distinct signal-to-noise ratios: 0 dB, 5 dB, 10 dB, and 15 dB. This combinatorial synthesis yielded a massive aggregate training corpus of 9.3 million labeled training examples. The dataset was partitioned into a 90 percent training split and a 10 percent validation set, optimized using the Adam algorithm with a fixed learning rate of 0.0001, a batch size of 128, and a training duration of 100 epochs implemented in PyTorch on an Ubuntu 22.04 computing environment powered by an Intel Xeon E5-2620 central processing unit and an NVIDIA GeForce RTX 3080 Ti graphics card.

The complete structural parameter count of the trained network totaled approximately 6.80 million trainable weights, with the vast majority distributed across the dense linear transformations. Specifically, the first one-dimensional convolutional layer accounts for 1,152 parameters, while the second and third convolutional stages each incorporate 24,576 parameters. The initial fully connected layer represents the primary parameter concentration, requiring 6,160,896 weights to interface with the flattened convolutional feature maps, followed by 262,656 parameters in the second dense layer, 262,656 parameters in the third dense layer, and 64,125 parameters connecting to the 125-neuron output stage. During operational inference, the network takes a real-valued observation matrix across 200 temporal snapshots, evaluates the forward pass, and isolates the K highest probability peaks from the predicted vector, assigning the corresponding discrete grid frequencies as the estimated signal components.

The authors evaluated the estimation accuracy and robustness of their network through comprehensive Monte Carlo simulation trials against the classical cyclic correlation algorithm, using the empirical root-mean-square error across estimated frequencies as the primary performance benchmark. In an initial experiment examining two off-grid frequencies restricted within the conventional interval from negative pi over two to positive pi over two at a signal-to-noise ratio of 5 dB and 200 snapshots, the signals were separated by an angular distance of 0.25 radians while sweeping the primary frequency from negative 1.552 to positive 1.302 in steps of 0.05. Because these tested frequencies were deliberately placed off-grid, neither frequency existed within the training set. Under these conditions, the estimation errors of the neural network remained tightly bounded within the interval between zero and 0.005, achieving an overall empirical root-mean-square error of 0.0021. In contrast, the cyclic correlation algorithm produced errors spanning up to 0.05, including severe outlier anomalies, resulting in a significantly higher root-mean-square error of 0.0126.

A second targeted experiment demonstrated the network’s capability to overcome the fundamental half-spectrum limitation inherent to classical cyclic correlation estimators. Testing two off-grid frequencies across the wider spectrum from negative pi to positive pi, the primary frequency was adjusted from negative 3.102 to positive 2.852 with a fixed component separation of 0.25 radians. Under these conditions, the cyclic correlation method suffered total failure due to spectral aliasing and could not resolve the frequencies. The proposed one-dimensional convolutional network, however, maintained stable estimation fidelity across the full operational bandwidth, recording an empirical root-mean-square error of 0.0087. Additional comparative evaluations across signal-to-noise ratios from 0 dB to 15 dB indicated that while both algorithms exhibited comparable accuracy at 0 dB, the neural network consistently outperformed cyclic correlation across intermediate and high signal-to-noise conditions. Similarly, evaluating frequency separations from 0.05 to 0.50 radians over 1,000 Monte Carlo repetitions confirmed that the deep learning model attained superior root-mean-square error performance across most separation intervals.

To rigorously substantiate the architectural choice of their three-channel input representation, the researchers executed an ablation study comparing seven distinct input configurations across identical network backbones, training hyperparameters, and evaluation datasets. Models trained solely on isolated single channels—using only the real component, only the imaginary component, or only the phase angle—exhibited large estimation errors and inferior accuracy, confirming that single-dimensional representations cannot sufficiently capture the underlying complex harmonic structure. Dual-channel configurations substantially improved estimation fidelity, demonstrating that Euclidean and phase coordinates provide complementary information. Ultimately, the complete three-channel architecture combining real, imaginary, and phase inputs achieved the lowest root-mean-square error and highest overall estimation accuracy among all tested groups, verifying that the multi-channel structure eliminates feature deficiency without introducing redundant overhead.

While the simulation findings confirm enhanced precision and resilience against complex noise disturbances, the authors identified defined scope conditions and limitations within the current study. The present architecture assumes that the exact number of active harmonic components, K, is known prior to inference, leaving the simultaneous estimation of component quantity and frequency coordinates to future investigations. Furthermore, the reported empirical performance gains were validated using synthetic Gaussian zero-mean noise processes and simulated signal models rather than experimental in-situ acoustic recordings. Nevertheless, by demonstrating that deep convolutional neural networks can eliminate spectral aliasing restrictions and overcome Rayleigh resolution limits, the study establishes a viable machine learning foundation for robust parameter estimation in non-stationary harmonic environments.

Subject of Research: Deep learning approaches for harmonic frequency estimation under multiplicative and additive noise

Article Title: Frequency estimation of harmonics in complex noise based on deep learning

Article References: Yang, S., Zhang, B., & Jiang, J. (2026). Frequency estimation of harmonics in complex noise based on deep learning. Journal on Wireless Communications and Networking, 2026(1), Article 122. https://doi.org/10.1186/s13638-026-02663-0

Image Credits: AI Generated

DOI: 10.1186/s13638-026-02663-0

Keywords: frequency estimation, harmonic signals, deep learning, convolutional neural networks, multiplicative noise, additive noise, signal processing, cyclic correlation, multi-label classification, cyclostationary signals, signal-to-noise ratio, root-mean-square error, supervised learning, ablation study, channel feature extraction, Monte Carlo simulation, underwater acoustics

Cite Scienmag News

Cassandra Pierce. (November 20, 2026). Deep Convolutional Network Accurately Estimates Harmonic Frequencies Across Full Angular Range in Complex Noise. Scienmag. https://scienmag.com/deep-convolutional-network-accurately-estimates-harmonic-frequencies-across-full-angular-range-in-complex-noise/

Cassandra Pierce. "Deep Convolutional Network Accurately Estimates Harmonic Frequencies Across Full Angular Range in Complex Noise." Scienmag, 20 November 2026, https://scienmag.com/deep-convolutional-network-accurately-estimates-harmonic-frequencies-across-full-angular-range-in-complex-noise/. Accessed 20 November 2026.

Cassandra Pierce. "Deep Convolutional Network Accurately Estimates Harmonic Frequencies Across Full Angular Range in Complex Noise." Scienmag. November 20, 2026. https://scienmag.com/deep-convolutional-network-accurately-estimates-harmonic-frequencies-across-full-angular-range-in-complex-noise/

Tags: ablation studyadditive noisechannel feature extractionconvolutional neural networkscyclic correlationcyclostationary signalsdeep learningfrequency estimationharmonic signalsMonte Carlo simulationmulti-label classificationmultiplicative noiseroot mean square errorSignal Processingsignal-to-noise ratiosupervised learningunderwater acoustics
Share26Tweet16
Previous Post

Low-Voltage Aluminum MEMS Hot Arm Actuator Delivers Fast Microscale Lateral Displacement

Next Post

Dual-Action Injectable Hydrogel Suppresses Pancreatic Tumors and Seals Postoperative Fistulas in Rodents

Related Posts

Climate and Glacier Geometry Dominate Ice Loss at Lake-Terminating Glaciers
Climate

Climate and Glacier Geometry Dominate Ice Loss at Lake-Terminating Glaciers

November 20, 2026
Exercise Prescriptions and Nearby Facilities Increase Stated Willingness in China
Medicine

Exercise Prescriptions and Nearby Facilities Increase Stated Willingness in China

November 20, 2026
Microsensors and Remote Sensing Track Spring Smoke Haze Across Northern Thailand
Athmospheric

Microsensors and Remote Sensing Track Spring Smoke Haze Across Northern Thailand

November 20, 2026
Review Assesses NINJ1 Oligomerization and Membrane Rupture in Regulated Cell Death
Biology

Review Assesses NINJ1 Oligomerization and Membrane Rupture in Regulated Cell Death

November 20, 2026
Federated Digital Twins Speed Edge Healthcare Monitoring While Protecting Patient Data
Technology and Engineering

Federated Digital Twins Speed Edge Healthcare Monitoring While Protecting Patient Data

November 20, 2026
Early Serum Uric Acid Decline Links to Survival After Cardiac Arrest
Medicine

Early Serum Uric Acid Decline Links to Survival After Cardiac Arrest

November 20, 2026
Next Post
Dual-Action Injectable Hydrogel Suppresses Pancreatic Tumors and Seals Postoperative Fistulas in Rodents

Dual-Action Injectable Hydrogel Suppresses Pancreatic Tumors and Seals Postoperative Fistulas in Rodents

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Climate and Glacier Geometry Dominate Ice Loss at Lake-Terminating Glaciers
  • Exercise Prescriptions and Nearby Facilities Increase Stated Willingness in China
  • Microsensors and Remote Sensing Track Spring Smoke Haze Across Northern Thailand
  • Review Assesses NINJ1 Oligomerization and Membrane Rupture in Regulated Cell Death

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading