Deep Learning ANC System for Nonstationary Noise Suppression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional active noise cancellation (ANC) systems face limitations in effectively addressing nonstationary and high nonlinear noise environments, particularly in construction sites and vehicles, due to their reliance on linear filters and single-sensor feedback systems, which struggle with stability and efficiency.
Innovation Solution
A deep learning-based ANC system, DNoiseNet, is developed, incorporating a convolution layer, atrous scaled convolution modules, a recurrent neural network, and fully connected layers to generate anti-noise signals, capable of processing complex noise patterns and reducing computational costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional linear filters and single-sensor feedback systems are used in ANC, then the system structure is simple, but the performance in suppressing nonstationary and high nonlinear noise is insufficient
Solution Approach 1:
The patent replaces traditional linear filter-based ANC systems with a deep learning-based system that uses neural networks to model and suppress nonstationary and high nonlinear noise. The neural network learns complex noise patterns from data and generates appropriate anti-noise signals, substituting the mechanical linear filtering approach with an intelligent computational model that adapts to varying noise conditions.
Solution Approach 2:
The patent changes the fundamental parameters of the ANC system by transitioning from fixed linear filter coefficients to learnable neural network parameters that can adapt to different noise environments. The system uses multiple sensors and dynamically adjusts its computational model based on real-time noise characteristics, enabling effective suppression of nonstationary noise while maintaining manageable system complexity through structured network architecture.
2Reliability
If deep learning-based ANC system with multiple sensors is used, then the noise cancellation efficiency is improved, but the computational cost and system complexity increase
Solution Approach 1:
The patent segments the deep learning model into multiple specialized components: convolutional layers for extracting spatial features from sensor data, recurrent layers for capturing temporal dependencies, and fully connected layers for integration and anti-noise signal generation. This segmentation allows each component to process specific aspects of the noise signal efficiently, reducing overall computational burden while maintaining high noise cancellation efficiency across multiple sensors.
3Measurement precision
If traditional narrow band noise cancelation systems are used, then the controller design is precise for specific noise ranges, but the system stability becomes problematic and adaptability to different noise environments is limited
Solution Approach 1:
The patent creates a universal ANC system using deep learning models that can handle multiple types of noise environments (stationary, nonstationary, linear, nonlinear) within a single framework. The neural network is trained on diverse noise data and can adapt its feature extraction and anti-noise generation to match the specific characteristics of any given environment, replacing the need for separate precise controllers for different noise bands with one versatile system.
Data Source
AI summary
A computer-implemented method for generating anti-noise using an anti-noise generator to suppress noise from a noise source in an environment comprises processing a sound signal, which is representative of ambient sound including noise, anti-noise and propagation noise from the environment, using a deep learning algorithm configured to generate an anti-noise signal to form anti-noise. The deep learning algorithm comprises a convolution layer; after the convolution layer, a series of atrous scaled convolution modules, wherein each of the atrous scaled convolution modules comprises an atrous convolution, a nonlinear activation function after the atrous convolution, and a pointwise convolution after the nonlinear activation function; after the series of atrous scaled convolution modules, a recurrent neural network; and after the recurrent neural network, a plurality of fully connected layers.