Method to identify acoustic sources for anti-submarine warfare
Inventors
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11769045-B2
Publication Date
2023-09-26
Expiration Date
Abstract
A method to detect the presence and location of submarines in a complex marine environment by wavelet denoising, wavelet signal enhancement, by autocorrelation and signal source identification a convolutional neural network.
Core Innovation
The described invention provides a method of identifying an underwater source of a sound based on an acquired signal from listening devices in noisy marine environments. The method decomposes the acquired signal into a plurality of resolution levels using wavelet multi-resolution feature extraction and a denoising process, and integrates at each resolution level to form a denoised signal. The denoised signal is then enhanced by autocorrelation to preserve repetitive signal patterns.
The invention isolates the signal source by extracting signal characteristics from the denoised signal using wavelet-based multi-resolution analysis and by waveform matching. In the disclosure, the isolating includes comparing a waveform by overlapping a mask onto a testing waveform and calculating a difference between the mask and the testing waveform to discern the source from background noise and other components.
For final source detection and classification, the invention processes the denoised signal with a convolutional neural network (CNN). The CNN is used to assign weights and biases, repeat processing until an error rate stabilizes, and detect the underwater source by classification based on spectro-temporal/time-frequency acoustic features. The disclosure also indicates that the system supports distinguishing source classes including submarines, interferers, and noise.
Claims Coverage
The provided claim set includes five independent claim groupings covering multiple aspects of the pipeline: multi-resolution decomposition and denoising, autocorrelation enhancement, CNN-based detection and training, reconstruction/back-projection selection with a trained model, and background-noise discrimination using mask overlap and difference calculations.
Wavelet multi-resolution decomposition to denoise acquired signals
Decomposing the acquired signal into a plurality of resolution levels by dividing into scaling and detail coefficient groups and further decomposing successive resolution levels, and integrating at each resolution level resulting in a denoised signal.
Autocorrelation enhancement of denoised signals for source identification
Enhancing the denoised signal by autocorrelation to support identification of the underwater source by extracting signal characteristics of the denoised signal.
CNN classification pipeline for detecting underwater sources
Processing the denoised signal with a convolutional neural network (CNN) having convolutional levels, activation functions levels, and pooling levels; training the CNN with a training data set; assigning weights and biases to each convolution level; measuring an error of classification by comparing the classification by the CNN to the training data set; adjusting the weights and biases; repeating processing until an error rate stabilizes; and detecting the underwater source by the CNN.
Reconstruction via trained-model selection of a resolution level
Decomposing a waveform signal in a signal domain into a plurality of resolution levels; selecting, with a trained model, a resolution level from the plurality of resolution levels to be reconstructed back to the signal domain; and reconstructing the waveform signal in the selected resolution level back to the signal domain through each previous resolution level to discern a source of the waveform signal from a background noise.
Mask overlap and difference calculation for noise discrimination
Discerning a source of the waveform signal from a background noise by overlapping a mask onto a testing waveform and calculating a difference between the mask and the testing waveform.
Training a model by iterative CNN error-rate stabilization
Training the CNN by processing denoised waveform signals, assigning weights and biases to each artificial neuron of the CNN, adjusting the weights and biases by measuring an error of classification, and repeating processing with the training data set until an error rate stabilizes.
Underwater listening acquisition with autocorrelation identification
Acquiring the waveform signal as an underwater waveform signal from underwater listening devices and identifying an underwater source of the underwater waveform signal with an autocorrelation.
The independent claims cover a composite approach in which wavelet-domain multi-resolution decomposition and integration produce a denoised signal, autocorrelation enhances repetitive signal patterns, and CNN processing detects or classifies the underwater/source signal. Additional independent claims further specify reconstruction back to the signal domain via trained-model selection of a resolution level, include discerning sources from background noise through mask overlap and difference calculations, and cover iterative CNN training until an error rate stabilizes.
Stated Advantages
Not explicitly described in patent.
Documented Applications
Not explicitly described in patent.
Interested in licensing this patent?