Excitation source identification method based on multi-scale channel attention module

Abstract

The invention discloses an excitation source recognition method based on a multi-scale channel attention module, belongs to the technical field of ship radiation noise excitation source recognition and deep learning algorithms, and can solve the problems of accuracy and rapidity of excitation source recognition in a real ship complex sound source feature environment. The method is mainly technically characterized by comprising the following steps: unifying the scale and format of data through preprocessing, extracting four spectrum features of radiation noise and excitation source signals, carrying out feature fusion based on an MS-CAM multi-scale channel attention module, inputting a fused result into a YOLOv8 classifier to obtain a probability vector, and carrying out feature fusion on the probability vector; the probability is the probability that the radiation noise feature belongs to a certain type of excitation sources, and finally an excitation source identification result is given. By adopting the excitation source identification method based on the multi-scale channel attention module, accurate and rapid identification of the ship radiation noise excitation source can be effectively realized, and the application requirements of ship acoustic control engineering are met.

Description

Technical Field

[0001] This invention relates to the field of excitation source identification of ship vibration and noise characteristics. In particular, it relates to an excitation source identification method based on a multi-scale channel attention module. Background Technology

[0002] Identification of excitation sources for ship vibration and noise characteristics is a crucial aspect of noise control technology and an important component of acoustic feature testing technology. Current traditional noise source analysis methods primarily rely on the physical mechanisms of feature generation and signal processing techniques for analysis. They achieve excitation source identification through comparative analysis. However, the accuracy of this method depends heavily on the analyst's technical skills, and it is time-consuming and inefficient. Therefore, there is an urgent need to research intelligent identification methods for ship vibration and noise characteristics. This research should fully utilize multi-dimensional information features such as time-frequency domain and cepstral domain, employ feature fusion technology, and establish excitation source classification and identification methods to achieve intelligent identification of ship vibration and noise characteristics, thus meeting the application needs of ship acoustic control engineering. Summary of the Invention

[0003] The purpose of this invention is to address the problems of limited accuracy and low analysis efficiency in identifying excitation sources under complex acoustic source characteristics on actual ships. This invention proposes an excitation source identification method based on a multi-scale channel attention module, which can effectively achieve accurate and rapid identification of ship radiated noise excitation sources and meet the application requirements of ship acoustic control engineering.

[0004] To achieve the above-mentioned objectives of this invention, the core of this invention is the design of an excitation source identification method based on the YOLOv8 algorithm using the MS-CAM multi-scale channel attention module, which is grounded in deep learning theory. By preprocessing data to unify scale and format, four spectral features of radiated noise and excitation source signals are extracted. Then, feature fusion is performed based on the MS-CAM multi-scale channel attention module. The fused result is input into the YOLOv8 classifier to obtain a probability vector, which represents the probability that the radiated noise feature belongs to a certain type of excitation source. Finally, the excitation source identification result is given, achieving accurate and rapid identification of ship radiated noise excitation sources.

[0005] The present invention provides a method for identifying excitation sources based on a multi-scale channel attention module, comprising the following steps:

[0006] Step 1: Data preprocessing. Framing and filtering methods are used to highlight the characteristics of the excitation source to be analyzed, and the raw data is standardized to a uniform scale and format. Specific steps are as follows:

[0007] (a) Feature extraction requires data of the same length. The main purpose of framing is to divide the continuous signal into multiple identical short segments so that each segment can be regarded as an approximately stationary signal. The frame length is selected as 10 seconds and the frame shift is 0.02 times the frame length.

[0008] (b) After framing, each frame is weighted by a window function and smoothed by the window function. The Hamming window is selected as the window function. This window function will affect the spectral leakage and resolution, and can effectively reduce the leakage phenomenon.

[0009] (c) Filters are used to remove irrelevant frequency components from the signal to highlight the key information to be analyzed, making the model easier to identify and classify. The filter range is selected based on the spectral characteristics of the radiated noise to be analyzed. For line spectrum analysis of LOFAR, MFCC, and GFCC spectra, the filter bandpass range in this invention is ±3Hz near the analysis characteristic frequency to highlight the frequency components of the signal. For broadband analysis, the filter range is set to the bandwidth to be analyzed. For DEMON spectra, the filter range is set to 5Hz to 5000Hz. Then, Hilbert transform is used for further feature extraction to meet the analysis requirements.

[0010] Step 2: Feature extraction. Considering the characteristics of underwater acoustic signals, four commonly used features for underwater acoustic target recognition are selected for extraction: LOFAR spectral features, DEMON spectral features, Mel frequency cepstral coefficients (MFCC) features, and Gamma-ton filter cepstral coefficients (GFCC) features.

[0011] Step 3: Feature Fusion. The four feature spectra from each frame are fused. First, a Transformer is used to convert the spectra into a standardized feature matrix. Then, a convolutional neural network with an attention mechanism is employed for feature fusion. The attention mechanism assigns different weights to features to highlight important features beneficial for classification, thereby helping the neural network efficiently complete the feature extraction and fusion tasks. The specific steps are as follows:

[0012] (a) Image loading and preprocessing: Read LOFAR, DEMON, MFCC and GFCC spectra, use a resize function and adaptive average pooling layer to resize the image to the input image with a consistent size, and use the Transformer feature transformation module to normalize the pixel values ​​of the resized image to a uniform range.

[0013] (b) Image format conversion: Each spectral image is converted into a tensor form, defined as (C, H, W), where C is the number of channels, set to 3 (corresponding to standard color), and H and W are the height and width of the image, respectively. These tensors are then stacked along a new dimension, adding a new dimension in front of each input tensor, resulting in a stacked shape of (4, C, H, W). Here, 4 indicates that there are four types of spectral images.

[0014] (c) Using the MS-CAM multi-scale channel attention module, the randomization seed of the iterative feature module is fixed by initializing the feature fusion module to ensure that the fusion result is the same each time, and the feature fusion result based on multi-scale channel attention is output in image format.

[0015] Step 4: Classifier design. Based on deep learning algorithms, the YOLOv8 classification model in the one-stage detection method is adopted. It uses convolutional layers combined with the ReLU activation function, and then outputs the probability of classifying and recognizing the image through the probability vector output layer (Soffmax layer), that is, the probability that the radiated noise feature belongs to a certain type of excitation source.

[0016] Step 5, Training and Recognition: The fused feature vector obtained in Step 3 is input into the YOLOv8 classifier to obtain a probability vector. Each component in this vector corresponds to the probability that the data belongs to a certain category, that is, the probability that the radiated noise feature belongs to a certain type of excitation source. The training set is input into the initial classification model to obtain the optimal parameters and generate the recognition model. The validation set is input into the recognition model to give the recognition result, realizing accurate and fast identification of ship radiated noise excitation sources.

[0017] The advantage of this invention is that, based on deep learning theory, it proposes an excitation source identification method based on the YOLOv8 algorithm of the MS-CAM multi-scale channel attention module, which can effectively achieve accurate and rapid identification of ship radiated noise excitation sources and meet the application requirements of ship acoustic control engineering. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall scheme of the excitation source recognition method based on the multi-scale channel attention module in this invention.

[0019] Figure 2 yes Figure 1 Structure diagram of the multi-scale channel attention module.

[0020] Figure 3 yes Figure 1 Diagram of the YOLOv8 network structure.

[0021] Figure 4 This is a curve showing the change in training model loss function and accuracy, as illustrated in an embodiment of the present invention. Detailed Implementation

[0022] The present invention provides a detailed description of an excitation source identification method based on a multi-scale channel attention module, with reference to specific embodiments. The dataset used in this embodiment includes radiated noise and vibration data from shipboard machinery, bearing bases, and hull parts. Each data segment is 16 seconds long. The radiated noise data represents the target to be identified, while the vibration data from the machinery, bearing bases, and hull parts represent the excitation source category.

[0023] Reference Figure 1-4 The specific implementation of the excitation source identification method based on the multi-scale channel attention module of the present invention is as follows:

[0024] Step 1: Perform data preprocessing on each sample in the dataset, including frame segmentation, filtering, windowing, etc., to standardize the original data to a uniform scale and format.

[0025] In frame splitting, the frame length is set to 10 seconds, and the frame shift is 0.02 times the frame length.

[0026] After framing, each frame is weighted by a window function, and then smoothed using the window function. The Hamming window is selected as the window function.

[0027] Filters are used to remove irrelevant frequency components from the signal to highlight the key information to be analyzed. The filter range is selected based on the spectral characteristics of the radiated noise to be analyzed. For line spectrum analysis of LOFAR, MFCC, and GFCC spectra, the filter bandpass range is ±3Hz around the characteristic frequency of the analysis to highlight the frequency components of the signal. For broadband analysis, the filter range is set to the bandwidth to be analyzed. For DEMON spectra, the filter range is set to between 5Hz and 5000Hz.

[0028] Step 2: Feature extraction is performed on the preprocessed samples. This invention extracts a total of 4 features, namely LOFAR spectral features, DEMON spectral features, Mel frequency cepstral coefficients (MFCC) features, and gamma-pass filter cepstral coefficients (GFCC) features.

[0029] Step 3: Perform feature fusion on the four feature spectra extracted from each frame of signal. First, convert them into a unified standard feature matrix using a Transformer, and then perform feature fusion using a convolutional neural network with an attention mechanism.

[0030] The image is resized to a uniform size using a resize function and an adaptive average pooling layer, and the pixel values ​​of the resized image are normalized to a uniform range using the Transformer feature transformation module.

[0031] Each spectral image is converted into tensor form, and each tensor is defined as (C, H, W), where C is the number of channels, set to a value of 3 related to standard colors, and H and W are the height and width of the image, respectively. These tensors are then stacked along a new dimension, adding a new dimension in front of each input tensor, so the stacked result has dimensions of (4, C, H, W).

[0032] A multi-scale channel attention module is employed to fuse features from various spectral tensors. This module includes local and global attention modules, and the feature fusion result is finally output using the sigmoid function. The calculation method for local attention is as follows:

[0033] M L (X)=σ(F L (X)) (1)

[0034] In the formula, M L Here, σ represents the weights of the local attention features, σ is the ReLU activation function, and F... L This is the learning function for a local attention neural network.

[0035] The global attention calculation method is as follows:

[0036] M G (X)=σ(F G [AvgPool(X)]) (2)

[0037] In the formula, M G Here, σ represents the weights of the local attention features, σ is the ReLU activation function, and F... G Here, AvgPool is the learning function for a local attention neural network, and AvgPool is the average pooling function.

[0038] The output of the multi-scale channel attention module can be expressed as:

[0039]

[0040] In the formula, σ s M is the sigmoid function. L M represents the weights of local attention features. G These are the weights for local attention features.

[0041] Step 4: Using the YOLOv8 classification model, the probability of classifying and recognizing the image is output through a convolutional layer combined with the ReLU activation function and then through a probability vector output layer (Softmax layer).

[0042] Step 5: Input the fused feature vector obtained in Step 3 into the YOLOv8 classifier to obtain a probability vector. Each component in this vector corresponds to the probability that the data belongs to a certain category, that is, the probability that the radiated noise feature belongs to a certain type of excitation source.

[0043] Eighty percent of the feature-fused map samples were used as the training set, and the remaining 20% ​​as the test set. A YOLOv8 classification model was used to train the training set. In the final output layer, a softmax function was used to provide the probability distribution for each category. The training effect was verified by comparing the training model's loss function and accuracy. After training, a recognition model was generated. The validation set was then input into the recognition model to provide the stimulus source recognition result.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the general process of implementing the present invention and are not intended to limit it. Although the present invention has been described in detail, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the present invention, and are all covered within the scope of the claims of the present invention.

Claims

1. A method for identifying excitation sources based on a multi-scale channel attention module, characterized in that... Includes the following steps: Step 1: Data preprocessing. Framing and filtering methods are used to highlight the characteristics of the excitation source to be analyzed, and the raw data is standardized to a uniform scale and format. Step 2: Feature extraction. Four features are selected for extraction: LOFAR spectral features, DEMON spectral features, Mel frequency cepstral coefficients (MFCC) features, and Gamma-ton filter cepstral coefficients (GFCC) features. Step 3: Feature fusion. The four feature spectra of each frame signal are fused. First, the Transformer is used to convert them into a unified standard feature matrix. Then, a convolutional neural network with attention mechanism is used to perform feature fusion. The attention mechanism highlights important features that are beneficial to classification by assigning different weights to the features, thereby helping the neural network to efficiently complete the feature extraction and feature fusion tasks. Step 4: Classifier design. Based on deep learning algorithms, the YOLOv8 classification model is adopted. It uses convolutional layers combined with the ReLU activation function, and then outputs the classification and recognition probability of the image through a probability vector output layer. Step 5: Training and Recognition. Input the training set into the initial classification model to obtain the optimal parameters, generate the recognition model, input the validation set into the recognition model, and give the recognition result.

2. The method for identifying excitation sources based on a multi-scale channel attention module according to claim 1, characterized in that, Step one specifically involves: (a) Feature extraction requires dividing the continuous signal into multiple identical short segments by framing the data of the same length. The frame length is selected as 10 seconds and the frame shift is 0.02 times the frame length. (b) After framing, each frame is windowed, with a Hamming window chosen as the window function. This window function affects spectral leakage and resolution, and can effectively reduce leakage. (c) Use a filter to remove irrelevant frequency components from the signal and highlight the key information to be analyzed. The filter bandpass range is ±3Hz around the characteristic frequency of the analysis to highlight the frequency components of the signal. For broadband analysis, the filter range is set to the bandwidth to be analyzed.

3. The method for identifying excitation sources based on a multi-scale channel attention module according to claim 1, characterized in that, Step three specifically involves: (a) Image loading and preprocessing: Read LOFAR, DEMON, MFCC and GFCC spectra, use a resize function and adaptive average pooling layer to resize the image to the input image with a consistent size, and use the Transformer feature transformation module to normalize the pixel values ​​of the resized image to a uniform range. (b) Image format conversion: convert each spectrum into tensor form; (c) Using the MS-CAM multi-scale channel attention module, the randomization seed of the iterative feature module is fixed by initializing the feature fusion module to ensure that the fusion result is the same each time, and the feature fusion result based on multi-scale channel attention is output in image format.

Images