A ship radiation noise recognition method based on semi-supervised collaborative learning

Abstract

The application discloses a ship radiation noise recognition method based on semi-supervised collaborative learning, frequency domain transformation is carried out on normalized audio through a Mel filter bank to obtain a Mel spectrogram of noise energy distribution, and the input of a model is audio and the Mel spectrogram. The data of each branch is divided into an unlabeled training set, a labeled training set and a test set. The unlabeled data is marked with a pseudo label and participates in collaborative learning together with the labeled data, a feature fusion parameter is dynamically adjusted, and final output classification is completed through a full connection layer. The method effectively improves the accuracy and generalization ability of ship radiation noise recognition in a small sample scene, and reduces the dependence of classification and recognition on large-scale labeled data.

Description

Technical Field

[0001] This invention belongs to the field of target classification technology, specifically relating to a method for identifying ship radiated noise based on semi-supervised collaborative learning. Background Technology

[0002] Classification and identification of ship radiated noise (SRN) is one of the core technologies in the field of underwater acoustic signal processing. It achieves automatic ship identification and classification by analyzing the characteristics of ship radiated noise in water. SRN mainly consists of three parts: mechanical noise, propeller noise, and hydrodynamic noise, containing rich feature information. However, the complex marine environment presents numerous challenges to high-precision classification and identification. With continuous technological breakthroughs, machine learning algorithms have been widely applied to SRN identification. The application of this method has further promoted technological change in this field, realizing intelligent high-precision classification and identification of SRN.

[0003] The classification and recognition of SRNs has evolved from simple feature analysis to complex and intelligent deep feature extraction. Early research methods primarily relied on manual extraction of surface features of radiated noise, but this method struggled to fully reflect the complex feature structure of ship noise. With technological advancements, the introduction of machine learning algorithms improved the automation level of recognition. Recently, the rise of deep learning algorithms has practically propelled the intelligent development of SRN classification and recognition. In general, the classification and recognition of ship radiated noise mainly falls into two categories: (1) based on traditional signal processing combined with classical classifiers; (2) based on intelligent feature extraction combined with deep learning.

[0004] Traditional machine learning methods for ship identification and classification mainly consist of the following three processes: preprocessing the target signal, extracting features manually using signal processing techniques, and finally feeding the signal into a classifier to complete the identification and classification task of the SRN.

[0005] i. Preprocessing; Currently, mainstream signal preprocessing methods mainly include the following: traditional time-domain denoising algorithms and noise suppression based on signal statistical features or probabilistic models. 1. Regarding traditional denoising methods, these mainly include wavelet analysis and Variational Mode Decomposition (VMD). Wavelet analysis, due to its simple principle and ease of implementation, has become an important method in the field of signal denoising. However, due to deviations in the selection of fixed thresholds for soft and hard thresholds, the reconstructed signal is prone to distortion problems such as edge blurring. For VMD, compared to traditional Empirical Mode Decomposition (EMD), VMD has the advantages of fast decomposition speed and high accuracy, and has been relatively maturely applied in signal denoising. However, VMD requires determining the number of decomposition modes and the penalty factor during signal decomposition, thus parameter selection becomes a challenge for denoising; relying solely on empirical selection will lead to poor signal denoising results. It is evident that the problem of parameter selection is prevalent in both of these methods, which is a major reason restricting signal denoising. 2. Statistical and signal estimation algorithms mainly include Wiener filtering and Kalman filtering. Wiener filtering has the advantages of low computational cost, ease of implementation, and good smoothing properties, but it is difficult to use in complex and non-stationary noisy environments. Kalman filtering is suitable for real-time estimation of dynamic signals, but its computational complexity is relatively high, resulting in significant computational overhead. In summary, traditional preprocessing methods can effectively suppress noise to a certain extent, but certain technical bottlenecks still exist.

[0006] ii. Feature extraction; Currently, the main feature extraction methods can be categorized into the following types: time-domain, frequency-domain, and combined time-frequency-domain feature extraction. 1. Time-domain feature extraction methods primarily analyze the original time series or obtain features through envelope analysis. These methods are mainly characterized by their simplicity, ease of implementation, and good performance on relatively stationary signals. However, this also indicates their limited adaptability to non-stationary signals and their sensitivity to noise. 2. Frequency-domain feature extraction methods mainly include Fast Fourier Transform (FFT) and power spectral analysis. These methods can extract parameters such as power spectral density and spectral centroid, but this further indicates their lack of relevant time-domain features. 3. Combined time-frequency-domain methods mainly include Short-Time Fourier Transform (SFT), EMD, and VMD. SFT is simple to implement and computationally efficient, and can simultaneously and intuitively display signal features in two dimensions. However, it lacks sensitivity to abrupt changes and is difficult to describe transient features. EMD and VMD methods can also be applied to feature extraction. The EMD method can adaptively decompose into intrinsic mode functions without pre-setting basis functions, and it has good adaptability to nonlinear and non-stationary signals. However, EMD is prone to mode aliasing and endpoint effects. The VMD method effectively overcomes the mode aliasing problem of EMD and has higher decomposition stability, but it requires pre-setting the number of modes and penalty factors, and the designer needs to have mature experience. At the same time, the higher algorithm complexity will increase the computation time. In summary, time-frequency domain feature extraction methods are relatively mature in underwater acoustic signal processing, but they also have certain limitations.

[0007] iii. Traditional classifiers; Currently, commonly used traditional classifiers include Support Vector Machines (SVMs), K-Nearest Neighbors (KNN) algorithms, decision trees, and Gaussian mixture models. SVMs are suitable for high-dimensional and non-linear processing, exhibiting good generalization ability, but their long training time makes them unsuitable for large-scale data processing. KNN algorithms are easy to implement and also applicable to non-linearly distributed data, but their computational cost is high, resulting in low classification efficiency. Decision trees offer strong interpretability and perform well on small datasets, but are prone to overfitting and have weak generalization ability. Gaussian mixture models are suitable for incompletely labeled data and have strong expressive power for continuous feature distributions, but they are prone to getting trapped in local optima. Therefore, it is necessary to choose the appropriate classifier based on the specific circumstances.

[0008] With the improvement of computing power and the continuous accumulation of large-scale data, deep learning technology has gradually become a research hotspot in modern intelligent information processing methods. Unlike the traditional machine learning algorithms mentioned above, deep learning algorithms can extract deeper feature information from raw data, further reducing the dependence on manual feature design. For example, the proposal of generative adversarial networks can efficiently handle small sample cases. Through adversarial training, a large number of samples can be generated in the preprocessing stage for model training, which is a key representative of the transition from recognition to generation, but its training process is not stable enough. Other scholars have proposed Transformer and self-attention mechanisms for deep feature extraction, realizing parallel modeling of global dependencies, which is suitable for processing complex time-series data, but the computational and memory overhead of this algorithm is too large. The proposal of autoencoders has shown good results under unsupervised training and can be used for feature compression and reconstruction, but it cannot generate new samples and is prone to overfitting. In order to solve the defect of autoencoders in not being able to generate new samples, variational autoencoders have been proposed, but the reconstructed results are often blurry, and the modeling of high-dimensional complex data such as SRN is still limited. Others have used improved HHT transforms and constructed hybrid neural network architectures for multi-scale feature extraction, which can effectively improve the recognition rate. However, this method relies entirely on large labeled samples, which limits its effectiveness. Therefore, while deep learning algorithms have become the mainstream technology for signal recognition and classification due to their powerful nonlinear modeling capabilities, they also suffer from problems such as high computational requirements, susceptibility to overfitting, and poor stability.

[0009] The emergence of deep learning algorithms has marked a significant step forward in SRN classification and recognition research. With continuous technological advancements, deep learning algorithms will mature further, offering broader application prospects for complex acoustic pattern recognition.

[0010] Existing solutions to the technical problem: Option 1: Acoustic signals, as an effective carrier of long-distance information transmission in underwater environments, are of great significance for target identification. Compared with traditional signal processing methods, deep learning in data-driven methods has higher computational complexity and cannot extract more relevant feature vectors with minimal domain transformation and expertise. Furthermore, the quality of acoustic signal acquisition depends to a large extent on various factors.

[0011] Limitations: 1. This method relies on a single teacher model to generate pseudo-labels. When the noise levels in the target domain vary significantly, the errors in the pseudo-labels can be amplified and propagated, leading to bias in the teacher prediction system. 2. The model structure remains fixed during training, which limits its adaptability to different noise environments. 3. On resource-constrained real-world platforms, the workload needs further reduction.

[0012] Option 2: Due to the time-varying nature of underwater acoustic channels and the increasingly complex marine environment, the identification of underwater acoustic targets has become extremely difficult.

[0013] Limitations: This method relies on a large amount of labeled data and 3D fusion input. Although it utilizes multi-domain features to achieve high recognition accuracy, labeled data is scarce in the actual marine environment, and the background noise is complex. The dependence on labeled data reduces its generalization ability and makes it prone to overfitting.

[0014] Option 3: Maritime traffic safety requires fine-grained ship radiated noise identification methods for different specific ships. Existing ship radiated noise identification systems have certain shortcomings. The imperfections in the feature extraction and identification algorithms for ship radiated noise result in the ability to classify ships rather than identify them.

[0015] Limitations: 1. While this scheme achieves multi-scale structural fusion, enabling complementary features across different levels, it also results in a complex network structure, high computational cost, and demanding hardware requirements. 2. This scheme still primarily relies on supervised training and does not consider the scarcity of labeled samples in ship radiated noise data. Summary of the Invention

[0016] To overcome the shortcomings of existing technologies, this invention provides a semi-supervised collaborative learning-based method for identifying ship radiated noise. Normalized audio is transformed in the frequency domain using a Mel filter bank to obtain the Mel spectrum of the noise energy distribution. The model input consists of the audio and the Mel spectrum. Data from each branch is divided into unlabeled training sets, labeled training sets, and test sets. The unlabeled data is pseudo-labeled and participates in collaborative learning along with the labeled data. Feature fusion parameters are dynamically adjusted, and the final output classification is achieved through a fully connected layer. This method effectively improves the accuracy and generalization ability of ship radiated noise identification in small sample scenarios and reduces the dependence of classification on large-scale labeled data.

[0017] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Data processing; Step 1-1: Audio Processing; The raw SRN data is normalized and then used as the input to one branch of the SSCL; the normalization formula is:

[0018] in, It is the original signal. The mean of the signal. The signal variance; Steps 1-2: Image processing; The normalized audio signal is converted into a Mel spectrum using the following formula:

[0019] in, It involves performing STFT processing on the audio. For the first Weighting coefficients of each Mel filter; K This indicates the total number of frequency points obtained after STFT processing; Indicates the center frequency of the filter; This indicates the first frame of the audio signal after frame processing. One time frame; Step 2: Feature extraction network construction; Step 2-1: Audio Feature Extraction Network; The SincResNet structure is used as the feature extractor. The input waveform is filtered through a Sinc convolutional layer to form a preliminary feature representation; then ReLU activation and max pooling are applied; subsequently, it is processed through multiple residual modules to complete feature extraction; the simplified feature extraction formula is as follows:

[0020] in It is a Sinc convolutional layer. It is a multi-layer residual module. The number of channels representing the feature; Indicates the sequence length of features in a single channel; Step 2-2: Image Feature Extraction Network; ResNet18 is used as the feature extractor; preliminary features in the frequency domain are obtained through basic initial sampling; subsequently, four residual layers are used to extract multi-scale spatial features from low to high levels; finally, the final features are generated by flattening; the simplified feature extraction formula is as follows:

[0021] in, This is the initial feature extraction module. For global average pooling GAP, There are four residual modules; , Indicates the batch size of the sample; This represents the feature dimension of a single sample after processing. Step 3: Adaptive pseudo-tag generation and selection mechanism; Step 3-1: Sample selection; The pseudo-label prediction probability vector obtained through feature training and classifier Choose the one with the highest probability. As candidate labels; pseudo-labels are selected as:

[0022] in The set confidence threshold; Indicates the first The unlabeled sample corresponds to the first... The probability of pseudo-label prediction for the category; Step 3-2: Loss Function; To utilize the prediction results from unlabeled samples for training supervision, a pseudo-label loss function is introduced; the pseudo-label loss function is:

[0023] in The crossover loss function, This represents the pseudo-label corresponding to the unlabeled sample. This represents the prediction results for unlabeled samples based on Mel spectral features. This represents the prediction results for unlabeled samples based on waveform features; Step 4: Collaborative learning mechanism; Step 4-1: Loss based on supervised learning; The loss function is:

[0024] in and These represent the total number of categories and the total number of samples, respectively. The corresponding predicted probability; Indicates the first The sample corresponds to the first The true label of the class; Indicates the first Input features of each sample; Step 4-2: Consistency learning loss; The loss function is:

[0025] in Indicates the Kullback-Leibler divergence. and The output distributions of the two branches are shown below. Let be the mean of the two branch distributions; the Kullback-Leibler divergence is defined as follows:

[0026] This formula represents a measure. and The differences between them; Represent one of the probability distributions; This indicates that the probability distribution of the other variable to be compared is... The probability value at that location; Step 4-3: Total Loss; The final total loss function is in the form of:

[0027] in, and These are weight parameters; Step 5: Dynamic feature fusion; The formula for obtaining dynamic feature weights is:

[0028] in, Represents the Mel-spectral characteristics of audio; Represents the waveform domain characteristics of audio; Represents the feature mapping function; This represents the Softmax activation function; Further feature-weighted fusion:

[0029] in , These represent the spectral characteristics of Mel. Corresponding dynamic weights and waveform domain features The corresponding dynamic weights; Step 6: Training; The input data is passed to the two branch models for forward propagation to obtain the feature fusion result after attention-weighted fusion; the total loss is backpropagated to dynamically update the relevant parameters of the model; the above steps are repeated until the early stopping round is triggered.

[0030] Step 7: The final output is the sample recognition accuracy.

[0031] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described ship radiated noise identification method.

[0032] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described ship radiated noise identification method.

[0033] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the aforementioned ship radiated noise identification method.

[0034] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described ship radiated noise identification method.

[0035] The beneficial effects of this invention are as follows: This invention maximizes the use of feature information from different domains of SRN for classification and recognition, thereby improving the overall recognition accuracy. This cross-feature domain joint learning method also enhances the model's generalization ability. Furthermore, the early stopping mechanism and the constraint of the loss function address the overfitting problem. The use of pseudo-labels solves the problem of insufficient labeled data in SRN datasets, reducing the data's dependence on the model. Collaborative learning ensures information sharing among models. In summary, this invention not only achieves high accuracy in classifying and recognizing SRN data but also solves potential problems in practical applications. Attached Figure Description

[0036] Figure 1 This is a technical flowchart of the method of the present invention; Figure 2 shows the time series diagrams and corresponding Mel diagrams of some signals in the DeepShip dataset, where Figure 2(a) is a cargo ship, Figure 2(b) is a passenger ship, Figure 2(c) is a tugboat, Figure 2(d) is a Mel diagram of a cargo ship, Figure 2(e) is a Mel diagram of a passenger ship, and Figure 2(f) is a Mel diagram of a tugboat. Figure 3 shows the timing diagrams and corresponding Mel plots of some signals in the ShipsEar dataset, where Figure 3(a) is signal A, Figure 3(b) is signal B, Figure 3(c) is signal C, Figure 3(d) is the Mel plot of signal A, Figure 3(e) is the Mel plot of signal B, and Figure 3(f) is the Mel plot of signal C. Figure 4 shows the timing diagrams and corresponding Mel plots of some signals in the QiandaoEar22 dataset, where Figure 4(a) is signal A, Figure 4(b) is signal B, Figure 4(c) is signal C, Figure 4(d) is the Mel plot of signal A, Figure 4(e) is the Mel plot of signal B, and Figure 4(f) is the Mel plot of signal C. Figure 5 shows the confusion matrix of the identification results of the method proposed in this invention. Figure 5(a) shows the DeepShip dataset, Figure 5(b) shows the ShipsEar dataset, and Figure 5(c) shows the QiandaoEar22 dataset. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] The technical problem to be solved by this invention is as follows: (1) The existing SRN dataset has few labeled samples, and the existing deep learning-based recognition methods will result in insufficient generalization ability of the model.

[0039] (2) The existing dataset has a small number of samples, which makes it easy for overfitting to occur during deep learning.

[0040] (3) Existing deep learning-based recognition methods do not make sufficient use of information from different feature domains.

[0041] (4) Existing deep learning-based recognition methods are computationally too complex, which prolongs the model training time.

[0042] In recent years, thanks to the development of deep learning algorithms, the accuracy of underwater acoustic target recognition has been significantly improved. Deep learning algorithms can utilize massive amounts of data to deeply extract sample features for recognition, demonstrating excellent performance. However, actual ship radiated noise data is scarce, especially labeled samples. Furthermore, environmental interference can lead to overfitting or insufficient feature learning under small sample conditions. To address the impact of these issues on ship recognition results, this invention proposes a SRN recognition method based on SSCL. Its core idea is to utilize cross-feature domain dual-branch collaborative learning combined with an adaptive pseudo-label generation and constraint mechanism. The technical flowchart of this invention is shown below. Figure 1 As shown.

[0043] This method mainly consists of three parts: dual-domain feature extraction model construction, semi-supervised collaborative learning, and feature fusion and classification. The detailed description of all steps in this invention is as follows: The normalized audio is transformed in the frequency domain using a Mel filter bank to obtain the Mel spectrum of the noise energy distribution. The input to the model is the audio and the Mel spectrum. The data from each branch are divided into unlabeled training set, labeled training set, and test set according to a ratio of 0.49:0.21:0.3. The unlabeled data is given pseudo-labels and participates in collaborative learning together with the labeled data. The feature fusion parameters are dynamically adjusted, and the final output classification is completed through a fully connected layer. The detailed description of all steps in this invention is as follows: Step 1: Data Processing; (1) Audio processing: The raw SRN data is normalized and then used as input to one branch of the SSCL. Normalization formula:

[0044] in, It is the original signal. The mean of the signal. This represents the signal variance. This method ensures consistent input scales for the model and eliminates amplitude differences between different samples.

[0045] (2) Image processing: The normalized audio signal is converted into a Mel spectrum. Conversion formula:

[0046] in, It involves performing STFT processing on the audio. For the first The weighting coefficients of the Mel filter are used. This method can better capture important low-frequency information of ship radiated noise while reducing high-frequency noise interference.

[0047] Step 2: Feature extraction network construction; (1) Audio feature extraction network: This module uses the SincResNet structure as the feature extractor. Filtering is performed through Sinc convolutional layers to form a preliminary feature representation. Feature processing also includes ReLU activation and max pooling. Subsequently, the input waveform is processed through multiple residual modules to complete feature extraction. Simplified feature extraction formula:

[0048] in It is a Sinc convolutional layer. It is a multi-layer residual module.

[0049] (2) Image feature extraction network: This module uses ResNet18 as the feature extractor. Preliminary features in the frequency domain are obtained through basic initial sampling. Subsequently, four residual layers are used to extract multi-scale spatial features from low to high levels. Finally, the final features are generated by flattening. Simplified feature extraction formula:

[0050] in, This is the initial feature extraction module. For GAP, There are four residual modules.

[0051] Step 3: Adaptive pseudo-tag generation and selection mechanism; For unlabeled samples, the quality of the pseudo-label selection directly determines the learning effect and stability of the model during iteration. Therefore, an adaptive selection strategy is chosen to ensure that high-confidence pseudo-labels participate in co-training.

[0052] (1) Sample selection: For the pseudo-label prediction probability vector obtained through feature training and classifier Choose the one with the highest probability. As candidate labels, the pseudo-labels selected are:

[0053] in The set confidence threshold.

[0054] (2) Loss Function: To utilize the prediction results of unlabeled samples to provide training supervision, a pseudo-label loss function is introduced. The pseudo-label loss function is:

[0055] in The cross-loss function takes (predicted probability, target label) as input.

[0056] Step 4: Collaborative Learning Mechanism; In the collaborative learning mechanism proposed in this invention, the loss function is crucial. It determines the overall optimization direction of the model during the collaborative learning process and is the core element ensuring that each branch can collaborate effectively and fully utilize both labeled and unlabeled data for efficient training. The relevant principles of each loss function are as follows: (1) Loss based on supervised learning: To measure the difference between the true label distribution of labeled samples and the model's predicted probability distribution, a loss function based on supervised training is introduced. The greater the difference, the higher the loss value. The core principle is the cross-entropy loss function. The loss function is:

[0057] in and These represent the total number of categories and the total number of samples, respectively. This represents the corresponding predicted probability.

[0058] (2) Consistency learning loss: To avoid a single model overfitting its own modality, consistency loss can be used to force the outputs of the two branches to remain consistent, allowing the two branch models to reach a consensus and generate more reliable pseudo-labels and model fusion parameters. The loss function is:

[0059] in Indicates the Kullback-Leibler divergence. and The output distributions of the two branches are shown below. Let be the mean of the two branch distributions. The Kullback-Leibler divergence is defined as follows:

[0060] Wherein represents measurement and The differences between them.

[0061] (3) Total loss; The final total loss function is a weighted sum of the above losses, and its form is:

[0062] in, and These are weight parameters.

[0063] Step 5: Dynamic Feature Fusion; During collaborative learning, SSCL continuously selects optimized attention weights through model iteration and performs feature fusion, which directly determines the final recognition result. The formula for obtaining dynamic feature weights is:

[0064] Further feature-weighted fusion:

[0065] in This is the result of mapping two feature domains to the same dimension; This is a general term for small neural networks used to calculate attention weights, which include fully connected layers and activation functions. It refers to this part of the code: the small neural network.

[0066] Step Six: Training; The input data is passed to the two branch models for forward propagation, initially yielding the feature fusion result after attention-weighted fusion. The total loss is then used for backpropagation to dynamically update the model's parameters. This process is repeated until an early stopping round is triggered.

[0067] Step 7: Output; The final output is the sample recognition accuracy.

[0068] Example: (1) Dataset used in the experiment; To verify the generalization performance of the proposed model, three ship radiated noise datasets were selected for validation: the DeepShip dataset, the ShipsEar dataset, and the QiandaoEar22 dataset. The corresponding datasets are described below: 1) Deepship dataset; The Deepship public dataset was developed to fill a gap in underwater acoustic datasets, containing records of various sea conditions and noise levels throughout the year. Data was recorded from May 2, 2016 to October 4, 2018, in one of the busiest shipping lanes on the Pacific Northwest coast, near Vancouver, one of Canada's busiest ports. The dataset contains 47 hours and 4 minutes of data on 265 different vessels, which can be categorized into four types: tankers, tugboats, passenger ships, and cargo ships. The data was recorded using an icListen AF hydrophone with a bandwidth of 1Hz-12kHz, a dynamic range of 120dB, and a sensitivity of [missing information]. 170 dB re 1V / Pa. Data was recorded at a sampling rate of 32 kHz. The data includes the ship name, recording time, duration, and distance of the ship from the sensor. Figure 2 shows the signal timing diagrams and corresponding Mel plots for some ships in the DeepShip dataset.

[0069] 2) ShipsEar dataset; The ShipsEar public dataset was developed to address the lack of data for underwater researchers. This dataset contains noise data from different areas along the Spanish Atlantic coast in northwestern Spain from autumn 2012 to summer 2013, including 11 categories of ship-radiated noise from fishing boats, ocean-going passenger ships, ferries of various sizes, and one background noise. The dataset was captured using a digitalHyd SR-1 hydrophone with a sensitivity of -193.5 dB re 1V / 1 uPa and a sampling rate of 52734 Hz. The authors categorized the dataset into five classes (A, B, C, D, and E) based on ship size, and included information such as GPS location, recording time and date, hydrophone gain, and in-situ measured wind speed and depth. Figure 3 shows the signal timing diagrams and corresponding Mel plots for some ships in the ShipsEar dataset.

[0070] 3) QiandaoEar22 dataset; The QiandaoEar22 public dataset was collected from June 24th to 28th, 2022, at the Xin'anjiang Experimental Station of the Institute of Acoustics, Chinese Academy of Sciences, located in Qiandao Lake Town, Zhejiang Province, China. The dataset includes vessel types primarily consisting of UUVs, speedboats, and the Kaiyuan passenger ship, with mixed background radiated noise from ships, totaling 20 vessel types. Recording was performed using a DigitalHyd SR-1 hydrophone with a receiving sensitivity of -162.2 ohms. -126.1 dB re 1V / 1 uPa, recorded at a sampling frequency of 52734 Hz, ultimately yielding a cumulative 21 hours and 58 minutes of background noise data, recorded according to the distance between the ship and the hydrophone, sound audibility, and audio content target. Figure 4 shows the signal timing diagrams and corresponding Mel plots of some ships in the QiandaoEar22 dataset.

[0071] (2) Experimental setup; This experiment was conducted on a Windows 11 system, using Python 3.11.11 as the software, with deep learning provided by PyTorch and accelerated by CUDA version 12.7. The hardware configuration included an Intel Core i7-13700H processor, an NVIDIA GeForce RTX 4060 graphics card (8GB VRAM), and 32GB of DDR5 system memory. The following key parameters were set during the experiment: 80 iterations, a pseudo-label confidence threshold of 0.97, a patience value of 20 for early stopping, the Adam optimization algorithm, an initial learning rate of 1e-3, and an L2 regularization coefficient of 1e-5. Other parameters are not listed here.

[0072] (3) Numerical simulation experiments; 1) Model performance testing experiments based on the DeepShip dataset; The two core mechanisms of the proposed method are an adaptive pseudo-label generation and selection mechanism and a collaborative learning mechanism. By pseudo-labeling unlabeled data, the problem of insufficient labeled data for ship radiated noise can be effectively addressed. However, due to the inherent characteristics of pseudo-labels, their reliability may decrease during model iteration, affecting recognition performance. Therefore, determining the proportion of unlabeled data is the core of this mechanism's research. To simultaneously verify the effectiveness of collaborative learning, a comparative experiment was conducted between the collaborative learning model and the single-branch model. Table 1 shows the experimental results, comparing the experiments with different amounts of unlabeled data with the comparative experiments with collaborative learning and the single-branch model. The unlabeled data proportion refers to the percentage of unlabeled data in the training set.

[0073] Table 1. Experimental results of the comparative experiment.

[0074] As shown in Table 1, the model exhibits the best recognition performance when the proportion of unlabeled data is 0.7. When the proportion of unlabeled data is too low, the complex distribution and small feature differences of ship radiated noise mean that the model can only learn local features due to its reliance on labeled data for training. The presence of false labels can lead to feature expansion. Conversely, when the proportion of unlabeled data is too high, the model is susceptible to misleading information from the accumulation of erroneous false labels, resulting in blurred classification boundaries and decreased recognition performance. Comparative experiments between the single-branch model and the collaborative learning model further demonstrate the effectiveness of this invention. Through knowledge transfer between the two branches, the model can dynamically fuse the feature information of both branches, achieving deep interaction of multimodal information and effectively improving recognition performance.

[0075] 2) Experiments on model generalization performance based on different datasets; To further demonstrate the applicability of the proposed method in practical engineering, experiments were conducted on publicly available datasets under different conditions. The DeepShip dataset contains relatively clean background noise, which can verify the basic performance of this method. The ShipsEar dataset contains a total of 11 classes of ship radiated noise, which can verify the performance of this method in multi-class classification. The QiandaoEar22 dataset was collected near shipping lanes used in tourism, resulting in a single noise dataset containing multiple types of ship radiated noise. Effectively extracting the required ship noise from the complex marine environment is a crucial step in verifying the recognition performance in practical applications. The experimental results tested on the above different datasets are shown in Table 2. The confusion matrix of the recognition results of the proposed method is shown in Figure 5.

[0076] Table 2 Experimental results for different datasets

[0077] Experimental results show that the proposed method exhibits good basic recognition performance, achieving an accuracy of 0.9365. Although performance slightly decreases on the ShipsEar dataset, this further demonstrates the method's good cross-dataset generalization performance. On the QiandaoEar22 dataset, performance remains around 0.82. This is because the adaptive pseudo-label generation and selection mechanism introduced in this experiment is affected by overly complex background noise during model iteration, impacting the reliability of pseudo-labels and leading to a decrease in pseudo-label generation quality, thus interfering with training. However, overall, the model maintains good recognition performance, demonstrating that this invention can reliably identify and classify ship radiated noise in real-world scenarios with strong noise interference and numerous data categories, possessing significant application value in practical maritime monitoring and identification engineering.

[0078] 3) Comparative experiments based on different existing models; This method is compared with five common existing recognition methods: Support Vector Machine, Deep Convolutional Neural Network, Multilayer Perceptron, Long Short-Term Memory Network, and Sequence Modeling Network based on Attention Mechanism, denoted as M1-M5 for simplicity. Experimental results of different methods on different datasets are shown in Table 3.

[0079] Table 3 Experimental results of different methods based on different datasets

[0080] As shown in Table 3, the proposed method outperforms other methods in recognition performance on the DeepShip, ShipsEar, and QiandaoEar22 datasets. On the DeepShip dataset, the accuracy of this invention is approximately 8 percentage points higher than the best contrast model M2, indicating that it can effectively distinguish ship types even under background noise. On the ShipsEar dataset, the accuracy of this invention is similar to that of the best contrast model M5, but still exhibits the best performance, demonstrating good performance in multi-class classification. On the QiandaoEar22 dataset, the accuracy of this invention is approximately 4 percentage points higher than the best contrast model M5. Although the accuracy is slightly lower, it shows that this invention can still effectively learn features under weakly labeled background noise and mixed noise from multiple ship types, further demonstrating the advantages of the semi-supervised collaborative learning framework.

[0081] The above experiments have verified that the SSCL method proposed in this invention has significant implications for practical engineering applications and has high potential for widespread application in ship target identification.

Claims

1. A method for identifying ship radiated noise based on semi-supervised collaborative learning, characterized in that, Includes the following steps: Step 1: Data processing; Step 1-1: Audio Processing; The raw SRN data is normalized and then used as the input to one branch of the SSCL; the normalization formula is: in, It is the original signal. The mean of the signal. The signal variance; Steps 1-2: Image processing; The normalized audio signal is converted into a Mel spectrum using the following formula: in, It involves performing STFT processing on the audio. For the first Weighting coefficients of each Mel filter; K This indicates the total number of frequency points obtained after STFT processing; Indicates the center frequency of the filter; This indicates the first frame of the audio signal after frame processing. One time frame; Step 2: Feature extraction network construction; Step 2-1: Audio Feature Extraction Network; The SincResNet structure is used as the feature extractor. The input waveform is filtered through a Sinc convolutional layer to form a preliminary feature representation; then ReLU activation and max pooling are applied; subsequently, it is processed through multiple residual modules to complete feature extraction; the simplified feature extraction formula is as follows: in It is a Sinc convolutional layer. It is a multi-layer residual module. The number of channels representing the feature; Indicates the sequence length of features in a single channel; Step 2-2: Image Feature Extraction Network; ResNet18 is used as the feature extractor; preliminary features in the frequency domain are obtained through basic initial sampling; subsequently, four residual layers are used to extract multi-scale spatial features from low to high levels; finally, the final features are generated by flattening; the simplified feature extraction formula is as follows: in, This is the initial feature extraction module. For global average pooling GAP, There are four residual modules; , Indicates the batch size of the sample; This represents the feature dimension of a single sample after processing. Step 3: Adaptive pseudo-tag generation and selection mechanism; Step 3-1: Sample selection; The pseudo-label prediction probability vector obtained through feature training and classifier Choose the one with the highest probability. As candidate labels; pseudo-labels are selected as: in The set confidence threshold; Indicates the first The unlabeled sample corresponds to the first... The probability of pseudo-label prediction for the category; Step 3-2: Loss Function; To utilize the prediction results from unlabeled samples for training supervision, a pseudo-label loss function is introduced; the pseudo-label loss function is: in The crossover loss function, This represents the pseudo-label corresponding to the unlabeled sample. This represents the prediction results for unlabeled samples based on Mel spectral features. This represents the prediction results for unlabeled samples based on waveform features; Step 4: Collaborative learning mechanism; Step 4-1: Loss based on supervised learning; The loss function is: in and These represent the total number of categories and the total number of samples, respectively. The corresponding predicted probability; Indicates the first The sample corresponds to the first The true label of the class; Indicates the first Input features of each sample; Step 4-2: Consistency learning loss; The loss function is: in Indicates the Kullback-Leibler divergence. and The output distributions of the two branches are shown below. Let be the mean of the two branch distributions; the Kullback-Leibler divergence is defined as follows: This formula represents a measure. and The differences between them; Represent one of the probability distributions; This indicates that the probability distribution of the other variable to be compared is... The probability value at that location; Step 4-3: Total Loss; The final total loss function is in the form of: in, and These are weight parameters; Step 5: Dynamic feature fusion; The formula for obtaining dynamic feature weights is: in, Represents the Mel-spectral characteristics of audio; Represents the waveform domain characteristics of audio; Represents the feature mapping function; This represents the Softmax activation function; Further feature-weighted fusion: in , These represent the spectral characteristics of Mel. Corresponding dynamic weights and waveform domain features The corresponding dynamic weights; Step 6: Training; The input data is passed to the two branch models for forward propagation to obtain the feature fusion result after attention-weighted fusion; the total loss is backpropagated to dynamically update the relevant parameters of the model; the above steps are repeated until the early stopping round is triggered. Step 7: The final output is the sample recognition accuracy.

2. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in claim 1.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in claim 1.

4. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in claim 1.

5. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in claim 1.

Images