A method, system, device and medium for identifying ship radiation noise

Abstract

The present invention relates to the technical field of underwater acoustic detection, and discloses a method, system, device and medium for identifying radiation noise of a ship. The method comprises: obtaining the radiation noise signal generated by the ship during navigation; constructing a wave equation according to the physical law followed by the radiation noise signal during propagation; constructing a first loss function based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and constructing a second loss function based on whether the predicted category output by the PINNs network satisfies the wave equation as the constraint condition of the PINNs network output; training the PINNs network using the radiation noise signal, and optimizing the network parameters of the PINNs network according to the first loss function and the second loss function during the training process; identifying the radiation noise of the ship through the trained PINNs network, and having good recognition performance.

Description

Technical Field

[0001] The present invention relates to the field of underwater acoustic detection technology, and in particular to a method, system, equipment and medium for identifying ship radiation noise. Background Art

[0002] Using hydrophones to collect passive radiation noise from ships and extract features to classify and identify ships is of great significance in promoting the development of the marine economy and protecting national marine rights and interests. Previous identification methods relied on expert experience, mainly by skilled technicians performing signal processing, feature extraction, and identification and confirmation of ship types. Due to the complex and changeable marine environment, there are many types of marine noise (such as marine animals, artificial operation noise, wave sound, etc.), which poses a great challenge to manual identification.

[0003] With the rise of artificial intelligence technology, the study of automatic and accurate identification methods using neural networks has become a development trend. At present, when identifying the passive radiation noise of ships based on deep learning (i.e., classification and identification of ships), the model establishment process only relies on the passive radiation noise data of ships, i.e., pure data, without considering the physical laws followed by signal propagation, resulting in poor recognition performance. Summary of the invention

[0004] The object of the present invention is to provide a method, system, device and medium for identifying ship radiation noise, which can improve the recognition performance of ship radiation noise.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for identifying ship radiated noise, comprising the following steps:

[0006] Acquire the radiation noise signal generated by the ship during navigation;

[0007] According to the physical laws followed by the radiation noise signal during the propagation process, the wave equation is constructed;

[0008] The first loss function is constructed based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and the second loss function is constructed based on whether the predicted category output by the PINNs network satisfies the wave equation as the constraint condition of the PINNs network output;

[0009] The PINNs network is trained using a radiation noise signal, and the network parameters of the PINNs network are optimized according to a first loss function and a second loss function during the training process;

[0010] Ship radiation noise is identified through the trained PINNs network.

[0011] Optionally, the wave equation is:

[0012] ;

[0013] In the formula, u and c represent the sound pressure and sound speed respectively, represents the Laplace operator.

[0014] Optionally, whether the prediction category output by the PINNs network satisfies the constraint condition that the wave equation is the output of the PINNs network is:

[0015] ;

[0016] The second loss function is:

[0017] ;

[0018] In the formula, and Respectively represent the second-order derivatives of the output of the PINNs network with respect to time and space, M Indicates the number of samples.

[0019] Optionally, the wave equation is expressed in a strong form or a weak form of a physical law followed by the radiation noise signal during propagation;

[0020] The first loss function comprises the residual of the wave equation if the wave equation is expressed in the strong form of the laws of physics and is derived from the energy functional of the wave equation if the wave equation is expressed in the weak form of the laws of physics.

[0021] Optionally, optimizing the network parameters of the PINNs network according to the first loss function and the second loss function during the training process includes:

[0022] During the training process, the network parameters of the PINNs network are optimized according to the respective weights of the first loss function and the second loss function with the goal of minimizing the first loss function and the second loss function; wherein the respective weights of the first loss function and the second loss function will be dynamically adjusted during the training process.

[0023] Optionally, the network parameters of the PINNs network are optimized by the following formula:

[0024] ;

[0025] In the formula, represents the network parameters, represents the optimized network parameters, and denote the exponentially weighted moving averages of the gradient and squared gradient, respectively. represents the learning rate, Represents a preset numerical stabilization term.

[0026] Optionally, the first loss function is constructed based on a cross entropy loss function.

[0027] An embodiment of the present invention further provides a ship radiated noise identification system, comprising:

[0028] A signal acquisition module is used to acquire the radiation noise signal generated by the ship during navigation;

[0029] An equation building module, which is used to build the wave equation according to the physical laws that the radiated noise signal follows during the propagation process;

[0030] A function construction module is used to construct a first loss function based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and to construct a second loss function based on whether the predicted category output by the PINNs network satisfies the wave equation as a constraint condition for the PINNs network output;

[0031] A network training module, used to train the PINNs network using a radiation noise signal, and optimize the network parameters of the PINNs network according to a first loss function and a second loss function during the training process;

[0032] The ship identification module is used to identify ship radiated noise through the trained PINNs network.

[0033] An embodiment of the present invention also provides a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned ship radiation noise identification method.

[0034] An embodiment of the present invention further provides a computer-readable storage medium storing a computer program, which implements the above-mentioned ship radiation noise identification method when executed by a processor.

[0035] The ship radiation noise identification method provided by the present invention has at least the following beneficial effects:

[0036] By introducing a wave equation constructed according to the physical laws followed by the radiation noise signal during propagation, the loss function includes not only the loss of training data (i.e., the first loss function), but also the physical loss based on the wave equation constructed by the wave equation (i.e., the second loss function). When the radiation noise signal generated by the ship during navigation is used to train the neural network, the training process is constrained by the physical loss, and the actual physical phenomena are taken into account. Finally, the trained neural network model has better recognition performance. In addition, the addition of the above physical loss makes the loss based on training data no longer the main factor affecting model training, so it is no longer necessary to use large-scale data sets for training, reducing the probability of model overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] One or more embodiments are exemplarily described by the pictures in the corresponding drawings, and these exemplary descriptions do not constitute limitations on the embodiments.

[0038] Figure 1 is a flow chart of a method for identifying ship radiated noise provided according to an embodiment of the present invention;

[0039] Figure 2 is a technical roadmap of a ship radiation noise identification method provided according to an embodiment of the present invention;

[0040] Figure 3 A signal timing is provided according to an embodiment of the present invention. Figure 1 ;

[0041] Figure 4 A signal timing is provided according to an embodiment of the present invention. Figure 2 . DETAILED DESCRIPTION

[0042] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. However, it will be appreciated by those skilled in the art that in the embodiments of the present invention, many technical details are proposed in order to enable the reader to better understand the present invention. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed in the present invention can be implemented. The division of the following embodiments is for the convenience of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined and referenced with each other without contradiction.

[0043] Current signal recognition methods are mainly divided into: (1) based on traditional signal processing methods; (2) based on modern intelligent information processing and deep learning.

[0044] (1) Based on traditional signal processing methods:

[0045] The central idea of ​​the recognition method based on traditional signal processing methods is "birds of a feather flock together", that is, similar samples are close to each other in the pattern space. There must be some information differences between samples of different patterns, and these differences are usually reflected by some measurement methods, such as entropy and distance function. This type of method is generally divided into three steps, namely data preprocessing, feature extraction and classification.

[0046] i. Data preprocessing:

[0047] The existing data preprocessing methods are mainly divided into two categories: filtering and wavelet analysis. Regarding filtering methods, there are mainly Wiener filtering, Kalman filtering and adaptive filtering. The advantage of Wiener filtering is that it can filter in real time, can effectively reduce noise, and is widely used, but it has the defects of high statistical feature requirements and difficulty in processing nonlinear and non-Gaussian signals. The advantage of Kalman filtering is that it can handle uncertainty and randomness, is suitable for real-time estimation, and has good estimation performance for random systems, but has high computational complexity. The advantage of adaptive filtering is that it is more adaptable, has better filtering performance, and can adaptively adjust its own filter parameters, but some "music noise" still exists, and background noise has a great impact on performance. Regarding wavelet analysis methods, there are mainly soft threshold noise reduction and hard threshold noise reduction. The advantage of soft threshold noise reduction is that it is simple to calculate, easy to implement, and can maintain the main characteristics of the signal. The advantage of hard threshold noise reduction is that it can retain more signal features, especially when the signal strength changes greatly. However, the common problem of these two methods is that it is difficult to select a suitable threshold, and the latter is prone to generate false information.

[0048] ii. Feature extraction

[0049] The existing feature extraction methods are mainly divided into time domain, time-frequency domain and cepstrum domain feature extraction. 1. Regarding the time domain feature extraction method, there are mainly noise envelope modulation detection, low-frequency analysis and recording spectrum. The advantage of noise envelope modulation detection is that it has strong nonlinear and non-stationary signal processing capabilities, as well as higher frequency and time resolution, but it may produce overfitting. The advantages of low-frequency analysis and recording spectrum are that they have strong real-time data processing capabilities, as well as higher time and frequency resolution, but the application range is limited by the frequency band. 2. Regarding the time-frequency domain feature extraction method, there are mainly empirical mode decomposition, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Variational Mode Decomposition (VMD) and short-time Fourier transform. The advantage of empirical mode decomposition is that it does not require a preset basis function, can process nonlinear and non-stationary signals, and the local characteristics of the signal can be retained, but it is easy to produce mode aliasing and endpoint effects. The advantage of CEEMDAN is its strong adaptability, which can avoid the influence of excessive noise on the decomposition results and is not restricted by prior assumptions, but it needs to select appropriate noise parameters, and large-scale data leads to longer training and processing time. The advantage of VMD is that it can avoid modal aliasing, has strong frequency resolution, and has a wider range of applications, but it needs to preset decomposition parameters. The advantage of short-time Fourier transform is that it can analyze the frequency components of the signal at different time points, the spectrum diagram can directly display the time and frequency distribution of the signal, and the calculation process is intuitive and easy to implement, but it is difficult to process signals with fast changes and variable frequencies. 3. Regarding the cepstrum domain feature extraction methods, there are mainly cepstrum mean subtraction, cepstrum mean normalization and Mel-scale Frequency Cepstral Coefficients (MFCC). The advantage of cepstrum mean subtraction is that it can effectively remove environmental noise, and the calculation is simple and efficient, but it is ineffective for non-stationary noise. The advantage of cepstrum mean normalization is strong stability and is applicable to a variety of conditions, but the processing effect of dynamic noise is poor. The advantage of MFCC is that it can effectively distinguish different speech or audio samples and has excellent feature extraction capabilities, but it is too sensitive to noise.

[0050] iii. Classification:

[0051] The existing mainstream classification methods mainly include decision trees, support vector machines, and K-nearest neighbors. The advantage of decision trees is that they are highly interpretable and can effectively handle nonlinear problems, but they are prone to overfitting and falling into local optimality. The advantage of support vector machines is that they have better generalization capabilities and can effectively handle high-dimensional data, but the training time is long. The advantage of K-nearest neighbors is that they are suitable for small-scale data sets and can effectively handle nonlinear problems, but they are prone to dimensionality disasters.

[0052] (2) Based on modern intelligent information processing and deep learning:

[0053] With the improvement of computer technology and big data, deep learning has made breakthroughs in many problems that were difficult to solve in the past. With the advantages of deep learning, recognition models based on modern intelligent information processing and deep learning have also made revolutionary progress, and recognition models based on traditional signal processing methods will gradually be replaced. For example, a recognition method based on attention mechanism has been proposed. Specifically, the residual neural network Resnet is used to extract the spectral features of the signal, and then the attention mechanism is introduced to enhance the features, and finally classification is performed based on the features. However, if the data set is too small, it will lead to serious overfitting. Based on this, a multi-scale Hough transform (MHT) transformation is proposed, and a new self-supervised recognition model is established on this basis. First, MHT is used to learn time-frequency information from the Mel Spectrogram, and then the deep features of the learning target are learned using MHT transformation. This model can reduce the dependence of the training process on large-scale data sets, but has poor adaptability to the environment. The recognition of ship radiation noise signals is transformed into open set recognition, and a recognition model based on fusion graph convolutional network and full parameterization is proposed. This model can not only resist the interference of background noise, but also classify various ship radiation noise signals. However, these advantages are achieved on the basis of large-scale data sets. Based on this, there is a new modeling idea, in which the one-dimensional target signal is converted into two dimensions and then recognized. However, this method is not ideal for the recognition of cargo ships. The recognition model based on WA-DS decision fusion mainly integrates MFCC, Gammatone Frequency Cepstral Coefficients (GFCC) and Time-Frequency (TF), and then uses Resnet18 and basic probability distribution function for classification decisions. This method combines various features to obtain high-precision recognition effects, but it requires high computing resources. In short, this type of method has problems such as lack of large-scale data sets, no connection between the model and actual physical phenomena, and high computational complexity.

[0054] Although there are some bottlenecks in the identification method based on modern intelligent information processing and deep learning, the key technology of feature extraction has made great progress, creating a good start for the in-depth exploration of ship radiated noise identification technology. Therefore, deep learning will be a powerful tool for studying ship radiated noise identification and is worthy of in-depth study.

[0055] There are several ship radiated noise identification solutions based on deep learning:

[0056] i. Solution 1: Due to the lack of labeled data collection, the influence of the inherent characteristics of time domain changes and the interference of other noise sources, underwater acoustic target recognition has always been one of the most challenging tasks in underwater signal processing. Although some deep learning methods have been shown to achieve state-of-the-art accuracy, the accuracy of the recognition task can be improved by designing residual networks and optimizing feature extraction. In order to more comprehensively express the underwater acoustic signal, this solution first proposes a data enhancement strategy of three-dimensional fusion features and SpecAugment. Then, a ResNet18 with a center loss function and an embedding layer is designed to train aggregate features with an adaptive learning rate. The solution is divided into five parts, namely data preprocessing, feature extraction, residual network and embedding layer, where feature extraction includes feature calculation, feature fusion and feature enhancement.

[0057] Disadvantages: 1. The processing flow of this solution is too long, and it takes a lot of time to perform feature calculation, feature fusion and feature enhancement. In addition, the network structure of ResNet18 is relatively complex, which ultimately leads to a long time for the entire solution. 2. This solution uses a data enhancement strategy to solve the problem of lack of labeled data sets, but the loss function in the training process is only based on the data loss, and does not embed physical constraints into the total loss, causing the training process to be out of touch with actual physical phenomena.

[0058] ii. Solution 2: Class imbalance is an objective problem in underwater acoustic datasets, but it rarely attracts people's attention. It often leads to low recognition accuracy of minority categories. The main purpose of this solution is to provide an effective method for the recognition of unbalanced underwater acoustic datasets. To this end, an exponentially weighted cross entropy loss is proposed as the loss function of the convolutional neural network, and an influence factor is added to the standard cross entropy loss according to the prediction probability of each sample.

[0059] Disadvantages: Although this solution solves the problem of imbalanced categories in the underwater acoustic dataset to a certain extent, the newly proposed exponentially weighted cross entropy loss as the loss function of the convolutional neural network is still trained based on pure data and does not take into account the physical laws that underwater acoustic signals follow during propagation.

[0060] iii. Solution 3: Ship radiated noise identification is an important and complex task in the construction of marine information systems and marine scientific research. Environmental noise, unstable frequency shifts and irregular multipath interference make accurate identification of ship radiated noise complicated. Existing identification methods have limited ability to identify the motion state of ships, resulting in unsatisfactory application results. In order to reduce the amount of calculation and effectively identify ship motion, this solution proposes a time-frequency Swin-Transformer network and a hierarchical self-attention module to extract multi-layer time-frequency features, so that the time-frequency Swin-Transformer network can learn the characteristics of moving targets in the time-frequency representation of the radiated noise of moving targets.

[0061] Disadvantages: Although this solution reduces the computational complexity of the algorithm to a certain extent and improves recognition accuracy, the physical laws followed by signal propagation are not considered in the process of model establishment.

[0062] Therefore, the existing ship radiation noise identification methods have a small data set size, which is prone to overfitting during the training process; the computational complexity is high and the required calculation time is long; and the data set is only used for training without combining the training process with actual physical phenomena.

[0063] Based on this, the present invention proposes a method for identifying ship radiation noise to solve the above technical problems.

[0064] An embodiment of the present invention relates to a method for identifying ship radiation noise. The implementation details of the method for identifying ship radiation noise of this embodiment are described in detail below. The following content is only the implementation details provided for easy understanding and is not necessary for implementing this solution.

[0065] The specific process of the ship radiation noise identification method of this embodiment can be as follows: Figure 1 As shown, including:

[0066] Step 101, obtaining the radiation noise signal generated by the ship during navigation.

[0067] Specifically, the radiation noise signal generated by the ship during navigation is first obtained. Then the original radiation noise signal is normalized, and the normalized radiation noise signal is used as the input of the subsequent PINNs network.

[0068] The normalization formula is as follows:

[0069] ;

[0070] In the formula, is the original radiated noise signal.

[0071] This approach ensures that all data are in the same range and helps eliminate scale differences between elements, especially for data with large amplitude variations.

[0072] Step 102: construct a wave equation according to the physical laws followed by the radiation noise signal during the propagation process.

[0073] Specifically, the physical laws followed by the radiated noise signal during propagation are described by the wave equation (WE). In the context of the ship classification task, the propagation process of sound waves can be described using WE, which is the basic physical law that describes the propagation of sound waves in the medium.

[0074] The wave equation constructed at this time is:

[0075] ;

[0076] In the formula, u and c represent the sound pressure and sound speed respectively, represents the Laplace operator.

[0077] Step 103, constructing a first loss function based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and constructing a second loss function based on whether the predicted category output by the PINNs network satisfies the wave equation as a constraint condition for the PINNs network output.

[0078] Specifically, Physics-Informed Neural Networks (PINNs) is a machine learning model that combines deep learning and physics knowledge. Unlike traditional data-driven neural networks, PINNs use physical laws to guide the model during the learning process, thereby improving the generalization ability of the model, especially when there is less data or more noise. Therefore, this embodiment uses the PINNs network as the model network architecture. During the network training process, the loss function is an important component. This embodiment uses the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal to construct the first loss function, and uses whether the predicted category output by the PINNs network satisfies the wave equation as the constraint condition of the PINNs network output to construct the second loss function. That is, the loss function includes not only data errors, but also residuals based on physical laws. When PINNs is applied to the feature extraction of ship radiation noise signals, the loss function consists of data-based losses and physics-based losses.

[0079] Among them, the establishment and initial setting of the PINNs network are: given an input domain and time interval The objective function on , PINNs network right Make predictions, Represents the trainable parameters of the network (i.e., network parameters). The framework of PINNs usually adopts a fully connected feedforward neural network consisting of an input layer, multiple hidden layers, and an output layer. For example, the network is fed into Predict the physical quantities required The output of the PINNs network can be expressed as:

[0080] ;

[0081] In the formula, , and Representation layer i The weights and biases of Represents the activation function.

[0082] In one example, since the loss function in PINNs is constructed by embedding physical equations, these equations are usually expressed as strong or weak forms of physical laws, that is, the wave equations constructed according to the physical laws followed by the radiation noise signal during the propagation process are expressed in the strong or weak form of the physical laws followed by the radiation noise signal during the propagation process.

[0083] If the wave equation is expressed in the strong form of the laws of physics, then the first loss function consists of the residual of the wave equation, e.g., for For a controlled system, the loss function consists of the residual of the control equation:

[0084] ;

[0085] If the wave equation is expressed in the weak form of the laws of physics, the first loss function is derived from the energy functional of the wave equation, for example, the energy functional is , then the goal becomes:

[0086] .

[0087] In the specific implementation, the data-based loss (ie, the first loss function) is:

[0088] Used to measure the difference between the output of a neural network and the actual class label. Cross entropy is particularly suitable for classification tasks, where the goal is to minimize the difference between the predicted probability and the actual label. The formula for cross entropy is:

[0089] ;

[0090] In the formula, and It is a sample i Category c The actual label and predicted probability of C and M is the number of categories and the total number of samples.

[0091] Physically based loss (i.e. the second loss function):

[0092] The physics-based loss is a special term introduced by PINNs to ensure that the model complies with relevant physical laws. By embedding the physical laws WE followed by the radiation noise signal during propagation as physical constraints into the loss function, known physical knowledge can be used to enhance the generalization ability of the model, even in the absence of a large amount of labeled data.

[0093] In order to ensure that the output characteristics of the network meet the WE requirements, it is necessary to impose constraints on the equations during the training process. The output of the network should satisfy the following constraints, that is, whether the predicted category of the PINNs network output satisfies the wave equation. The constraints for the PINNs network output are:

[0094] ;

[0095] Therefore, the calculation formula of the physics-based loss (i.e., the second loss function) is:

[0096] ;

[0097] In the formula, and Respectively represent the second-order derivatives of the output of the PINNs network with respect to time and space, M Indicates the number of samples.

[0098] By summing and minimizing the physical residuals, the network is forced to adjust its output so that it not only fits the data but also satisfies the constraints of WE.

[0099] Therefore, the final total loss function of the network is a weighted sum of data-based loss and physics-based loss, which is in the form of:

[0100] ;

[0101] In the formula, and Represents the preset weight parameter.

[0102] Step 104: Use the radiation noise signal to train the PINNs network, and optimize the network parameters of the PINNs network according to the first loss function and the second loss function during the training process.

[0103] Specifically, during the training process, the network parameters of the PINNs network are optimized according to the respective weights of the first loss function and the second loss function, with the goal of minimizing the first loss function and the second loss function; wherein the respective weights of the first loss function and the second loss function will be dynamically adjusted during the training process.

[0104] In the specific implementation, in order to balance the impact of different loss components, PINNs can dynamically adjust weights during training. One way is to increase the boundary condition weights to ensure that the model first meets the boundary conditions, and then adjust the weights of other parts accordingly. Use an optimization algorithm to find the optimal neural network parameters with the loss function minimized as the optimization goal.

[0105] During training, the loss function and its gradient are calculated to update the network parameters. For the Adaptive Moment Estimation (Adam) optimization algorithm, each parameter is updated as follows:

[0106] ;

[0107] In the formula, represents the network parameters, represents the optimized network parameters, and denote the exponentially weighted moving averages of the gradient and squared gradient, respectively. represents the learning rate, Represents a preset numerical stabilization term.

[0108] It is understandable that the key links of passive underwater target recognition mainly include feature extraction, feature selection and classification decision. That is, several key links of target recognition have made great progress, laying a good foundation for the development of this technology. However, when facing practical applications, these research fields still face many problems, such as small data volume, low matching degree between recognition models and actual physical problems, etc. In order to reduce the impact of the above problems on the recognition results, this embodiment is equivalent to proposing a ship radiation noise recognition method based on PINNs. This method introduces physical losses based on wave equations to constrain the training process. The technical flow chart of this method is shown as follows: Figure 2 shown.

[0109] This method is mainly divided into two parts, namely feature extraction and classification prediction. The input is the ship radiation noise signal after standardization, that is, a series of time series data. The output is the prediction result of the classification label. For the model training process, the strategy of first division and then enhancement is adopted. First, the data set is divided into training set, validation set and test set in a ratio of 8:1:1, and then the data is enhanced by randomly adding 10dB-15dB Gaussian white noise. This operation can avoid the leakage of feature information of the training set to the test set, greatly reducing the probability of overfitting.

[0110] During training, domain and boundary points are fed into the network, which undergoes forward and backward propagation iterations to minimize the loss. Also, predictions are made on a set of data not used in training in order to assess the physical consistency and numerical accuracy of the model.

[0111] Step 105, identifying the ship radiation noise through the trained PINNs network.

[0112] Specifically, the final output of the trained PINNs network is is a solution that approximates the true physical behavior. This output can be compared with known analytical solutions or experimental results to assess accuracy.

[0113] In this embodiment, by introducing a wave equation constructed according to the physical laws followed by the radiation noise signal during propagation, the loss function includes not only the loss of training data (i.e., the first loss function), but also the physical loss based on the wave equation constructed by the wave equation (i.e., the second loss function). When the radiation noise signal generated by the ship during navigation is used to train the neural network, the training process is constrained by the physical loss, and the actual physical phenomena are taken into account, so that the finally trained neural network model has better recognition performance. In addition, the addition of the above physical loss makes the loss based on the training data no longer the main factor affecting the model training, so it is no longer necessary to use a large-scale data set for training, which reduces the probability of overfitting of the model.

[0114] The following is a specific example to verify the ship radiation noise identification method of the present invention:

[0115] (1) Dataset used in the experiment

[0116] The dataset consists of two parts: the ShipsEar dataset and the NWPU dataset. In order to increase the diversity of samples, the ShipsEar dataset and the NWPU dataset are combined to create a new dataset, called the Integral dataset. In addition, due to the significant differences between the two datasets, this approach also helps to demonstrate that our model has strong versatility and adaptability. A detailed introduction to the ShipsEar dataset and the NWPU dataset is as follows:

[0117] i. ShipsEar dataset

[0118] The ShipsEar dataset was developed to address the shortage of available audio data for underwater acoustic research, providing a rich collection of underwater ship noise recordings. The dataset includes recordings from 11 different types of ships, such as fishing boats, ferries, ocean liners, and container ships, as well as natural background noise such as wind, rain, and waves. The recordings were captured using a digitalHyd SR-1 autonomous acoustic recorder equipped with a hydrophone with a sensitivity of -193.5 dB re 1V / 1 uPa and a frequency response of 1Hz to 28kHz. The recording system operates at a sampling rate of 52734Hz, and each recording is recorded in detail, including the type of ship, environmental conditions, GPS position, and depth of the hydrophone. In order to apply the dataset to classification experiments, the dataset authors divided it into five categories and named them Class A, Class B, Class C, Class D, and Class E. Figure 3 shows timing plots of some signals in the ShipsEar dataset.

[0119] ii. NWPU dataset

[0120] The ship radiated noise data in the NWPU dataset comes from the ship signal acquisition experiment. In this experiment, 61 ship radiated noise data were collected, and all signals were sampled at 20kHz. The NWPU dataset includes three types of signals, named Class F, Class G, and Class H. Figure 4 shows the timing diagrams of some signals in the NWPU dataset.

[0121] iii. Integral dataset

[0122] Since there are differences in the sampling frequency and signal length of the two data sets, the following steps are taken to achieve the experimental goals: 1. Merge the data sets. 2. Ensure that the Integral data set is suitable for comparative experiments. 3. Reduce the impact of uneven sample distribution on the experimental results. First, 400,000 points are intercepted from each sample as new samples. Then, each type of signal is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. Finally, the data enhancement method is used to increase the number of samples of each type of signal to 50, that is, Gaussian white noise with a signal-to-noise ratio of 10dB-15dB is randomly added to each sample.

[0123] (2) Experimental configuration

[0124] This experiment was conducted on a Windows 11 operating system with an AMD EPYC 7542 CPU and an NVIDIA RTX 4090 GPU (24GB VRAM), as well as 256GB of system memory. The software environment includes MATLAB 2022a and Python 3.12, and the deep learning framework PyTorch supports CUDA 11.7 acceleration. The experiment utilizes the PINNs architecture, which consists of a feature extraction module and a classifier module using a Sin activation function. The batch size is set to 256, the number of iterations is set to 100, and the initial learning rate is set to 0.001. The initial weights of the data-based loss term and the physics-based loss term are both set to 1, and the optimizer is Adam.

[0125] i. Recognition experiments based on different datasets

[0126] The prepared original data set is used as the input of the model after normalization. The purpose of normalization is to reduce the impact of data differences on feature extraction and subsequent classification prediction. At this time, the minimum-maximum normalized data fluctuation range is limited to [0, 1].

[0127] Theoretically, due to the addition of physical loss in PINNs, the wave equation followed by the ship radiation noise signal during propagation can be taken into account. The data-based loss is no longer the main factor affecting model training. Therefore, the proposed recognition model no longer needs to be trained using a large-scale data set, and the probability of overfitting is also reduced. These advantages have greatly reversed the disadvantages of the difficulty in collecting underwater acoustic signal data and the lack of large-scale data sets. In order to link theory with practice and verify whether the performance of the proposed recognition model is consistent with reasoning, numerical experiments were conducted on the models with and without embedded physical losses using the ShipsEar dataset, NWPU dataset, and Integral dataset. The experimental results for different datasets are shown in Table 1.

[0128] Table 1

[0129]

[0130] On all data sets, the performance of the proposed method is significantly better than the baseline method, and there are comprehensive improvements in the four key indicators of Precision, Recall, F1-Score and AUC. In particular, on the ShipsEar data set and the Integral data set, the model embedded with physical loss has improved by more than 10% in AUC and F1-Score, respectively, indicating that it has a strong learning ability for small-scale data sets and can better avoid overfitting. In addition, the AUC value of each data set is high because AUC is a global indicator that measures the model's ability to distinguish among all categories. Even if the accuracy of some classifications is low, AUC can still show the good performance of the model as a whole. Although the indicator values ​​of each data set in the baseline method are above 0.7, in the actual training process, the probability of overfitting is high, and the parameters need to be adjusted multiple times to avoid overfitting. These results fully verify the effectiveness of introducing physical constraints in the network, highlight the robustness and generalization ability of the method embedded with physical loss when data collection is difficult and the data set is small, and further verify the consistency between theoretical derivation and practical application.

[0131] ii. Recognition experiments based on different methods

[0132] Four recognition methods based on deep learning (CNN and RNN) and modal decomposition (VMD and CEEMDAN) are compared. For ease of distinction, these four methods are referred to as RM1, RM2, RM3 and RM4 respectively. RM5 adds denoising to the original signal on the basis of the method proposed in the present invention, RM6 uses the MFCC features of the original signal as the input of PINNs, and RM7 is a combination of RM5 and RM6. Since the data set of this embodiment is not a standard data set, the network structure of other methods is modified to adapt to the experimental data set, and efforts are made to ensure that the core ideas of these methods are realized. After conducting multiple experiments and optimizing the hyperparameters, the experimental results of different methods are shown in Table 2.

[0133] Table 2

[0134]

[0135] As can be seen from Table 2, the method proposed in the present invention is significantly better than other methods on the Integral dataset, and is ahead in all four indicators, among which AUC reaches 0.9193, far exceeding the optimal value of 0.8325 of the comparison method. After extracting the MFCC features of the signal, some noise features may overwhelm the original features, and the dimension of the data is greatly reduced, which is not conducive to the training of the neural network-based method, so RM6 and RM7 are seriously overfitting. Compared with RM6 and RM7, the method proposed in the present invention significantly reduces the risk of overfitting by introducing physical constraints, while improving the generalization ability, especially in the case of data scarcity, showing stronger robustness and practical application potential. Considering the computational and time complexity, other comparison methods need to spend a lot of time and computing resources to preprocess the original dataset, while the proposed method only needs to normalize it. The above experimental results fully verify the effectiveness of PINNs and its adaptability to the field of ship radiated noise signals.

[0136] The step division of the various methods above is only for clear description. When implemented, they can be combined into one step or some steps can be split and decomposed into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of the present invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process without changing the core design of the algorithm and process are all within the protection scope of the invention.

[0137] Another embodiment of the present invention relates to a ship radiation noise identification system. The implementation details of the ship radiation noise identification system of this embodiment are specifically described below. The following content is only for the convenience of understanding the implementation details provided, and is not necessary for the implementation of this solution. The ship radiation noise identification system of this embodiment includes:

[0138] A signal acquisition module is used to acquire the radiation noise signal generated by the ship during navigation;

[0139] An equation building module, which is used to build the wave equation according to the physical laws that the radiated noise signal follows during the propagation process;

[0140] A function construction module is used to construct a first loss function based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and to construct a second loss function based on whether the predicted category output by the PINNs network satisfies the wave equation as a constraint condition for the PINNs network output;

[0141] A network training module, used to train the PINNs network using a radiation noise signal, and optimize the network parameters of the PINNs network according to a first loss function and a second loss function during the training process;

[0142] The ship identification module is used to identify ship radiated noise through the trained PINNs network.

[0143] It is not difficult to find that this embodiment is a system embodiment corresponding to the above method embodiment, and this embodiment can be implemented in conjunction with the above method embodiment. The relevant technical details and technical effects mentioned in the above embodiment are still valid in this embodiment, and in order to reduce repetition, they are not repeated here. Accordingly, the relevant technical details mentioned in this embodiment can also be applied in the above embodiment.

[0144] It is worth mentioning that all modules involved in this embodiment are logic modules. In practical applications, a logic unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of the present invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other units in this embodiment.

[0145] Another embodiment of the present invention relates to a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the ship radiation noise identification method in the above-mentioned embodiments.

[0146] Among them, the memory and the processor are connected in a bus manner, and the bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be one element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices on a transmission medium. The data processed by the processor is transmitted on a wireless medium via an antenna, and further, the antenna also receives data and transmits the data to the processor.

[0147] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory can be used to store data used by the processor when performing operations.

[0148] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program, which implements the above method embodiment when executed by a processor.

[0149] That is, those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, and the program is stored in a storage medium, including a number of instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as: ROM), random access memory (Random Access Memory, referred to as: RAM), disk or optical disk and other media that can store program codes.

[0150] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present invention, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for identifying ship radiated noise, characterized in that: The method comprises: Acquire the radiation noise signal generated by the ship during navigation; According to the physical laws followed by the radiation noise signal during the propagation process, the wave equation is constructed; The first loss function is constructed based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and the second loss function is constructed based on whether the predicted category output by the PINNs network satisfies the wave equation as the constraint condition of the PINNs network output; The PINNs network is trained using a radiation noise signal, and the network parameters of the PINNs network are optimized according to a first loss function and a second loss function during the training process; Ship radiation noise is identified through the trained PINNs network; Wherein, the wave equation is: ； In the formula, u and c represent the sound pressure and sound speed respectively, represents the Laplace operator.

2. The ship radiation noise identification method according to claim 1, characterized in that: Whether the prediction category of the PINNs network output satisfies the constraint condition that the wave equation is the PINNs network output is: ； The second loss function is: ； In the formula, and Respectively represent the second-order derivatives of the output of the PINNs network with respect to time and space, M Indicates the number of samples.

3. The ship radiation noise identification method according to claim 1, characterized in that: The wave equation is expressed in a strong form or a weak form of the physical law followed by the radiation noise signal during the propagation process; The first loss function comprises the residual of the wave equation if the wave equation is expressed in the strong form of the laws of physics and is derived from the energy functional of the wave equation if the wave equation is expressed in the weak form of the laws of physics.

4. The ship radiation noise identification method according to claim 1, characterized in that: The method of optimizing the network parameters of the PINNs network according to the first loss function and the second loss function during the training process includes: During the training process, the network parameters of the PINNs network are optimized according to the respective weights of the first loss function and the second loss function with the goal of minimizing the first loss function and the second loss function; wherein the respective weights of the first loss function and the second loss function will be dynamically adjusted during the training process.

5. The ship radiation noise identification method according to claim 4, characterized in that: The network parameters of the PINNs network are optimized by the following formula: ； In the formula, represents the network parameters, represents the optimized network parameters, and denote the exponentially weighted moving averages of the gradient and squared gradient, respectively. represents the learning rate, Represents a preset numerical stabilization term.

6. The ship radiation noise identification method according to claim 1, characterized in that: The first loss function is constructed based on the cross entropy loss function.

7. A ship radiation noise identification system, characterized in that: The system comprises: A signal acquisition module is used to acquire the radiation noise signal generated by the ship during navigation; An equation building module, which is used to build the wave equation according to the physical laws that the radiated noise signal follows during the propagation process; A function construction module is used to construct a first loss function based on the difference between the actual category of the ship indicated by the radiation noise signal and the predicted category of the ship output by the physical information neural network PINNs according to the radiation noise signal, and to construct a second loss function based on whether the predicted category output by the PINNs network satisfies the wave equation as a constraint condition for the PINNs network output; A network training module, used to train the PINNs network using a radiation noise signal, and optimize the network parameters of the PINNs network according to a first loss function and a second loss function during the training process; Ship identification module, used to identify ship radiation noise through the trained PINNs network; Wherein, the wave equation is: ； In the formula, u and c represent the sound pressure and sound speed respectively, represents the Laplace operator.

8. A computer device, characterized in that: include: at least one processor; And, a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the ship radiation noise identification method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the ship radiation noise identification method according to any one of claims 1 to 6 is implemented.

Images