Underwater effective positioning method of multiple vibration sources based on stockwell transform time-frequency diagram

Abstract

The application discloses a kind of underwater effective positioning methods of multiple vibration sources based on Stockwell transform time-frequency diagram.The method includes three steps of data acquisition, data preprocessing and model training.In the process of data acquisition, the vibration pressure signal of the set frequency is collected using the side line array, and the data is saved in real time by the control software of computer;In the data preprocessing stage, the signal of the sensor is removed noise, and the signal is subjected to Stockwell transform time-frequency diagram;In the process of using model training, the normalized combined time-frequency diagram is input, and the effective vibration source position prediction value is obtained.The method makes full use of the spectral characteristics of time series data, solves the problem that the original time series data characteristics are not obvious and difficult to extract, and the process is simple and easy to operate.The prediction algorithm used can be used for underwater vehicle cluster, detection, concealment and other tasks, avoids the danger of exposing the position by using sound wave positioning, and improves the success rate of underwater vehicle task execution.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of underwater positioning, in particular to a multi-vibration source underwater effective positioning method based on Stockwell transform time-frequency diagram. BACKGROUND

[0002] Underwater positioning technology is the most important of underwater detection technology, how to accurately position in the complex marine environment is a difficult problem to be solved.

[0003] Fish and some amphibians have very sensitive underwater sensing ability, which comes from the lateral line system on both sides of the fish body. The lateral line system is an important organ for these animals to obtain information about the surrounding environment. With the help of the lateral line organ, fish and other animals can achieve behaviors such as group activity, prey positioning, obstacle avoidance, target recognition, etc. in complex underwater environments. Based on the characteristics and mechanism of the lateral line system of fish and other animals to perceive the surrounding flow, a new idea is provided for the accurate perception of the surrounding flow field by underwater unmanned underwater vehicles.

[0004] In traditional underwater detection technology, underwater positioning technology based on acoustics is widely studied by researchers and has achieved good positioning results. However, for near-field targets, it is easy to be affected by mutual interference, such as noise between targets. Sonar and visual detection technology have poor effect in specific environments (narrow space, turbid water). Moreover, due to the natural barrier effect of seawater, GPS positioning of land navigation cannot be used for underwater positioning. Traditional optical detection equipment is limited by factors such as underwater light refraction and insufficient light, which greatly reduces the positioning effect. Therefore, it is of great strategic significance to develop a new type of underwater positioning technology to ensure the successful execution of underwater vehicle missions.

[0005] At present, the signal processing for underwater positioning is generally divided into two directions. One is to establish a dipole source flow field model and use various constraint methods to analyze the model. Some of the ideas of these methods are: using Gauss-Newton iteration and Newton-Raphson method to solve linear equations to estimate the vibration direction and amplitude of the dipole source (vibration source), or using beamforming algorithm to solve the positioning problem. The second is to establish a neural network model to process the received signal, to simulate the behavior of fish and other animals in complex environments through years of training to complete various activities, to build a neural network model to complete the positioning task after data training, such as BP neural network, multilayer perception network, etc. By establishing a nonlinear relationship between the extracted features and the positioning coordinates, the trained model is further used to predict the features of the coordinates as the input of the model, and the predicted value of the test coordinates is obtained. However, the existing methods have the problem that the information contained in the data is not fully tapped, resulting in poor prediction effect. At the same time, the existing features use strong prior knowledge to artificially extract features, such as extracting the amplitude of the vibration frequency, which does not match the actual positioning conditions of the environment, resulting in its inability to be used in practice. And when there are multiple targets vibrating at the same time, the existing feature extraction method is not enough to reflect the complex flow field environment. SUMMARY

[0006] In order to fully utilize the pressure data collected by the artificial lateral line array, solve the problem that the existing underwater positioning method has large positioning error for multiple vibration sources and cannot be used in practice due to the feature extraction method. The present application provides a feature extraction method that does not require strong prior knowledge, does not require the use of complex mathematical models for positioning method, can fully tap the information in the data, and can accurately establish the nonlinear relationship between the flow field pressure data generated by vibration and the vibration source position coordinates. The predicted vibration source position coordinates provide a method for guiding the near-field target positioning of underwater vehicles.

[0007] The technical scheme adopted by the present application to solve its technical problems is: a multi-vibration source underwater effective positioning method based on Stockwell transform time-frequency diagram, comprising the following steps:

[0008] Step 1: Use 3D printing technology to print the shell of the designed artificial lateral line sensor array, and assemble commercial pressure sensors to build the sensor array.

[0009] Step 2: Randomly generate a plurality of groups of vibration source position groups, Target = (x1, y1, x2, y2), where (x1, y1) and (x2, y2) are two-dimensional coordinates of two vibration sources, respectively. Move the vibration source according to the generated coordinates to determine the vibration source at the set target position.

[0010] Step 3, at each position, use the designed artificial side line sensor array to obtain the underwater flow field pressure signal in a period of time, use the host computer software and data acquisition card to save the pressure signal with a sampling frequency of more than 2 times the vibration frequency for 10-20 seconds on the computer, and the related pressure signal data is X, X=[x i (t)],i=1…,N;Use x i (:) to represent the sampling input of the sensor array once.

[0011] Step 4, data preprocessing is performed on the pressure signal data obtained in step 3, specifically, the signal of each sensor is subtracted by the average value to remove the noise, and the coordinates randomly generated in step 2 are also normalized to make the size of the target value within the same scale range.

[0012] Further, in step 4, data preprocessing refers to:

[0013] x i (t)=x i (t)-mean(x i (t))

[0014] Wherein, x i (t) is a column component in X, representing the signal data collected by the corresponding sensor, and mean(x i (t)) is the average value of x i (t).

[0015] Further, in step 4, normalization refers to:

[0016]

[0017] Wherein, (x i ,y i ) represents the two-dimensional coordinates of the vibration source, max((x i ,y i )) and min((x i ,y i )) are the maximum and minimum values of (x i ,y i ) respectively.

[0018] Step 5, in order to expand the sample number of data, use data augmentation method to divide the original data M data points into q equal parts, and each part of sample data contains m data points.

[0019] Further, in step 5, the data augmentation method refers to:

[0020] x(i,j)[N,M] / q=x(x,j′)[N,m]

[0021] Wherein, x(i,j) represents the data matrix collected, the size is N*M, x(i,j') represents the data matrix after segmentation, the size is N*m.

[0022] Step 6, using Stockwell transform of multi-frequency analysis to analyze the signal in time-frequency, using discrete Stockwell transform to convert one-dimensional form of pressure change data into two-dimensional time-frequency diagram format, which is convenient for subsequent use of the model to extract features and nonlinear fitting of the features.

[0023] Further, in step 6, Stockwell transform refers to:

[0024]

[0025] Wherein, τ is the time related to the time axis of the control window function, h(t) is the analysis signal, f is the frequency, and S(τ,f) is the time-frequency spectrum matrix obtained by transformation.

[0026] Further, in step 6, the discrete Stockwell transform refers to:

[0027] For discrete signal, let τ=jT, N is the length of discrete signal, T is the sampling time interval, then the Stockwell transform of discrete signal is:

[0028]

[0029] Wherein, m,n=0,1,…,N-1 represents the discrete data points, and H represents the discrete Fourier spectrum.

[0030] Step 7, the resolution of the time-frequency diagram generated in step 6 is normalized to W0*H0, and N sensor arrays correspond to N time-frequency diagrams.

[0031] Further, in step 7, the image resolution is normalized to W0*H0, which refers to:

[0032]

[0033] Wherein, F(i',j') represents a new pixel point, M,N represents the pixel field range, dx represents the decimal coordinate in x direction, dy represents the decimal coordinate in y direction, and R represents the pixel value.

[0034] Step 8, combining the N time-frequency diagrams normalized in step 7 into one diagram, the resolution of which is W*H.

[0035] Step 9, dividing the sample image data set obtained in step 8 into training set, validation set and test set according to certain proportion.

[0036] Step 10, a positioning prediction model (SCFANet) of residual convolutional neural network based on channel self-attention mechanism and spatial self-attention mechanism is constructed, all inputs are preprocessed by a preprocessing module before feature extraction, the model is trained by using a training set, the training weight of the model is verified by using a verification set, and the optimal weight is selected according to the verified performance.

[0037] Further, the model constructed in step 10 comprises:

[0038] a preprocessing module, a feature extraction module and a fully connected regression module; the preprocessing module adopts a 2D convolutional layer, and a batch normalization layer and a ReLu function are added; the feature extraction module comprises four block structures, one block structure is composed of a local residual learning and a feature attention (FA) module, the local residual learning part is composed of two convolutional layers, and a batch normalization layer, a ReLu function and a Dropout layer are followed after each convolutional layer; the fully connected comprises a fully connected layer.

[0039] Further, the preprocessing process of the preprocessing module comprises:

[0040] scaling all time-frequency graph image data in the training set, the verification set and the test set to a unified size.

[0041] normalizing all time-frequency graph image data in the training set, the verification set and the test set to a distribution state with the same mean and the same variance.

[0042] Step 11, according to the optimal weight trained in step 10, the test set is tested to obtain the predicted value of the underwater vibration source position.

[0043] Compared with the prior art, the present application has the beneficial effects that:

[0044] The present application fully mines the collected flow field pressure data, and in view of the fact that the existing neural network model underwater positioning method generally adopts the method of manually extracting the vibration frequency amplitude, the Stockwell time-frequency analysis method is selected, the defect of the fixed window function width of the short-time Fourier method (STFT) is overcome, a Gaussian window function varying with frequency is adopted, the characteristics of multi-frequency analysis are utilized, a narrow window is used for high frequency and a wide window is used for low frequency, so that the time-frequency diagram can obtain the frequency information at a certain time and the amplitude information of the signal at a certain frequency. The convolutional neural network model of the self-attention mechanism is constructed to automatically extract features from the time-frequency diagram, and the defect of manually extracting features with prior knowledge is solved. Moreover, the local residual structure of the network allows less important feature information to be filtered out, so that the overall model pays more attention to effective information. At the same time, the self-attention mechanism adopted makes it quite flexible to process different features, suppresses the information that is not important to the result, and increases the influence of the significant features of the result. The more accurate and detailed features are extracted from the time-frequency image data, and the accuracy of underwater multi-vibration source positioning is improved. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 Schematic diagram of an artificial lateral line sensory array designed using 3D printing technology;

[0046] Figure 2 Schematic diagram of an experimental platform for collecting data;

[0047] Figure 3 Schematic diagram of pressure data collected when the vibration source is located at (2, -8, 12, -6);

[0048] Figure 4 Flowchart of a multi-vibration source underwater effective positioning method based on Stockwell transform time-frequency diagram;

[0049] Figure 5 Stockwell transform time-frequency diagram and network input image after combination thereof;

[0050] Figure 6 Schematic diagram of the structure of the network model and schematic diagram of the basic block structure;

[0051] Figure 7 Comparison diagram of the predicted values and the true values of the first four groups in the test set. DETAILED DESCRIPTION

[0052] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0053] The present application provides a multi-vibration source underwater effective positioning method based on Stockwell transform time-frequency diagram, comprising the following steps:

[0054] Step 1: Use 3D printing technology to print the housing of the designed artificial lateral line sensor, and assemble the commercial pressure sensor MPXV5004 to build a sensor array.

[0055] Figure 1 The diagram shows an artificial lateral line sensor array designed using 3D printing technology. The sensor array is parallel to the surface, which is designed to better mimic the distribution of nerve mounds in the lateral line system of fish to sense flow pressure information, thus enabling the sensor array to acquire sufficient flow field information.

[0056] Step 2: Randomly generate multiple sets of vibration source positions, Target = (x1, y1, x2, y2), where (x1, y1) and (x2, y2) are the two-dimensional coordinates of two vibration sources, respectively. Move the vibration sources according to the generated coordinates to determine the vibration source at the set target position. To better obtain information about the flow field during vibration, the coordinates of multiple vibration sources are randomly generated instead of manually set, making the experimental environment more realistic.

[0057] Figure 2 The diagram shows a schematic of the experimental platform for data collection.

[0058] Step 3: At each location, use the designed artificial lateral line sensor array to obtain the underwater flow field pressure signal over a period of time. Use DAQmax software and an NIX6002 acquisition card to save the pressure signal data for 14 seconds at a sampling frequency of 1kHz on the computer. The relevant pressure signal data is denoted as X, where X = [x i [(t)], i = 1…, 8; using x i (:) represents a single sampling input of the sensor array.

[0059] Figure 3 The diagram shows a time-domain representation of pressure data collected by the sensor array.

[0060] Step 4 involves preprocessing the pressure signal data obtained in Step 3. Specifically, this involves subtracting the average value from the signal from each sensor to remove background noise. This is because the static pressure measured by the pressure sensors is very large, while the collected pressure variation is relatively small. To facilitate better feature extraction later, it is necessary to remove static pressure background noise. Specifically, this refers to:

[0061] x i (t)=x i (t)-mean(x i (t))

[0062] Where, x i (t) is a column component of X, representing the signal data acquired by the corresponding sensor, mean(x i (t) is xi The average value of (t).

[0063] Step 5: Normalize the coordinates randomly generated in Step 2 to ensure that the target values ​​are within the same scale range, facilitating model fitting. Specifically, this means:

[0064]

[0065] Among them, (x i ,y i ) represents the two-dimensional coordinates of the vibration source in the plane, max((x i ,y i )) and min((x i ,y i )) are respectively (x i ,y i The maximum and minimum values ​​of ).

[0066] Step 6: The amount of sample data collected by the experimental platform is too small for the neural network model, which can easily lead to overfitting. To increase the number of samples, data augmentation is used to divide the original data into equal parts, with each part containing 1000 data points. Specifically:

[0067] x(i,j)[8,14000] / 14=x(x,j′)[8,1000]

[0068] Where x(i,j) represents the collected data matrix with a size of 8×14000, and x(x,j′) represents the matrix after data segmentation with a size of 8×1000.

[0069] Step 7: The Stockwell transform overcomes the limitation of the short-time Fourier transform window being invariant, adaptively adjusting the analysis window width according to frequency changes and providing intuitive time-frequency characteristics. The Stockwell method with multi-frequency analysis is used to perform time-frequency analysis on the acquired pressure signal. Considering that the pressure signal is discrete data points, discrete Stockwell is used to convert the one-dimensional pressure change data into a two-dimensional time-frequency graph format, facilitating subsequent feature extraction using the built model and nonlinear fitting of the features. Specifically, the Stockwell method refers to:

[0070]

[0071] Where τ is the time related to the control window function on the time axis, h(t) is the analyzed signal, f is the frequency, and S(τ,f) is the transformed time-spectrum matrix.

[0072] For discrete signals, let T, N is the length of the discrete signal, and T is the sampling time interval. The discrete Stockwell transform here has the specific form:

[0073]

[0074] where m, n = 0, 1, …, N-1 represent discrete data points, and H represents the discrete Fourier spectrum.

[0075] Step 8, in order to facilitate the combination of subsequent images, the generated time-frequency map is selected to be unified into the same scale. Here, the picture resolution is unified, specifically referring to:

[0076]

[0077] where F(i', j') represents a new pixel point, M, N represents the pixel field range, dx represents the decimal coordinate in the x direction, dy represents the decimal coordinate in the y direction, and R represents the pixel value.

[0078] Step 9, the eight time-frequency maps normalized in step 8 are combined into one map, and the resolution of this map is 256x512.

[0079] Figure 5 The time-frequency map after Stockwell transform and the combined network input image.

[0080] Step 10, the sample image dataset obtained in step 8 is divided into training set, validation set and test set according to the ratio of 7:1:2. In order to better reflect the fitting performance of the model, the training dataset, the validation dataset and the test dataset are randomly generated from the image database according to the corresponding ratio. Finally, the average value of the test set is taken as the evaluation index of the model.

[0081] Step 11, a residual convolutional neural network positioning prediction model (SCFANet) based on channel self-attention mechanism and spatial self-attention mechanism is constructed, all inputs are preprocessed before feature extraction, the model is trained using the training set, the training weight of the model is verified using the validation set, and the optimal weight is selected according to the verification performance. The preprocessing module, the feature extraction module and the fully connected regression module; the preprocessing module adopts a 2D convolutional layer, and adds a batch normalization layer and a ReLu function; the feature extraction module includes four block structures, and each block structure is composed of a local residual learning and a feature attention (FA) module. The local residual learning part is composed of two convolutional layers, and each convolutional layer is followed by a batch normalization layer, a ReLu function and a Dropout layer, and the Dropout layer is set to prevent the model from overfitting; the fully connected includes a fully connected layer.

[0082] Step 12, optionally, as an embodiment of the application, the method further comprises: shrinking the values of the pixels of the time-frequency images in the training set, the validation set and the test set to a fixed interval; adjusting all time-frequency image data in the training set, the validation set and the test set to the same mean and the same variance distribution state; and scaling all time-frequency images in the training set, the validation set and the test set to a uniform size. Specifically, in this embodiment, the input image data is a color three-channel image, and the mean and variance of the three channels of RGB are [0.5, 0.5, 0.5] and [0.5, 0.5, 0.5] respectively, so that the pixel points in the image are shrunk to between 0 and 1, and all input images of 256x512 are scaled to a size of 256x256.

[0083] Step 13, in order to strengthen the fusion between the bottom layer and the high layer features, a jump connection is added on the basis of the basic block residual structure, the features after the attention module are fused with the features before the basic block structure, so that the pre-trained model retains sufficient local feature information and improves the generalization of the model.

[0084] Step 14, in order to make the model pay more attention to the features significant for positioning, a channel self-attention mechanism and a spatial self-attention mechanism are added after the residual structure feature extraction, realizing the separation of the channel and spatial dimensions. First, the feature map is globally pooled to obtain the channel corresponding weight through the MLP network, so that the feature map of the significant channel increases in proportion. Then, the feature map is maximally pooled and averagely pooled in the channel dimension, and the weight of the feature map in the spatial dimension is obtained through training, so that the weight of the spatial feature map increases, which is conducive to the extraction of the features significant for positioning.

[0085] Figure 6 The structure of the network model and the basic block structure are shown, wherein, in this embodiment, the convolution kernel size K=3, Padding=1, the activation function is selected as ReLu function, and the ratio of the Dropout layer is selected as 0.2 for the pre-processing convolution layer and the basic block structure convolution layer. The model is constructed by four basic block structures.

[0086] Step 15, finally, the features are regressed and predicted through a fully connected layer, and the pre-trained model is saved.

[0087] Step 16, according to the optimal weight trained in step 15, the test set is tested to obtain the predicted value of the underwater vibration source position.

[0088] Figure 7 The comparison chart of the predicted values and the true values of the first four groups in the test set is shown.

[0089] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for effective underwater positioning of multiple vibration sources based on Stockwell transform time-frequency map, characterized in that, The method comprises the following steps: Step 1, printing the shell of the designed artificial side line sensor array by using 3D printing technology, assembling the pressure sensor to build the sensor array; Step 2, randomly generating a plurality of groups of position groups of vibration sources, Target=(x1,y1,x2,y2), wherein (x1,y1) and (x2,y2) are two-dimensional coordinates of the two vibration sources respectively, moving the vibration sources according to the generated coordinates, and determining that the vibration sources are at the set target positions; Step 3, at each position, use the designed artificial side line sensor array to obtain the underwater flow field pressure signal in a period of time, use the host computer software and data acquisition card to save the pressure signal with a sampling frequency of more than 2 times the vibration frequency for 10 to 30 seconds on the computer, and the related pressure signal data is X, X=[x i (t)],i=1…,N; use x i (:) to represent the sampling input of the artificial side line sensor array once. Step 4, data preprocessing of the pressure signal data obtained in step 3, specifically, subtracting the average value of the signal of each sensor to remove the noise, and normalizing the coordinates randomly generated in step 2, so that the size of the target value is within the same scale range; Step 5, using a data augmentation method to divide the original data M data points into q equal parts, and each sample data contains m data points; Step 6, using a Stockwell transform of multi-frequency analysis to perform time-frequency analysis on the signal, using a discrete Stockwell transform to convert one-dimensional pressure change data into a two-dimensional time-frequency graph format, facilitating subsequent feature extraction using the built model, and performing nonlinear fitting on the features; Step 7, uniformly normalizing the resolution of the time-frequency graph generated in step 6 to W0xH0, and N sensors correspond to N time-frequency graphs; Step 8, combining the N normalized time-frequency graphs in step 7 into one graph, and the size of the combined image is WxH; Step 9, dividing the sample image data set obtained in step 8 into a training set, a validation set and a test set according to a certain proportion; Step 10, constructing a residual convolutional neural network positioning prediction model based on channel self-attention mechanism and spatial self-attention mechanism, the positioning prediction model comprising a preprocessing module, a feature extraction module and a full connection regression module, all inputs pass through the preprocessing module before feature extraction, the model is trained using the training set, the training weights of the model are verified using the validation set, and the optimal weights are selected according to the verification performance; Step 11, testing the test set according to the optimal weights trained in step 10 to obtain the predicted value of the underwater vibration source position.

2. The Stockwell transform time-frequency map based multi-vibratory source underwater effective positioning method according to claim 1, characterized in that, In step 4, data preprocessing refers to: x i (t) = x i (t) - mean(x i (t)) where x i (t) is the column component in X, representing the signal data collected by the corresponding sensor, mean(x i (t)) is the average value of x i (t). In step 4, normalization refers to: where (x i ,y i ) represents the planar two-dimensional coordinates of the vibration source, max((x i ,y i )) and min((x i ,y i )) are the maximum and minimum values of (x i ,y i ), respectively.

3. The Stockwell transform time-frequency map based multi-vibratory source effective underwater localization method according to claim 1, characterized in that, In step 5, the data augmentation method refers to: x(i,j)[N,M] / q=x(x,j′)[N,m] Wherein, x(i,j) represents the collected data matrix, the size is N*M, x(x,j′) represents the matrix after data segmentation, the size is N*m.

4. The Stockwell transform time-frequency map based multi-vibratory source effective underwater localization method according to claim 1, wherein, In step 6, the Stockwell transform refers to: Wherein, τ is the time related to the time axis of the control window function, h(t) is the analysis signal, f is the frequency, and S(τ,f) is the time-frequency spectrum matrix obtained by transformation; In step 6, the discrete Stockwell transform refers to: For a discrete signal, let τ = jT, N is the length of the discrete signal, and T is the sampling time interval, then the Stockwell transform of the discrete signal is Wherein, m,n=0,1,…,N-1 represents the discrete data points, and H represents the discrete Fourier spectrum.

5. The Stockwell transform time-frequency map based multi-vibratory source effective underwater localization method according to claim 1, wherein, In step 7, the image resolution is uniformly normalized to W0xH0, which refers to: Wherein, F(i', j') represents a new pixel point, M, N represents the pixel field range, dx represents the decimal coordinate in the x direction, dy represents the decimal coordinate in the y direction, and R represents the pixel value.

6. The Stockwell transform time-frequency map based multi-vibratory source effective underwater localization method according to claim 1, wherein, In the step 10, The preprocessing module adopts a 2D convolution layer, and a batch normalization layer and a ReLu function are added; The feature extraction module includes four block structures, and one block structure is composed of a local residual learning and a feature attention module; the local residual learning part is composed of two convolution layers, and each convolution layer is followed by a batch normalization layer, a ReLu function and a Dropout layer; The full connection regression module contains a full connection layer.

7. The method of claim 6, wherein, In the step 10, the preprocessing process of the preprocessing module includes: Scale all time-frequency graph image data in the training set, the validation set and the test set to a unified size; All time-frequency graph image data in the training set, the validation set and the test set are normalized to the same mean and the same variance distribution state.

Images