POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method

Abstract

The invention discloses a POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method, which comprises the following steps: firstly, establishing a pressure data acquisition system, utilizing an excitation unit to cooperate with a disturbance source to simulate disturbance features of a target object, and utilizing a pressure sensor array to acquire pressure data samples at different time periods at the same position of the disturbance source; constructing a feature matrix by using the pressure data sample, performing signal feature extraction based on POD decomposition of the feature matrix, and constructing a pressure data set by using a reconstruction matrix and the real position of the disturbance source as new sample data; an MLP network is adopted as a positioning model, a pressure data set is used for training and testing the MLP network, the precision of the MLP network is verified through pressure data samples before and after reconstruction under the same position label, a trained positioning model is finally obtained, and positioning is carried out based on pressure data collected by an unknown disturbance source. According to the method, the underwater flow field disturbance source position detection precision can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of underwater flow field feature recognition and target detection, specifically involving a method for signal feature extraction and reconstruction based on POD and disturbance source localization combined with MLP. Background Technology

[0002] With the exploration and development of the ocean, the development of underwater object detection technology has played a major role. Sensitive organs such as the lateral line of aquatic vertebrates like fish and the whiskers of seals can detect the hydrodynamic wakes generated by the movement of other objects, enabling them to track prey, avoid obstacles, and swarm. Inspired by this, non-acoustic underwater target localization technology has gradually developed. Combining machine learning methods with training neural network models using large amounts of data to predict the size, shape, and location of targets based on their hydrodynamic characteristics has become a new research trend. Currently, the flow field feature recognition and target detection technology using hydrodynamic information has the following shortcomings: (1) The low signal-to-noise ratio of the signal has become the main factor that inhibits the improvement of the accuracy of underwater flow field feature recognition and target detection based on machine learning methods.

[0003] (2) Existing signal preprocessing methods are mainly based on traditional filtering methods, which are highly dependent on prior information such as the characteristic frequency range of the signal.

[0004] (3) During the detection process, there is a certain noise interference in different input signals under single-location samples, and there is random error in the prediction process of the neural network model. Therefore, the accuracy and reliability of the detection results need to be improved. Summary of the Invention

[0005] The purpose of this invention is to provide a method for signal feature extraction and reconstruction based on POD and for locating disturbance sources by combining MLP, so as to effectively improve the accuracy of detecting the location of disturbance sources in underwater flow fields.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A method for signal feature extraction and reconstruction based on POD and disturbance source localization combined with MLP includes: A pressure data acquisition system is established, comprising a frame and a water tank within the frame. A lead screw guide rail is mounted above the frame, driving a slider to move horizontally. An excitation unit is mounted on the slider, and a disturbance source for simulating the disturbance characteristics of the target object is mounted below the excitation unit via a connecting rod. By utilizing the cooperation of the lead screw guide rail and the slider, and adjusting the length of the connecting rod, the disturbance source can be positioned at different locations within the water tank. The excitation unit generates an excitation signal, driving the disturbance source to perform one-dimensional reciprocating motion within the water tank, thereby generating a disturbance source flow field. A measuring plate parallel to the tank wall is placed within the water tank, and a pressure sensor array is embedded in the measuring plate. The center point of the pressure sensor array is used as the origin of the coordinate system to define the data acquisition area. The disturbance source is positioned at different locations within the data acquisition area, with defined position intervals and distances from the pressure sensor array. Pressure data at different times is collected at each location using the pressure sensor array. At each location of the disturbance source, the pressure data collected by the pressure sensor array are collectively considered as a pressure data sample. The pressure data contained in each pressure data sample are used to form a feature matrix, and intrinsic orthogonal mode decomposition is performed on the feature matrix to extract the eigenvalues. These eigenvalues ​​are then arranged in descending order of energy to select the main spatiotemporal distribution features contained in the pressure data sample, and a corresponding reconstruction matrix is ​​constructed. The location of the disturbance source when the pressure data sample is collected is used as the location label, which, together with the reconstruction matrix, forms a new pressure data sample, thus constructing a pressure dataset. Construct a multilayer perceptron neural network; wherein the input layer dimension of the multilayer perceptron neural network is... ,in This represents the number of pressure sensors in the pressure sensor array. This represents the number of pressure data collected by the pressure sensor at different initial sampling times; multiple hidden layers are used, and the output layer outputs the predicted coordinates of the disturbance source in three-dimensional space. After normalizing the pressure data samples in the pressure dataset, the multilayer perceptron neural network is trained and validated, and the trained multilayer perceptron neural network is saved. For a disturbance source with an unknown location, pressure data is acquired using a pressure sensor array and a corresponding reconstruction matrix is ​​constructed. Multiple pressure data collected by all pressure sensors at different initial sampling times under the reconstruction matrix are input into a trained multilayer perceptron neural network to obtain the location prediction result of the disturbance source.

[0007] Furthermore, the excitation unit includes a modal exciter, a power amplifier, and a signal generator.

[0008] Furthermore, by changing the sampling starting point when the pressure sensor array collects pressure data, pressure data samples at different times at each location of the disturbance source are collected, thus constructing different pressure datasets.

[0009] Furthermore, the construction process of the reconstruction matrix is ​​as follows: Let the number of pressure sensors in the pressure sensor array be . The number of times pressure data was collected at each location of the disturbance source was: Then the characteristic matrix is First, the feature matrix needs to be decentralized: ; in, Representation of the characteristic matrix The decentralized matrix; The mean vector of each pressure sensor at all times; Subsequently, the covariance matrix is ​​constructed. , is represented as: ; Among them, superscript Indicates matrix transpose; For covariance matrix Perform eigenvalue decomposition: ; In the formula, For the first There are 10 eigenvalues, arranged in descending order; For the first Each singular value represents the energy of a different mode; For the first One time feature vector; Spatial eigenvector, i.e., the first Each spatial mode is represented as: ; Spatial modes encompass a certain correlation between pressure data from different pressure sensors, while This includes information about the temporal changes of the signal, and the corresponding modal evolution can be represented as: ; in, Indicates in At the moment The time evolution coefficient of the mode, representing the . Each mode in The level of activity at any given moment; Indicates the first Singular values ​​of each mode, Indicates the first Time feature vectors The One component; Feature matrix It can be approximately represented as the former A linear combination of modalities, reconstructing the matrix exist The column vector at time t is represented as: ; in, For the first One spatial mode; Define the order of the modes to construct the corresponding reconstruction matrix. .

[0010] Furthermore, the input layer dimension of the multilayer perceptron neural network is... The pressure data collected by the nine pressure sensors at different initial sampling times under the corresponding reconstruction matrix are collected at 25 time points. Five hidden layers are used, with the number of neurons in each hidden layer from input to output being 162, 108, 54, 27, and 9, respectively.

[0011] Furthermore, when evaluating the multilayer perceptron neural network, the predicted values ​​of different pressure data samples of the same location label are averaged to obtain the average statistical result of the location coordinates, thereby evaluating the positioning accuracy.

[0012] Furthermore, during the training of the multilayer perceptron neural network, the mini-batch stochastic gradient descent algorithm is used to update the network parameter values. The mean square error between the predicted value obtained from the output layer and the location label is used as the loss function. The loss function is calculated in batches and different parameters are updated along the negative gradient direction of the loss function to achieve the global minimization of the loss function. Through a certain number of iterative training steps, the final parameters of the entire neural network model can be determined.

[0013] A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the POD-based signal feature extraction and reconstruction and the disturbance source localization method combined with MLP.

[0014] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the POD-based signal feature extraction and reconstruction and disturbance source localization method combined with MLP.

[0015] Compared with the prior art, the present invention has the following technical features: 1. This invention utilizes a pressure sensor array to acquire water pressure signals at different times on a measuring plate surface, and trains a multilayer perceptron model to predict the location of different disturbance sources in three-dimensional space. It leverages the hydrodynamic information generated by the disturbance of the target object, effectively aiding in target localization within a small near-field range without interfering with or impacting the marine environment or other aquatic organisms. This provides a new approach for non-acoustic underwater detection and target identification in both civilian and military applications.

[0016] 2. This invention employs intrinsic orthogonal mode decomposition (IOD) to process time-series data acquired by a pressure sensor array. By using a matrix theory-based method, the pressure data is decomposed into several orthogonal bases, which contain the temporal and spatial characteristics of the data. This allows for the extraction and reconstruction of key features while filtering out most noise interference in the signal, without requiring prior knowledge of signal spectral characteristics. As feature value inputs to a multilayer perceptron model, this significantly reduces the prediction error of disturbance source locations.

[0017] 3. This invention uses the average of the prediction results of multiple samples collected at the same location as the final predicted position coordinate value. This method uses statistical averaging to suppress the random interference generated by a single signal during the acquisition process and the random error in position coordinate prediction caused by the instability of the neural network model during the prediction process. It achieves a prediction accuracy that is almost equivalent to that of a multilayer perceptron model based on the input of the signal's main frequency component. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the experimental device used in this invention to generate the underwater disturbance source flow field and acquire pressure signals. Figure 2 This is a schematic diagram of the structure of the multilayer perceptron neural network used in the embodiments of the present invention; Figure 3 This is a flowchart illustrating the disturbance source localization method of the present invention; Figure 4 This is a power spectral density distribution diagram of the pressure signal measured by the sensor at a certain location of the disturbance source in an embodiment of the present invention; Figure 5 (a) to (d) represent the relative prediction errors of the test set data in the three-dimensional position coordinates and straight-line distance of the disturbance source in the embodiments of the present invention. Probability density distribution diagram; Figure 6 These are the relevant fitting parameters for the prediction relative error of the test set data in terms of the three-dimensional position coordinates and straight-line distance of the disturbance source in this invention. Detailed Implementation

[0019] This invention provides a signal feature extraction and reconstruction method based on Point of View (POD) and a disturbance source localization method combined with Multilayer Perceptron (MLP). By reconstructing the signal, the reliance on prior signal information is reduced. The reconstructed high signal-to-noise ratio (SNR) signal and the corresponding disturbance source location are combined to form a dataset, which is then trained and tested using a multilayer perceptron model. Averaging of prediction results for different input signals at the same location is combined to effectively improve the accuracy of underwater flow disturbance source location detection. The steps of this invention are as follows: Step 1: Construct a pressure data acquisition system; use the pressure data acquisition system to acquire pressure data from disturbance sources and perform POD decomposition to construct a pressure dataset.

[0020] (1) Pressure data acquisition system.

[0021] The pressure data acquisition system includes a frame and a water tank within the frame. Above the frame are lead screw guides in the X and Y directions. The Y-direction lead screw guide is movably mounted on the X-direction lead screw guide. An excitation unit is mounted on the Y-direction lead screw guide via a movable slider. Below the excitation unit (in the Z direction), a disturbance source for simulating the disturbance characteristics of a target object is mounted via a connecting rod. By utilizing the cooperation of the lead screw guide and slider, and adjusting the length of the connecting rod, the disturbance source can be positioned at different locations within the water tank. The excitation unit generates an excitation signal, driving the disturbance source to perform one-dimensional reciprocating motion within the water tank, thus generating a disturbance source flow field. A measuring plate parallel to the tank wall is placed within the water tank, and a pressure sensor array is embedded in the measuring plate. The center point of the pressure sensor array is used as the origin of the coordinate system to define the data acquisition area. The disturbance source is positioned at different locations within the data acquisition area, with defined position intervals and distances from the pressure sensor array. Pressure data at different times is collected at each location using the pressure sensor array.

[0022] The excitation unit includes a modal exciter, a power amplifier, and a signal generator.

[0023] See Figure 1In this embodiment of the invention, a steel sphere with a diameter of 0.03 m is used as the disturbance source. The sphere undergoes one-dimensional reciprocating motion in the vertical direction (i.e., the Z-direction) underwater to generate the disturbance source flow field. Its key flow field characteristics, such as pressure and velocity distribution, are highly similar to those generated by the target object, i.e., the wagging tail fin of an underwater fish. The reciprocating motion is achieved using an SA-JZ040 modal exciter, a power amplifier, and a signal generator. The motion frequency is set to 35 Hz, and the amplitude is 0.004 m. A range of 10 to 12 times the characteristic length (diameter) of the sphere is used as the data acquisition area. The sphere's position is moved using a lead screw and guide rail, and the one-dimensional reciprocating motion of the sphere is achieved through the excitation unit. In this embodiment, an SMP-6009-1 type pressure sensor is used, forming a 3×3 two-dimensional pressure sensor array.

[0024] The geometric center of the measuring plate is used as the origin of the XYZ three-dimensional coordinate system and the center point of the pressure sensor array. The minimum spacing between the pressure sensors is 0.06m. Data acquisition areas of 0.315m × 0.36m × 0.12m are selected in the XYZ directions of three-dimensional space. The SMP-6009-1 pressure sensor has a range of 0~100Kpa, an accuracy of ±0.25%FS, and a frequency response of 10kHz. The pressure sensor's force-bearing surface diameter is 15mm, and its waterproof performance is ensured by O-rings, silicone adhesive sealing, and waterproof materials. When the sphere is arranged at different positions within the data acquisition area, the minimum spacing between positions is 0.03m, i.e., its characteristic length; the minimum spacing between the sphere and the pressure sensor array in the direction perpendicular to the measuring plate is 0.045mm. The position change of the sphere in the XY plane is achieved through a lead screw guide, while the position change in the vertical direction is achieved through the transmission of motion via stainless steel connecting rods of different lengths, with a connecting rod diameter of 2mm. The analog signals measured by the pressure sensor array are converted into digital signals using an Altair Technology USB data acquisition card and efficiently transmitted to the host computer acquisition software for real-time display and data storage. The acquisition range is ±5V, and the external clock sampling frequency is 100kHz. In this embodiment, the lead screw guide has a static load of 20kg, a movable distance range of 1.1m, and is controlled by a servo motor with a position error controlled within 0.5mm.

[0025] To ensure the stability of the sphere's motion and the full development of the flow field during the pressure sensor array measurement process, data from the first 3 seconds of recording are discarded, and only pressure data from the last 7 seconds are recorded. The sampling frequency of the pressure sensors is 10 kHz, and each pressure data point is obtained by changing the initial sampling time while keeping the sampling duration the same. Using time-series pressure data measured by 9 pressure sensors at different locations, 75 pressure data points are collected at each location, forming a total of 48,750 sets of pressure data, denoted as "Raw data".

[0026] (2) Construct a stress dataset.

[0027] At each location of the disturbance source, the pressure data collected by the pressure sensor array are collectively considered as a single pressure data sample; the location of the disturbance source is known; and the pressure data contained in each pressure data sample are used to form a feature matrix. and the characteristic matrix Proper Orthogonal Decomposition (POD) is performed to extract the orthogonal basis of the feature matrix, and the bases are arranged in descending order of energy to select the main spatiotemporal distribution features contained in the pressure data sample, and the corresponding reconstruction matrix is ​​constructed. The location of the disturbance source during the pressure data sample collection is used as the location label, and then compared with the reconstruction matrix. Together, they form a new stress data sample, thus constructing a stress dataset.

[0028] By changing the sampling starting point when collecting pressure data from the pressure sensor array, pressure data samples at different times at each location of the disturbance source can be collected, thereby constructing different pressure datasets. These different pressure datasets can then be used for training and testing of multilayer perceptron neural networks.

[0029] Let the number of pressure sensors in the pressure sensor array be . The number of times pressure data was collected at each location of the disturbance source was: Then the characteristic matrix is , To represent the real number space, the characteristic matrix must first be decentered. ; in, Representation of the characteristic matrix The decentralized matrix represents the portion of the signal that has changed (i.e., the data after removing the mean). The mean vector of each pressure sensor at all times. .

[0030] Subsequently, the covariance matrix is ​​constructed. , can be represented as: ; Among them, superscript This indicates the matrix transpose.

[0031] For covariance matrix Perform eigenvalue decomposition, that is: ; In the formula, For the first There are 10 eigenvalues, arranged in descending order. The larger the value, the greater the contribution of that mode to the signal characteristics; For the first Each singular value represents the energy of a different mode; For the first A time feature vector.

[0032] And the spatial eigenvector is the first... Each spatial mode can be represented as: ; It contains a certain correlation between pressure data from different pressure sensors, and This includes information about the temporal changes of the signal, and the corresponding modal evolution can be represented as: ; in, Indicates in At the moment The time evolution coefficient of the mode, representing the . Each mode in The level of activity at any given moment; Indicates the first Singular values ​​of each mode, Indicates the first Time feature vectors The Each component.

[0033] Feature matrix It can be approximately represented as the former A linear combination of modalities, reconstructing the matrix exist The column vector at time t can be represented as: ; in, For the first Spatial modes.

[0034] In this scheme, the first 1-3 modes are selected, i.e. =1,2,3, thus constructing the corresponding reconstruction matrix. .

[0035] Therefore, by selecting high-energy modes to reconstruct the original sensor array signal, the main features of the disturbance source flow field were extracted, and low-energy noise components in the signal were filtered out, thus improving the signal-to-noise ratio of the input data of the disturbance source localization model.

[0036] Step 2: Construct a multilayer perceptron (MLP) neural network.

[0037] See Figure 2 The smallest unit neuron in an MLP neural network processes weighted input values ​​through a linear combination and activation function before outputting them to the next layer of neurons. Taking the first neuron in the first hidden layer as an example, the expression between its output and input is as follows: ; in, Representing the One input value, ; The dimension of the input value; Representing the The weight values ​​of each input value for the first neuron. The threshold representing the neuron; The activation function, typically a nonlinear function, transforms the linear combination of neurons into a nonlinear form. In this invention, the ReLU function is chosen as the activation function for the hidden layer neurons, and its mathematical expression is: ; Connecting different neurons pairwise in layers creates a neural network, thereby improving its ability to handle complex mapping relationships. Multiple fully connected neural networks with multiple layers constitute a multilayer perceptron neural network; for example... Figure 2 As shown, each sphere represents a neuron, and a vertical row of neurons forms a layer. The leftmost layer is the input layer, where neurons have no thresholds or activation functions; their function is limited to feeding input signals into the network. The rightmost layer is the output layer, and the middle layers are hidden layers, where neurons have thresholds and activation functions, enabling them to process information.

[0038] The input layer dimension of the multilayer perceptron neural network designed in this invention is . The corresponding reconstruction matrix contains pressure data collected by nine pressure sensors at 25 time points starting from different initial sampling times. Five hidden layers are used, with the number of neurons in each hidden layer from input to output being 162, 108, 54, 27, and 9, respectively. The output of the output layer is the predicted value of the coordinates of the disturbance source in three-dimensional space.

[0039] Step 3: Divide the stress dataset into training, testing, and validation sets, and normalize the stress data samples. The training set is responsible for updating and optimizing the network parameters of the multilayer perceptron neural network. The validation set contains stress data samples that do not overlap with the training set and does not participate in the parameter update process. It is responsible for verifying the generalization ability of the multilayer perceptron neural network after each iteration of network parameter updates. By observing the loss function value calculated by the validation set after each iteration, the convergence speed and convergence level of the multilayer perceptron neural network can be judged and evaluated. The testing set is used to test the trained network model. The predicted values ​​of different stress data samples with the same location label (i.e., stress data collected at different time periods at the same location of the disturbance source) are averaged to obtain the average statistical result of the location coordinates, thereby evaluating the positioning accuracy.

[0040] During the training of the multilayer perceptron neural network, the mini-batch stochastic gradient descent algorithm is used to update the network parameter values, i.e., weights and thresholds. The mean squared error (MSE) between the predicted value obtained from the output layer and the true position (label) is used as the loss function. The loss function is calculated in batches with a batch size of 16, and different parameters are updated along the negative gradient direction of the loss function to achieve the global minimization of the loss function, thereby determining the final parameters of the entire neural network model. The number of iterations in the training process is set to 8000 steps.

[0041] Step 4: In practical applications, for disturbance sources with unknown locations, pressure data is acquired using a pressure sensor array, and a corresponding reconstruction matrix is ​​constructed. The reconstructed matrix Pressure data collected by all pressure sensors at different initial sampling times are input into a trained multilayer perceptron neural network to obtain the location prediction result of the disturbance source.

[0042] Example: In one embodiment of the present invention, for each pressure data sample, the pressure dataset obtained by reconstructing the first 3 modes is denoted as "PODMr3"; the pressure dataset obtained by reconstructing the first 2 modes is denoted as "PODMr2"; and the pressure dataset obtained by reconstructing the first mode is denoted as "PODMr1".

[0043] like Figure 4As shown, the power spectral density distribution curves of the original pressure signal, the signal reconstructed from the first three modes, and the signal reconstructed from the first mode, measured by the sensor array, are displayed when the location coordinates of the disturbance source are (0.165, 0.09, 0.03) m. The energy of different signals at the characteristic frequency (the disturbance frequency of the sphere) is basically the same, while the energy at other frequencies (the noise frequencies caused by various interference factors) shows significant differences. The energy of the noise components in the reconstructed signals is effectively suppressed compared to the original signals. Among them, the noise signal reconstructed from the first mode has the lowest energy.

[0044] For the stress dataset, 70% of the data was randomly selected as the training set, 10% as the validation set, and 20% as the test set. The reconstructed matrix and positional label values ​​were normalized by obtaining the maximum and minimum values ​​of the feature or label values ​​included in the samples and mapping them to the numerical range of 0 to 1 to prevent data overflow. During training, five hidden layers were used, with 162, 108, 54, 27, and 9 neurons in each hidden layer from input to output, respectively. A mini-batch stochastic gradient descent algorithm was employed with a batch size of 16, a learning rate of 0.01, and a mean squared error (MSE) loss function. Each set of data was iterated for 8000 steps. After training, the weight parameters between neurons and the threshold parameters for each neuron were determined, resulting in a final multilayer perceptron neural network (MLP) model.

[0045] The localization performance of the trained multilayer perceptron model was tested and analyzed to verify its generalization ability and accuracy. The sample data from the test set was normalized using the minimum and maximum values ​​used in the normalization of the training set sample data, and then input into the model to output the corresponding normalized predicted values. After inverse normalization, the predicted 3D position coordinates of the disturbance source were obtained. The relative error between the predicted coordinates and the true values ​​was calculated, expressed as: ; in, The true values ​​representing the three-dimensional position coordinates and straight-line distance of the disturbance source. Represents the respective predicted values ​​obtained after processing by the MLP model, and the straight-line distance. That is, the distance between the location of the disturbance source and the center of the pressure sensor array, expressed as: ; Relative error of coordinate values , and It can reflect the proportion of positional deviation in each direction relative to the overall detection distance, and the relative error of straight-line distance. It can reflect the deflection angle and deflection distance between the predicted and actual locations of the disturbance source.

[0046] For the test set sample data, pressure sensor array signals collected in four time periods are randomly selected from pressure data samples at the same disturbance source location. The average value of the disturbance source location coordinate prediction values ​​obtained by processing these four sample signals by the multilayer perceptron model is taken as the final prediction value. Taking the test set data obtained after reconstructing the first-order mode as an example, it is denoted as "PODMr1a".

[0047] Figure 5 This study investigates the probability density distribution of the relative error in predicting coordinate values ​​and straight-line distances in each direction after testing the original data, the data obtained after reconstructing the first three modes, the data obtained after reconstructing the first two modes, the data obtained after reconstructing the first mode, and the data after averaging these parameters. Figure 5 (a) to (d) are the probability density distribution curves of the relative errors of the x, y, z coordinates and the straight-line distance r, respectively.

[0048] The probability density distribution of the predicted relative errors of the coordinate values ​​in the three directions approximates a symmetrical Gaussian distribution with a mean of 0. The standard deviation of the error distribution can be obtained by fitting the distribution with a Gaussian function. The probability density distribution of the predicted relative error of the straight-line distance approximates a Karnaugh distribution, with the Karnaugh distribution having the following degrees of freedom (…). This determines the shape of the distribution function, which reflects the mean and standard deviation of the error distribution. The smaller the value, the more concentrated the distribution, and the higher the peak value. For example... Figure 6 The figure shows the relative error correlation fitting parameters obtained after testing the test set data. The results obtained from different processing methods correspond to each row in the figure. The first three columns are the standard deviation parameters of the predicted relative error of the three-dimensional position coordinates of the disturbance source after Gaussian fitting, and the fourth column is the degrees of freedom of the predicted relative error on the straight-line distance after fitting with a chi-square distribution. The parameter is defined as follows: (The value in parentheses represents the percentage reduction of that parameter relative to the original data, except for the raw data.)

[0049] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for signal feature extraction and reconstruction based on POD and disturbance source localization combined with MLP, characterized in that, include: A pressure data acquisition system is established, comprising a frame and a water tank within the frame. A lead screw guide rail is mounted above the frame, driving a slider to move horizontally. An excitation unit is mounted on the slider, and a disturbance source for simulating the disturbance characteristics of the target object is mounted below the excitation unit via a connecting rod. By utilizing the cooperation of the lead screw guide rail and the slider, and adjusting the length of the connecting rod, the disturbance source can be positioned at different locations within the water tank. The excitation unit generates an excitation signal, driving the disturbance source to perform one-dimensional reciprocating motion within the water tank, thereby generating a disturbance source flow field. A measuring plate parallel to the tank wall is placed within the water tank, and a pressure sensor array is embedded in the measuring plate. The center point of the pressure sensor array is used as the origin of the coordinate system to define the data acquisition area. The disturbance source is positioned at different locations within the data acquisition area, with defined position intervals and distances from the pressure sensor array. Pressure data at different times is collected at each location using the pressure sensor array. At each location of the disturbance source, the pressure data collected by the pressure sensor array are collectively considered as a pressure data sample. The pressure data contained in each pressure data sample are used to form a feature matrix, and intrinsic orthogonal mode decomposition is performed on the feature matrix to extract the eigenvalues. These eigenvalues ​​are then arranged in descending order of energy to select the main spatiotemporal distribution features contained in the pressure data sample, and a corresponding reconstruction matrix is ​​constructed. The location of the disturbance source when the pressure data sample is collected is used as the location label, which, together with the reconstruction matrix, forms a new pressure data sample, thus constructing a pressure dataset. Construct a multilayer perceptron neural network; wherein the input layer dimension of the multilayer perceptron neural network is... ,in This represents the number of pressure sensors in the pressure sensor array. This represents the number of pressure data collected by the pressure sensor at different initial sampling times; multiple hidden layers are used, and the output layer outputs the predicted coordinates of the disturbance source in three-dimensional space. After normalizing the pressure data samples in the pressure dataset, the multilayer perceptron neural network is trained and validated, and the trained multilayer perceptron neural network is saved. For a disturbance source with an unknown location, pressure data is acquired using a pressure sensor array and a corresponding reconstruction matrix is ​​constructed. Multiple pressure data collected by all pressure sensors at different initial sampling times under the reconstruction matrix are input into a trained multilayer perceptron neural network to obtain the location prediction result of the disturbance source.

2. The method for signal feature extraction and reconstruction based on POD and disturbance source localization combined with MLP according to claim 1, characterized in that, The excitation unit includes a modal exciter, a power amplifier, and a signal generator.

3. The method for signal feature extraction and reconstruction based on POD and disturbance source localization combined with MLP according to claim 1, characterized in that, By changing the sampling starting point when collecting pressure data from the pressure sensor array, pressure data samples at different times at each location of the disturbance source can be collected, thus constructing different pressure datasets.

4. The signal feature extraction and reconstruction method based on POD and the disturbance source localization method combined with MLP according to claim 1, characterized in that, The process of constructing the reconstruction matrix is ​​as follows: Let the number of pressure sensors in the pressure sensor array be . The number of times pressure data was collected at each location of the disturbance source was: Then the characteristic matrix is First, the feature matrix needs to be decentralized: ； in, Representation of the characteristic matrix The decentralized matrix; The mean vector of each pressure sensor at all times; Subsequently, the covariance matrix is ​​constructed. , is represented as: ； Among them, superscript Indicates matrix transpose; For covariance matrix Perform eigenvalue decomposition: ； In the formula, For the first There are 1 eigenvalues, arranged in descending order; For the first Each singular value represents the energy of a different mode; For the first One time feature vector; Spatial eigenvector, i.e., the first Each spatial mode is represented as: ； Spatial modes encompass a certain correlation between pressure data from different pressure sensors, while This includes information about the temporal changes of the signal, and the corresponding modal evolution can be represented as: ； in, Indicates in At the moment The time evolution coefficient of the mode, representing the . Each mode in The level of activity at any given moment; Indicates the first Singular values ​​of each mode, Indicates the first Time feature vectors The One component; Feature matrix It can be approximately represented as the former A linear combination of modalities, reconstructing the matrix exist The column vector at time t is represented as: ； in, For the first One spatial mode; Define the order of the modes to construct the corresponding reconstruction matrix. .

5. The method for signal feature extraction and reconstruction based on POD and disturbance source localization combined with MLP according to claim 1, characterized in that, The input layer dimension of a multilayer perceptron neural network is The pressure data collected by the nine pressure sensors at different initial sampling times under the corresponding reconstruction matrix are collected at 25 time points. Five hidden layers are used, with the number of neurons in each hidden layer from input to output being 162, 108, 54, 27, and 9, respectively.

6. The signal feature extraction and reconstruction method based on POD and the disturbance source localization method combined with MLP according to claim 1, characterized in that, When evaluating a multilayer perceptron neural network, the predicted values ​​of different pressure data samples at the same location are averaged to obtain the average statistical result of the location coordinates, thereby evaluating the positioning accuracy.

7. The signal feature extraction and reconstruction method based on POD and the disturbance source localization method combined with MLP according to claim 1, characterized in that, During the training of a multilayer perceptron neural network, a mini-batch stochastic gradient descent algorithm is used to update the network parameter values. The mean square error between the predicted values ​​obtained from the output layer and the location labels is used as the loss function. The loss function is calculated in batches and different parameters are updated along the negative gradient direction of the loss function to achieve global minimization of the loss function. Through a certain number of iterative training steps, the final parameters of the entire neural network model can be determined.

8. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the signal feature extraction and reconstruction method based on POD and the disturbance source localization method combined with MLP as described in any one of claims 1-7.

9. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the POD-based signal feature extraction and reconstruction and disturbance source localization method combined with MLP as described in any one of claims 1-7.

Images