CN-121996943-A - POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method

CN121996943ACN 121996943 ACN121996943 ACN 121996943ACN-121996943-A

Abstract

The invention discloses a POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method, which comprises the steps of firstly establishing a pressure data acquisition system, utilizing an excitation unit to match a disturbance source to simulate disturbance features of a target object, utilizing a pressure sensor array to acquire pressure data samples at different periods under the position of the disturbance source, utilizing the pressure data samples to construct a feature matrix, carrying out signal feature extraction based on POD decomposition of the feature matrix, constructing a pressure data set by taking the true position of the reconstruction matrix and the disturbance source as new sample data, adopting an MLP network as a positioning model, utilizing the pressure data set to train and test the MLP network, verifying the accuracy of the MLP network through pressure data samples before and after reconstruction under the same position label, finally obtaining a trained positioning model, and positioning the pressure data acquired by an unknown disturbance source. The invention can effectively improve the detection precision of the position of the disturbance source of the underwater flow field.

Inventors

CHEN XIAOPENG
BAO YI
REN FENG
CHEN YUPENG
HUANG QINGHUA
DU PENG
HU HAIBAO
Xie Shaozhang

Assignees

西北工业大学

Dates

Publication Date: 20260508
Application Date: 20251222

Claims (9)

1. The POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method is characterized by comprising the following steps: The method comprises the steps of establishing a pressure data acquisition system, wherein the pressure data acquisition system comprises a frame and a water tank arranged in the frame, a screw guide rail capable of driving a sliding block to move in a horizontal plane is arranged above the frame, an excitation unit is assembled on the sliding block, disturbance sources for simulating disturbance characteristics of a target object are arranged below the excitation unit through connecting rods, the disturbance sources can be positioned at different positions in the water tank by utilizing the matching of the screw guide rail and the sliding block and the adjustment of the length of the connecting rods, excitation signals are generated by the excitation unit to drive the disturbance sources to do one-dimensional reciprocating motion in the water tank so as to generate a disturbance source flow field, a measuring plate parallel to the wall surface of the water tank is arranged in the water tank, a pressure sensor array is embedded in the measuring plate, a central point of the pressure sensor array is used as an origin of a coordinate system, the data acquisition area is set, the disturbance sources are positioned at different positions in the data acquisition area, the distance between the set positions and the pressure sensor array is used for acquiring pressure data at different moments at each position; when the disturbance source is positioned at each position, the pressure data collected by the pressure sensor array are taken as a pressure data sample together, the pressure data contained in each pressure data sample are formed into a characteristic matrix, the characteristic matrix is subjected to intrinsic orthogonal modal decomposition, the characteristic values of the characteristic matrix are extracted, and the characteristic values are arranged according to the energy from large to small so as to select main time-space distribution characteristics contained in the pressure data sample and construct a corresponding reconstruction matrix; constructing a multi-layer perceptron neural network, wherein the dimension of the input layer of the multi-layer perceptron neural network is as follows Wherein For the number of pressure sensors in the array of pressure sensors, Adopting a plurality of hidden layers, wherein the output result of the output layer is a predicted value of the coordinate of the disturbance source in the three-dimensional space; after normalizing pressure data samples in the pressure data set, training and verifying the multi-layer perceptron neural network, and storing the trained multi-layer perceptron neural network; And for the disturbance source with unknown position, acquiring pressure data by using a pressure sensor array, constructing a corresponding reconstruction matrix, and inputting a plurality of pressure data acquired by all pressure sensors under the reconstruction matrix at different initial sampling moments into a trained multi-layer perceptron neural network to obtain a position prediction result of the disturbance source.
2. The POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method of claim 1, wherein said excitation unit comprises a modal exciter, a power amplifier and a signal generator.
3. The POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method according to claim 1, wherein pressure data samples of different time periods under each position of the disturbance source are acquired by changing a sampling start point when the pressure sensor array performs pressure data acquisition, and different pressure data sets are constructed.
4. The method for extracting and reconstructing signal features based on POD and combining with MLP for positioning disturbance sources according to claim 1, wherein the construction process of the reconstruction matrix is as follows: the number of pressure sensors in the pressure sensor array is recorded as The number of times the pressure data acquisition is performed when the disturbance source is at each position is The feature matrix is Firstly, carrying out decentration treatment on the feature matrix: ; Wherein, the Representing feature matrices A matrix after decentralization; A mean value vector for each pressure sensor at all times; subsequently, a covariance matrix is constructed Expressed as: ; Wherein, superscript Representing a matrix transpose; for covariance matrix And (3) performing eigenvalue decomposition: ; in the formula, Is the first The characteristic values are arranged from large to small; Is the first Singular values representing energy of different modes; Is the first A temporal feature vector; Spatial feature vectors, i.e. first The spatial modes are expressed as: ; the spatial mode comprises a certain correlation existing between pressure data of different pressure sensors The time-varying information of the signal is included and the corresponding modal evolution can be expressed as: ; Wherein, the Is shown in At the moment of time The time evolution coefficient of each mode represents the first The individual modes are The degree of activity at the moment; Represent the first The singular values of the individual modes, Represent the first Time feature vector Is the first of (2) A component; feature matrix Can be approximately represented as a front Linear combination of individual modes, reconstructing matrix At the position of The column vector at time instant is expressed as: ; Wherein, the Is the first A spatial modality; setting the order of the mode, thereby constructing a corresponding reconstruction matrix 。
5. The method for locating disturbance sources based on POD signal feature extraction and reconstruction and MLP combination according to claim 1, wherein the input layer dimension of the multi-layer perceptron neural network is The number of neurons on each hidden layer in the direction from input to output is 162, 108, 54, 27 and 9 respectively by using 5 hidden layers.
6. The method for locating disturbance sources based on POD signal feature extraction and reconstruction and MLP combination according to claim 1, wherein when the multi-layer perceptron neural network is evaluated, the predicted values of different pressure data samples of the same position label are subjected to average processing to obtain an average value statistical result of position coordinates, so that the evaluation of locating precision is performed.
7. The POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method according to claim 1, wherein in the multi-layer perceptron neural network training process, a small batch of random gradient descent algorithm is adopted to update network parameter values, the mean square error of a predicted value obtained by an output layer and a position label is used as a loss function, the loss function is calculated in batches, different parameters are updated along the negative gradient direction of the loss function, so that the global minimization of the loss function is achieved, and the final parameters of the whole neural network model can be determined through iterative training of a certain number of steps.
8. A terminal device comprising a processor, a memory and a computer program stored in the memory, characterized in that the processor, when executing the computer program, implements the POD-based signal feature extraction and reconstruction and the disturbance source localization method in combination with MLP according to any one of claims 1-7.
9. A computer readable storage medium, in which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the POD-based signal feature extraction and reconstruction and MLP-combined disturbance source localization method according to any one of claims 1 to 7.

Description

POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method Technical Field The invention belongs to the field of underwater flow field feature identification and target detection, and particularly relates to a POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method. Background With the exploration and development of the ocean, the development of the underwater object detection technology plays a main role. Sensitive organs such as side lines of aquatic vertebrates such as fish and beard of seal can sense hydrodynamic wake generated by movement of other objects, and behavior such as hunting, obstacle avoidance and clustering can be completed by identifying the wake. According to the inspired, the non-acoustic underwater target positioning technology is gradually developed, a neural network model is trained through a large amount of data by combining a machine learning method, so that the information such as the size, the shape and the position of a target object can be predicted through hydrodynamic characteristics generated by the target object, and the method becomes a new research trend. The current flow field characteristic identification and target detection technology utilizing hydrodynamic information has the following defects: (1) The low signal-to-noise ratio characteristic of the signal has become a main influencing factor for inhibiting the underwater flow field characteristic identification and the target detection accuracy improvement based on the machine learning method. (2) The existing signal preprocessing method is mainly a traditional filtering method, and has high dependence on priori information such as characteristic frequency ranges of signals. (3) In the detection process, different input signals under a single-position sample have certain noise interference, and random errors exist in the neural network model prediction process, so that the accuracy and the reliability of a detection result are required to be improved. Disclosure of Invention The invention aims to provide a POD-based signal feature extraction and reconstruction and MLP-combined disturbance source positioning method so as to effectively improve the detection precision of the position of an underwater flow field disturbance source. In order to realize the tasks, the invention adopts the following technical scheme: a disturbance source positioning method based on POD signal feature extraction and reconstruction and combined with MLP comprises the following steps: The method comprises the steps of establishing a pressure data acquisition system, wherein the pressure data acquisition system comprises a frame and a water tank arranged in the frame, a screw guide rail capable of driving a sliding block to move in a horizontal plane is arranged above the frame, an excitation unit is assembled on the sliding block, disturbance sources for simulating disturbance characteristics of a target object are arranged below the excitation unit through connecting rods, the disturbance sources can be positioned at different positions in the water tank by utilizing the matching of the screw guide rail and the sliding block and the adjustment of the length of the connecting rods, excitation signals are generated by the excitation unit to drive the disturbance sources to do one-dimensional reciprocating motion in the water tank so as to generate a disturbance source flow field, a measuring plate parallel to the wall surface of the water tank is arranged in the water tank, a pressure sensor array is embedded in the measuring plate, a central point of the pressure sensor array is used as an origin of a coordinate system, the data acquisition area is set, the disturbance sources are positioned at different positions in the data acquisition area, the distance between the set positions and the pressure sensor array is used for acquiring pressure data at different moments at each position; when the disturbance source is positioned at each position, the pressure data collected by the pressure sensor array are taken as a pressure data sample together, the pressure data contained in each pressure data sample are formed into a characteristic matrix, the characteristic matrix is subjected to intrinsic orthogonal modal decomposition, the characteristic values of the characteristic matrix are extracted, and the characteristic values are arranged according to the energy from large to small so as to select main time-space distribution characteristics contained in the pressure data sample and construct a corresponding reconstruction matrix; constructing a multi-layer perceptron neural network, wherein the dimension of the input layer of the multi-layer perceptron neural network is as follows WhereinFor the number of pressure sensors in the array of pressure sensors,Adopting a plurality of hidden layers, wherein the output result of the output layer is a