Multi-wave detection and imaging system for fish school

Abstract

The invention discloses a multi-wave detection and imaging system for fish schools, which relates to the field of detection and imaging of fish schools and comprises a metamaterial acoustic emission module, a self-adaptive signal receiving module, a three-dimensional point cloud generation module, an image enhancement and generation module and an intelligent decision control module. According to the multi-wave detection and imaging system for the fish school, clustering parameters can be dynamically adjusted based on the real-time fish school density, the neighborhood radius is automatically expanded in a high-density area to avoid segmentation errors, and the core point judgment threshold is reduced in a low-density area to reduce missing detection; an acoustic attenuation model is constructed in combination with water body environment parameters, emission parameters are calibrated in real time, images can be enhanced, space and time dimension features are fused, the image resolution and definition are improved, accurate recognition of multiple fish species is achieved, the fish species can be better distinguished, meanwhile, a generative network is adopted to directly convert acoustic data into visual images, and the recognition accuracy is improved. Therefore, the generated image can reflect the actual distribution condition of the fish school more truly and accurately.

Description

Technical Field

[0001] The present invention relates to the field of fish school detection and imaging, and in particular to a multi-wave fish school detection and imaging system. Background Art

[0002] Fish are aquatic animals with skulls and gills but no limbs with toes.

[0003] Traditional fish monitoring systems rely on the fusion of multiple types of sensors, such as acoustics and optics. The hardware cost is high and the deployment is complex. Optical equipment is prone to failure, especially when the water is turbid. Traditional sonar has a short detection distance and low resolution in turbid water, which cannot meet the needs of deep-sea aquaculture scenarios. At the same time, the parameters of traditional clustering algorithms are fixed, and they are prone to misjudgment when faced with changes in fish density. Dense fish schools may be segmented or sparse fish schools may be missed.

[0004] Therefore, it is necessary to propose a multi-wave detection and imaging system for fish schools to solve the above problems. Summary of the Invention

[0005] The main purpose of the present invention is to provide a multi-wave detection and imaging system for fish schools, which can effectively solve the problems in the background technology.

[0006] To achieve the above object, the technical solution adopted by the present invention is: A multi-wave detection and imaging system for fish schools, comprising a metamaterial acoustic emission module, an adaptive signal receiving module, a three-dimensional point cloud generation module, an image enhancement and generation module, and an intelligent decision-making control module. The metamaterial acoustic emission module is used to emit adaptively adjustable acoustic wave signals into the water body. The adaptive signal receiving module is used to efficiently receive sonar echo signals, suppress environmental interference and enhance fish school characteristics; The three-dimensional point cloud generation module is used to convert the sonar echo signal into fish school point cloud data in three-dimensional space, providing a basis for imaging; The image enhancement and generation module is used to convert the point cloud data generated by the three-dimensional point cloud generation module into a high-resolution visual image to improve the identification of fish distribution; The intelligent decision-making control module realizes fish species identification, quantity estimation and automatic control of breeding equipment based on image analysis.

[0007] Preferably, the metamaterial acoustic emission module includes a metamaterial composite transducer array, an environmental parameter sensing unit, and an environmental parameter data preprocessing unit; The metamaterial composite transducer array is used to transmit acoustic wave signals and achieve long-distance detection and energy focusing; The environmental parameter sensing unit includes a turbidity sensor, a temperature sensor and a salinity sensor, and the data acquisition frequency of the environmental parameter sensing unit is ten times per second; The environmental parameter data preprocessing unit is used to perform outlier filtering and data normalization processing on the data collected by the environmental parameter sensing unit.

[0008] Preferably, an environment-acoustic attenuation model is established based on the data processed by the environmental parameter data preprocessing unit, and the formula is: ; in The water sound wave attenuation coefficient, expressed in dB / m, is used to directly determine the energy in sound wave propagation; Turbidity NTU, which indicates the concentration of suspended particles in water. Every increase of 1 NTU will increase the attenuation by 1.5 dB / m. is salinity PSU, which reflects the electrolyte content of water. Every increase of 1 PSU leads to an increase of 0.8 dB / m in attenuation. is the temperature °C. The increase in temperature will reduce the viscosity of water. The attenuation will decrease by 0.3 for every 1 °C increase in temperature. when When the metamaterial composite transducer array uses the 150-200kHz low frequency band to balance the detection distance and energy consumption; when When , the metamaterial composite transducer array switches to the 200-250kHz mid-frequency band to balance penetration and resolution; when , metamaterial composite transducer array enables high frequency band of 250-300kHz and energy focusing; Based on The power compensation of the metamaterial composite transducer array is calculated to ensure that the acoustic wave energy density at the target depth is maintained above 10μW / cm². The formula is: ; in is the power compensation of the metamaterial composite transducer array; is the basic power, which is 200W here, and 0.1 is the power adjustment coefficient based on experience.

[0009] Preferably, the adaptive signal receiving module specifically includes the following steps: The water-borne sonar echo signal is received via a sonar transducer and amplified with low noise using a superconducting quantum interference device preamplifier with a liquid helium sealed chamber, with a gain controlled at 20-40dB. The signal is sampled using a 16-bit ADC with a sampling rate of 1MHz. After conversion to a digital signal, the signal is subjected to a short-time Fourier transform using an FPGA preprocessing unit. This generates a time-frequency spectrum with a 512-point time window and a 75% overlap rate, which is used to extract the time-frequency distribution characteristics of the echo. The digital signal energy mean and variance are simultaneously calculated to provide a preliminary assessment of the echo intensity and noise level. The preprocessed sonar echo signal is divided into 512-point frames according to the time series, expanded to 128-channel input, and formed into an input tensor of dimension (512, 128); Extract convolutional features and introduce attention mechanism; Phase conjugate weight generation: Then pass it through the fully connected layer to calculate the 128×128 phase conjugate weight matrix , generate conjugate matrices through complex operations , for the input signal matrix Perform element-wise dot product: ; The phase inversion of the echo signal is achieved to compensate for the phase distortion caused by multipath propagation. The signal after phase conjugation processing is then output after denormalization to preliminarily suppress the regular interference of water surface and bottom reflections. Multipath interference classification and mask generation: The processed signal is subjected to short-time Fourier transform again to generate a time-frequency spectrum with a resolution of 512×256, which is input into a pre-trained CNN classifier. The CNN classifies each pixel of the time-frequency spectrum into the time-frequency region of fish, surface reflection, bottom reflection, and noise, and generates a three-dimensional mask matrix of time × frequency × category. The background noise mean is calculated based on the classification results. and standard deviation , calculate the adaptive threshold: ; Dynamically adjusted by the estimated fish density, ranging from 1.5 to 3.0; Interference signal suppression is performed by applying a threshold mask to suppress signal energy in the time-frequency regions classified as surface and bottom reflections; and applying wavelet threshold denoising to the noise regions. Real-time estimation of noise power spectrum: The processed signal is input into the noise feature extraction network, and the noise power spectrum of the current frame is output. Sum signal power ; Frequency domain enhancement: Perform fast Fourier transform on the signal and calculate the dynamic Wiener filter coefficient according to the formula. This is used to focus on enhancing the energy of the fish characteristic frequency band of 100-200kHz and suppress the noise frequency band. The formula is: ; Time domain enhancement: The frequency-domain enhanced signal is converted back to the time domain via inverse fast Fourier transform and input into a bidirectional LSTM network to capture the temporal correlation of fish echoes and further suppress burst noise. The final output is a pure echo signal with a signal-to-noise ratio improved by 35-40dB, retaining the phase and amplitude information of the fish echo.

[0010] Preferably, the signal quality after processing according to claim 4 is analyzed: if the fish school echo energy is lower than the self-set threshold, the transmission power is increased by 10-20% and the frequency is adjusted by ±5kHz, and the signal is fed back to the metamaterial acoustic transmission module for subsequent transmission of the acoustic wave signal.

[0011] Preferably, the three-dimensional point cloud generation module performs band-pass filtering on the echo signal output by the adaptive signal receiving module, retaining only the effective frequency band of 50-300kHz, and eliminating environmental low-frequency noise and high-frequency electronic interference; based on the phase conjugation result of the adaptive signal receiving module, the multi-channel echo signal is calibrated for phase consistency to eliminate the phase error caused by the installation deviation of the transducer array.

[0012] Preferably, the three-dimensional point cloud generation module includes a time-frequency feature processing submodule, a point cloud generation submodule, a point cloud optimization submodule, and a motion modeling submodule. The time-frequency feature processing submodule performs pixel-level classification on the time-frequency spectrum using a deep learning model, identifies four types of features: fish echo, water surface reflection, bottom reflection, and environmental noise, generates a probability mask map, and automatically filters out non-fish echo points. The time-frequency feature processing submodule is electrically connected to the point cloud generation submodule. The point cloud generation submodule is used to calculate the three-dimensional coordinates of the valid echo points after the probability mask screening through the beamforming algorithm, wherein the distance is calculated based on the echo delay and the water sound speed, and needs to be corrected in real time in combination with the temperature, salinity and turbidity data; the azimuth and pitch angles are determined by the beam pointing of the phased array transducer; based on this, the echo energy intensity of each point is synchronously recorded to form the initial point cloud data; The time-frequency feature processing submodule is electrically connected to the point cloud optimization submodule, and the point cloud optimization submodule is used to calculate the local density of each point, automatically identify and eliminate outliers with a density significantly lower than that of the neighborhood, and retain the real fish echo points; Adaptive DBSCAN clustering algorithm is used to dynamically adjust according to the real-time point cloud density: the neighborhood radius is automatically reduced in high-density fish areas to avoid the accidental merging of multiple fish schools; the neighborhood radius is expanded in low-density areas to ensure that scattered fish schools are not segmented; finally, the point cloud is divided into multiple fish clusters, each cluster corresponds to the spatial distribution of a group of fish. The adaptive DBSCAN clustering algorithm formula includes the dynamic neighborhood radius formula and the dynamic minimum point number formula. The dynamic neighborhood radius formula is: ; in is the neighborhood radius of the current cluster, which represents the distance threshold for judging whether two points are neighbors in the point cloud space; is the basic neighborhood radius, which is pre-set by the sensor resolution and water environment; The real-time fish density estimation is calculated by point cloud density using the formula: ; in To detect the volume of water; is the number of valid fish echo points in the point cloud; The conversion coefficient is determined by calibration test, specifically: place a 10*10*5m pool Fish of known size, collect point clouds and count points at the same time , calculate the conversion coefficient according to the formula: ; When the fish density increases, the neighborhood radius is expanded to avoid high-density fish schools being divided into multiple clusters; when the fish density decreases, the neighborhood radius is reduced to distinguish scattered fish schools; The dynamic minimum point formula is: ; in The minimum number of neighborhood points for determining a core point; 5 is the minimum number of points to ensure that noise is not mistakenly identified as a core point in low-density scenes; is a floor function to ensure is an integer; In low-density scenes, the core point determination threshold is lowered to avoid missing scattered fish schools; in high-density scenes, the core point requirement is increased to avoid mistaken merging of multiple fish schools. The motion modeling submodule is used to perform spatial registration on continuous frame point clouds, and find the optimal transformation relationship between frames through an iterative nearest point algorithm to achieve alignment of point clouds in a unified coordinate system, laying the foundation for motion analysis; the three-dimensional point cloud is projected onto a two-dimensional plane, and the optical flow algorithm is used to calculate the point cloud motion vector between adjacent frames, and the horizontal and vertical velocity components of each point are obtained. The three-dimensional velocity is calculated in combination with the distance change rate; the process noise parameters of the Kalman filter are dynamically adjusted based on the motion state of the fish school, specifically: the noise covariance of fast-swimming fish schools is automatically increased to improve the tracking response speed; the noise covariance of stationary or slowly moving fish schools is reduced to improve the trajectory smoothness; based on this, a smooth and continuous fish school motion trajectory is generated.

[0013] Preferably, the point cloud in the metamaterial composite transducer array coordinate system is converted to the world coordinate system, and the point clouds of different observation positions are unified based on GPS and attitude sensor data; radial basis function interpolation is used to complete the sparse point cloud area to ensure the complete outline of the fish school; and the moving least squares method is applied to smooth the surface to reduce noise interference while retaining the geometric features.

[0014] Preferably, the image enhancement and generation module includes a point cloud projection imaging submodule, an image enhancement submodule, an image generation and completion submodule, and an image post-processing and feature extraction submodule; The point cloud projection imaging submodule is used to complete the projection transformation of the three-dimensional point cloud into a two-dimensional image, generate an initial fish school distribution image, and output a preliminary two-dimensional fish school image; The image enhancement submodule is used to improve image contrast, compensate for illumination, remove noise and shadows, enhance the display of fish school details, and output an enhanced high-quality two-dimensional fish school image; The image generation and completion submodule generates new fish school images based on the generative adversarial network and completes and repairs the images with missing information to output complete and realistic fish school images; The image post-processing and feature extraction submodule is used to sharpen image edges, extract image features and perform dimensionality reduction processing, and is used to output an image with clear edges and a feature vector after dimensionality reduction.

[0015] Preferably, the intelligent decision control module includes a data fusion submodule, a decision model submodule, an optimization and evaluation submodule, and an execution control submodule; The data fusion submodule is used to extract features, stitch and weightedly fuse the above-mentioned point cloud, image and environmental data, providing comprehensive and high-quality data input for subsequent decision-making models; The decision model submodule generates a preliminary decision plan based on the fused data by constructing and training an appropriate rule-based and machine learning decision model. The machine learning algorithm library includes random forest, LSTM time series prediction, and YOLOv5 target detection. The optimization and evaluation submodule is used to perform multi-objective optimization on the preliminary decision plan, evaluate the decision risk, and generate the final optimized and low-risk decision plan; The execution control submodule is used to convert the decision plan into control instructions and send them to the corresponding execution device, and receive feedback on the execution effect in real time to make decision adjustments and re-execution.

[0016] Compared with the prior art, the present invention provides a multi-wave detection and imaging system for fish schools, which has the following beneficial effects: This multi-wave detection and imaging system for fish schools adopts a pure acoustic detection architecture. Through the optimization of metamaterial composite transducers and deep learning algorithms, it eliminates optical modules, effectively reducing hardware costs. It can dynamically adjust the sound wave emission parameters and adaptively adjust the frequency and energy focusing according to the turbidity of the water body, significantly extending the detection distance in high-turbidity water bodies. Combined with deep learning beamforming technology, it can accurately control the beam angle and width, improve azimuth resolution, and identify smaller fish individuals.

[0017] The fish school multi-wave detection and imaging system dynamically adjusts clustering parameters based on real-time fish school density, automatically expands the neighborhood radius in high-density areas to avoid segmentation errors, and lowers the core point determination threshold in low-density areas to reduce missed detections. It builds an acoustic attenuation model based on water environment parameters and calibrates the emission parameters in real time, which can effectively solve the error problem of traditional empirical formulas, suppress the interference of water surface and bottom reflections, and enhance the fish school echo signal. It can also enhance the image, integrate spatial and temporal dimension features, improve image resolution and clarity, achieve accurate identification of multiple fish species, and better distinguish fish species.

[0018] The fish school multi-wave detection and imaging system uses an end-to-end generation network to directly convert acoustic data into visual images, eliminating multiple intermediate processing links in traditional acoustic imaging. It not only shortens the image generation time and reduces processing delays, but also reduces information loss during data transmission and processing, making the generated images more realistic and accurate in reflecting the actual distribution of fish schools. It can also optimize images from the spatial and temporal dimensions. In the spatial dimension, it improves the clarity and resolution of the image by enhancing the details and edge information of the image; in the temporal dimension, it uses the correlation between adjacent frames to smooth the image, reduce noise interference, and make the movement trajectory of the fish school more coherent and clear. Based on this, it can provide more intuitive and easier-to-analyze monitoring results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0021] like Figure 1 As shown, a multi-wave detection and imaging system for fish schools includes a metamaterial acoustic emission module, an adaptive signal receiving module, a three-dimensional point cloud generation module, an image enhancement and generation module, and an intelligent decision-making control module. The metamaterial acoustic emission module is used to emit adaptively adjustable acoustic wave signals into the water body. The metamaterial acoustic emission module includes a metamaterial composite transducer array, an environmental parameter sensing unit, and an environmental parameter data preprocessing unit. The metamaterial composite transducer array is used to transmit acoustic signals and achieve long-distance detection and energy focusing. The metamaterial composite transducer array includes a 256-element ring array with a diameter of 30 cm and a silicon-based metamaterial lattice structure. The metamaterial lattice has a characteristic size of 50 μm and is used to achieve the phonon quantum tunneling effect. Its operating frequency range is 50-300 kHz; the transmission power is adaptively adjusted from 0 to 1000 W; and the beam width is dynamically controllable from 1.5° to 15°. The environmental parameter sensing unit includes a turbidity sensor, a temperature sensor, and a salinity sensor. The data acquisition frequency of the environmental parameter sensing unit is ten times per second; The environmental parameter data preprocessing unit is used to perform outlier filtering and data normalization processing on the data collected by the environmental parameter sensing unit; The environment-acoustic attenuation model is established based on the data processed by the environmental parameter data preprocessing unit. The formula is: ; in The water sound wave attenuation coefficient, expressed in dB / m, is used to directly determine the energy in sound wave propagation; Turbidity NTU, which indicates the concentration of suspended particles in water. Every increase of 1 NTU will increase the attenuation by 1.5 dB / m. is salinity PSU, which reflects the electrolyte content of water. Every increase of 1 PSU leads to an increase of 0.8 dB / m in attenuation. is the temperature °C. The increase in temperature will reduce the viscosity of water. The attenuation will decrease by 0.3 for every 1 °C increase in temperature. when When the metamaterial composite transducer array uses the 150-200kHz low frequency band to balance the detection distance and energy consumption; when When , the metamaterial composite transducer array switches to the 200-250kHz mid-frequency band to balance penetration and resolution; when , metamaterial composite transducer array enables high frequency band of 250-300kHz and energy focusing; Based on The power compensation of the metamaterial composite transducer array is calculated to ensure that the acoustic wave energy density at the target depth is maintained above 10μW / cm². The formula is: ; in is the power compensation of the metamaterial composite transducer array; is the basic power, which is 200W here, and 0.1 is the power adjustment coefficient based on experience.

[0022] The adaptive signal receiving module is used to efficiently receive sonar echo signals, suppress environmental interference, and enhance fish school characteristics. The adaptive signal receiving module specifically includes the following steps: The water-borne sonar echo signal is received via a sonar transducer and amplified with low noise using a superconducting quantum interference device preamplifier with a liquid helium sealed chamber, with a gain controlled at 20-40dB. The signal is sampled using a 16-bit ADC with a sampling rate of 1MHz. After conversion to a digital signal, the signal is subjected to a short-time Fourier transform using an FPGA preprocessing unit. This generates a time-frequency spectrum with a 512-point time window and a 75% overlap rate, which is used to extract the time-frequency distribution characteristics of the echo. The digital signal energy mean and variance are simultaneously calculated to provide a preliminary assessment of the echo intensity and noise level. The preprocessed sonar echo signal is divided into 512-point frames according to the time series, expanded to 128-channel input, and formed into an input tensor of dimension (512, 128); Convolutional features are extracted and an attention mechanism is introduced, including the CBAM module. The weight of the echo energy is first calculated in the channel dimension, and then a position attention map is generated in the spatial dimension to achieve directional enhancement of fish echoes. Phase conjugate weight generation: Then pass it through the fully connected layer to calculate the 128×128 phase conjugate weight matrix , generate conjugate matrices through complex operations , for the input signal matrix Perform element-wise dot product: ; The phase inversion of the echo signal is achieved to compensate for the phase distortion caused by multipath propagation. The signal after phase conjugation processing is then output after denormalization to preliminarily suppress the regular interference of water surface and bottom reflections. Multipath interference classification and mask generation: The processed signal is subjected to short-time Fourier transform again to generate a time-frequency spectrum with a resolution of 512×256, which is input into a pre-trained CNN classifier. The CNN classifies each pixel of the time-frequency spectrum into the time-frequency region of fish, surface reflection, bottom reflection, and noise, and generates a three-dimensional mask matrix of time × frequency × category. The background noise mean is calculated based on the classification results. and standard deviation , calculate the adaptive threshold: ; Dynamically adjusted by the estimated fish density, ranging from 1.5 to 3.0; Interference signal suppression execution: For time-frequency regions classified as surface and bottom reflections, a threshold mask is applied to suppress signal energy; for noisy regions, wavelet threshold denoising is used to preserve the time-frequency characteristics of fish school echoes, and the suppression mask is updated in real time to adapt to changes in the water environment; Real-time estimation of noise power spectrum: The processed signal is input into the noise feature extraction network, and the noise power spectrum of the current frame is output. Sum signal power ; Frequency domain enhancement: Perform fast Fourier transform on the signal and calculate the dynamic Wiener filter coefficient according to the formula. This is used to focus on enhancing the energy of the fish characteristic frequency band of 100-200kHz and suppress the noise frequency band. The formula is: ; Time Domain Enhancement: The frequency-domain enhanced signal is converted back to the time domain via inverse fast Fourier transform and input into a bidirectional LSTM network to capture the temporal correlation of fish echoes, such as the periodic characteristics of fish tail wagging, to further suppress sudden noise. The final output is a pure echo signal with a signal-to-noise ratio improved by 35-40dB, preserving the phase and amplitude information of the fish echo, providing high-quality data for subsequent point cloud generation and signal quality analysis: If the fish echo energy is lower than the self-set threshold, the transmission power is increased by 10-20% and the frequency is adjusted by ±5kHz, and the data is fed back to the metamaterial acoustic transmission module for subsequent sound wave signal transmission.

[0023] The 3D point cloud generation module is used to convert the sonar echo signal into fish point cloud data in three-dimensional space, providing a basis for imaging. The 3D point cloud generation module performs bandpass filtering on the echo signal output by the adaptive signal receiving module, retaining only the 50-300kHz effective frequency band, and eliminating environmental low-frequency noise and high-frequency electronic interference; based on the phase conjugation result of the adaptive signal receiving module, the multi-channel echo signal is calibrated for phase consistency to eliminate the phase error caused by the installation deviation of the transducer array. The 3D point cloud generation module includes a time-frequency feature processing submodule, a point cloud generation submodule, a point cloud optimization submodule, and a motion modeling submodule. The time-frequency feature processing submodule uses a deep learning model to perform pixel-level classification on the time-frequency spectrum, identify four types of features: fish echo, water surface reflection, bottom reflection, and environmental noise, generate a probability mask map, and automatically filter out non-fish echo points; The time-frequency feature processing submodule is electrically connected to the point cloud generation submodule. The point cloud generation submodule is used to calculate the three-dimensional coordinates of the valid echo points after the probability mask is screened using a beamforming algorithm. The distance is calculated based on the echo delay and the water sound velocity, and needs to be corrected in real time with the temperature, salinity and turbidity data. The azimuth and elevation angles are determined by the beam pointing of the phased array transducer. Based on this, the echo energy intensity of each point is synchronously recorded to form the initial point cloud data. The time-frequency feature processing submodule is electrically connected to the point cloud optimization submodule. The point cloud optimization submodule is used to calculate the local density of each point, automatically identify and eliminate outliers with a density significantly lower than that of the neighborhood, such as noise interference points and isolated scattering points, and retain the real fish echo points; Adaptive DBSCAN clustering algorithm is used to dynamically adjust according to the real-time point cloud density: the neighborhood radius is automatically reduced in high-density fish areas to avoid the accidental merging of multiple fish schools; the neighborhood radius is expanded in low-density areas to ensure that scattered fish schools are not segmented; finally, the point cloud is divided into multiple fish clusters, each cluster corresponds to the spatial distribution of a group of fish. The adaptive DBSCAN clustering algorithm formula includes the dynamic neighborhood radius formula and the dynamic minimum point number formula. The dynamic neighborhood radius formula is: ; in is the neighborhood radius of the current cluster, which represents the distance threshold for judging whether two points are neighbors in the point cloud space; is the basic neighborhood radius, which is pre-set by the sensor resolution and water environment; The real-time fish density estimation is calculated by point cloud density using the formula: ; in To detect the volume of water; is the number of valid fish echo points in the point cloud; The conversion coefficient is determined by calibration test, specifically: place a 10*10*5m pool Fish of known size, collect point clouds and count points at the same time , calculate the conversion coefficient according to the formula: ; When the fish density increases, the neighborhood radius is expanded to avoid high-density fish schools being divided into multiple clusters; when the fish density decreases, the neighborhood radius is reduced to distinguish scattered fish schools; The dynamic minimum point formula is: ; in The minimum number of neighborhood points for determining a core point; 5 is the minimum number of points to ensure that noise is not mistakenly identified as a core point in low-density scenes; is a floor function to ensure is an integer; In low-density scenes, the core point determination threshold is lowered to avoid missing scattered fish schools; in high-density scenes, the core point requirement is increased to avoid mistaken merging of multiple fish schools. The motion modeling submodule is used to perform spatial registration on point clouds of consecutive frames. It uses an iterative nearest point algorithm to find the optimal transformation relationship between frames, which is used to achieve alignment of point clouds in a unified coordinate system, laying the foundation for motion analysis. The three-dimensional point cloud is projected onto a two-dimensional plane, and the optical flow algorithm is used to calculate the point cloud motion vector between adjacent frames. The horizontal and vertical velocity components of each point are obtained, and the three-dimensional velocity is calculated based on the distance change rate. The process noise parameters of the Kalman filter are dynamically adjusted based on the motion state of the fish school. Specifically, the noise covariance of fast-swimming fish schools is automatically increased to improve the tracking response speed; the noise covariance of stationary or slowly moving fish schools is reduced to improve the trajectory smoothness. Based on this, a smooth and continuous fish school motion trajectory is generated, and the point cloud in the metamaterial composite transducer array coordinate system is converted to the world coordinate system. At the same time, based on GPS and attitude sensor data, the attitude sensor data includes installation position, pitch angle, and roll angle, to achieve point cloud unification at different observation positions. Radial basis function interpolation is used to complete the sparse point cloud area to ensure the completeness of the fish school outline. At the same time, the moving least squares method is applied for surface smoothing to reduce noise interference while preserving geometric features.

[0024] The image enhancement and generation module is used to convert the point cloud data generated by the 3D point cloud generation module into a high-resolution visual image to improve the identification of fish distribution. The image enhancement and generation module includes a point cloud projection imaging submodule, an image enhancement submodule, an image generation and completion submodule, and an image post-processing and feature extraction submodule. The point cloud projection imaging submodule is used to complete the projection transformation of the three-dimensional point cloud to a two-dimensional image, generate an initial fish distribution image, and output a preliminary two-dimensional fish image; The image enhancement submodule is used to improve image contrast, compensate for illumination, remove noise and shadows, enhance the display of fish school details, and output enhanced high-quality two-dimensional fish school images; The image generation and completion submodule generates new fish school images based on a generative adversarial network and completes and repairs images with missing information to output complete and realistic fish school images. The image post-processing and feature extraction submodule is used to sharpen image edges, extract image features and perform dimensionality reduction processing, and output an image with clear edges and a feature vector after dimensionality reduction.

[0025] The intelligent decision-making and control module realizes fish species identification, quantity estimation and automatic control of aquaculture equipment based on image analysis. The intelligent decision-making and control module includes a data fusion submodule, a decision model submodule, an optimization and evaluation submodule, and an execution control submodule; The data fusion submodule is used to extract features, stitch and weightedly fuse the above-mentioned point cloud, image and environmental data, providing comprehensive and high-quality data input for subsequent decision-making models; The decision model submodule generates preliminary decision plans based on the integrated data by constructing and training appropriate rule-based and machine learning decision models; The optimization and evaluation submodule is used to perform multi-objective optimization on the preliminary decision plan, evaluate the decision risk, and generate the final optimized and low-risk decision plan; The execution control submodule is used to convert decision plans into control instructions and send them to the corresponding execution devices, and receive feedback on the execution effect in real time to make decision adjustments and re-execution.

[0026] Example 1: A multi-wave detection and imaging system for fish schools, which uses an environmental parameter sensing unit to collect temperature, salinity, and turbidity in real time. The environmental parameter data preprocessing unit preprocesses the collected data and calculates the water body sound wave attenuation coefficient. The formula is: ; The water body acoustic wave attenuation coefficient is output to the metamaterial composite transducer array for adaptive adjustment according to the water body acoustic wave attenuation coefficient. The power compensation amount is calculated to enable the metamaterial composite transducer array to transmit acoustic signals. The formula is obtained by fitting the least squares method using 200 sets of different water samples with acoustic attenuation experiments at temperatures of 5-30°C, salinity of 0-35PSU, and turbidity of 0-100NTU. 、 and , coefficient of determination R²=0.96; Example 2: A multi-wave fish detection and imaging system receives water-borne sonar echo signals via a sonar transducer and performs low-noise amplification on a superconducting quantum interference device preamplifier with a liquid helium sealed chamber, with a gain controlled at 20-40 dB. The signal is sampled by a 16-bit ADC at a sampling rate of 1 MHz, converted to a digital signal, and then subjected to a short-time Fourier transform based on an FPGA preprocessing unit. This generates a time-frequency spectrum with a 512-point time window and a 75% overlap rate, which is used to extract the time-frequency distribution characteristics of the echo. The digital signal energy mean and variance are simultaneously calculated to preliminarily determine the echo intensity and noise level. The preprocessed sonar echo signal is divided into 512-point frames according to the time series, expanded to 128-channel input, and formed into an input tensor of dimension (512, 128); Convolutional features are extracted and an attention mechanism is introduced, including the CBAM module. The weight of the echo energy is first calculated in the channel dimension, and then a position attention map is generated in the spatial dimension to achieve directional enhancement of fish echoes. Phase conjugate weight generation: Then pass it through the fully connected layer to calculate the 128×128 phase conjugate weight matrix , where the fully connected layer input is a flattened vector of a 512×128 tensor and the output is a 128×128 complex matrix The real and imaginary parts of the conjugate matrix are generated by complex operations , for the input signal matrix Perform element-wise dot product: ; Using a multipath interference dataset of 10,000 samples, the weights are optimized through backpropagation with the goal of minimizing the signal distortion compensation error. The phase inversion of the echo signal is achieved to compensate for the phase distortion caused by multipath propagation. The signal after phase conjugation processing is then output after denormalization to preliminarily suppress the regular interference of water surface and bottom reflections. Multipath interference classification and mask generation: The processed signal is short-time Fourier transformed again to generate a time-frequency spectrum with a resolution of 512×256, which is input into the pre-trained CNN classifier. The network structure of the CNN classifier is: four-layer convolution, fully connected layer and Softmax classifier; training data: 20,000 time-frequency spectrum graphs are annotated, including manual segmentation labels of fish schools, water surface reflections, bottom reflections, and noise areas; CNN classifies each pixel of the time-frequency spectrum graph into the time-frequency areas of fish schools, water surface reflections, bottom reflections and noise, and generates a three-dimensional mask matrix of time × frequency × category. Three-dimensional mask generation: For the probability map of fish schools, water surface reflections, bottom reflections and noise areas output by CNN, pixels with probability >0.8 are taken to generate binary masks, which are stacked into a three-dimensional matrix along the time, frequency and category dimensions, and the background noise mean is calculated according to the classification results. and standard deviation , calculate the adaptive threshold: ; Dynamically adjusted by the estimated fish density, ranging from 1.5 to 3.0; Interference signal suppression execution: For time-frequency regions classified as surface and bottom reflections, a threshold mask is applied to suppress signal energy; for noisy regions, wavelet threshold denoising is used to preserve the time-frequency characteristics of fish school echoes, and the suppression mask is updated in real time to adapt to changes in the water environment; Real-time estimation of noise power spectrum: The processed signal is input into the noise feature extraction network, and the noise power spectrum of the current frame is output. Sum signal power ; Frequency domain enhancement: Perform fast Fourier transform on the signal and calculate the dynamic Wiener filter coefficient according to the formula. This is used to focus on enhancing the energy of the fish characteristic frequency band of 100-200kHz and suppress the noise frequency band. The formula is: ; Time Domain Enhancement: The frequency-domain enhanced signal is converted back to the time domain via inverse fast Fourier transform and input into a bidirectional LSTM network to capture the temporal correlation of fish echoes, such as the periodic characteristics of fish tail wagging, to further suppress sudden noise. The final output is a pure echo signal with a signal-to-noise ratio improved by 35-40dB, preserving the phase and amplitude information of the fish echo, providing high-quality data for subsequent point cloud generation. The following is a verification table: Example 3: A multi-wave detection and imaging system for fish schools, analyzing the quality of passing signals: if the fish school echo energy is lower than a self-set threshold, the transmission power is increased by 10-20% and the frequency is adjusted by ±5kHz, and the feedback is fed back to the metamaterial acoustic transmission module for subsequent transmission of acoustic wave signals, wherein the self-set threshold is set based on 90% of the historical echo energy average; The echo signal output by the adaptive signal receiving module is band-pass filtered, retaining only the 50-300kHz effective frequency band to eliminate environmental low-frequency noise and high-frequency electronic interference. Based on the phase conjugation results of the adaptive signal receiving module, the multi-channel echo signal is calibrated for phase consistency to eliminate phase errors caused by transducer array installation deviations. By using a deep learning model to perform pixel-level classification on the time-frequency spectrum, four types of features, namely fish echoes, surface reflections, bottom reflections, and ambient noise, are identified. A probability mask is generated to automatically filter out non-fish echo points. For valid echo points filtered by the probability mask, the three-dimensional coordinates are calculated using a beamforming algorithm. The distance is calculated based on the echo delay and water sound velocity, and needs to be corrected in real time with temperature, salinity and turbidity data. The azimuth and elevation angles are determined by the beam pointing of the phased array transducer. Based on this, the echo energy intensity of each point is synchronously recorded to form the initial point cloud data. Calculate the local density of each point, automatically identify and remove outliers with significantly lower density than the neighborhood, such as noise interference points and isolated scattering points, and retain the real fish echo points; Adaptive DBSCAN clustering algorithm is used to dynamically adjust according to the real-time point cloud density: the neighborhood radius is automatically reduced in high-density fish areas to avoid the accidental merging of multiple fish schools; the neighborhood radius is expanded in low-density areas to ensure that scattered fish schools are not segmented; finally, the point cloud is divided into multiple fish clusters, each cluster corresponds to the spatial distribution of a group of fish. The adaptive DBSCAN clustering algorithm formula includes the dynamic neighborhood radius formula and the dynamic minimum point number formula. The dynamic neighborhood radius formula is: ; in is the neighborhood radius of the current cluster, which represents the distance threshold for judging whether two points are neighbors in the point cloud space; is the basic neighborhood radius, which is pre-set by the sensor resolution and water environment; The real-time fish density estimation is calculated by point cloud density using the formula: ; in To detect the volume of water; is the number of valid fish echo points in the point cloud; The conversion coefficient is determined by calibration test, specifically: place a 10*10*5m pool Fish of known size, collect point clouds and count points at the same time , calculate the conversion coefficient according to the formula: ; When the fish density increases, the neighborhood radius is expanded to avoid high-density fish schools being divided into multiple clusters; when the fish density decreases, the neighborhood radius is reduced to distinguish scattered fish schools; The dynamic minimum point formula is: ; in The minimum number of neighborhood points for determining a core point; 5 is the minimum number of points to ensure that noise is not mistakenly identified as a core point in low-density scenes; is a floor function to ensure is an integer; In low-density scenes, the core point determination threshold is lowered to avoid missing scattered fish schools; in high-density scenes, the core point requirement is increased to avoid mistaken merging of multiple fish schools. The point clouds of consecutive frames are spatially registered, and the optimal transformation relationship between frames is found through the iterative nearest point algorithm to achieve alignment of the point clouds in a unified coordinate system, laying the foundation for motion analysis; the three-dimensional point cloud is projected onto a two-dimensional plane, and the optical flow algorithm is used to calculate the point cloud motion vector between adjacent frames, obtain the horizontal and vertical velocity components of each point, and calculate the three-dimensional velocity based on the distance change rate; based on the movement state of the fish school, such as velocity variance, the process noise parameters of the Kalman filter are dynamically adjusted, specifically: fast-swimming fish automatically increase the noise covariance to improve the tracking response speed; stationary or slowly moving fish reduce the noise covariance to improve the trajectory smoothness; based on this, a smooth and continuous fish movement trajectory is generated.

[0027] The point cloud in the metamaterial composite transducer array coordinate system is converted to the world coordinate system. Based on GPS and attitude sensor data, including installation position, pitch angle, and roll angle, point clouds from different observation positions are unified. Radial basis function interpolation is used to complete sparse point cloud areas to ensure the complete outline of the fish school. Moving least squares is also used for surface smoothing to reduce noise interference while preserving geometric features. Embodiment 4: A multi-wave detection and imaging system for fish schools, wherein the point cloud projection imaging submodule is used to complete the projection transformation of the three-dimensional point cloud into a two-dimensional image, generate an initial fish school distribution image, and output a preliminary two-dimensional fish school image. Specifically, according to the observation requirements, the horizontal, i.e., XY plane, or the vertical, i.e., XZ, YZ plane, is selected for the projection plane, and the three-dimensional point cloud coordinates are converted to the two-dimensional projection plane using the homogeneous transformation matrix. According to the point cloud density, grayscale values ​​and color values ​​are assigned to each projection point, so that high-density, high-energy fish school areas appear bright or vivid in the image, and vice versa. The mapped pixel values ​​are filled into the two-dimensional image matrix to generate the initial fish school distribution image, and Gaussian filtering is applied for smoothing to remove isolated noise points.

[0028] The image enhancement submodule is used to enhance image contrast, compensate for illumination, remove noise and shadows, and enhance the display of fish details. It then outputs an enhanced, high-quality, two-dimensional fish image. Specifically, the module calculates the grayscale value distribution of the image, calculates the cumulative distribution function, and maps the original image grayscale values ​​to a new distribution range, expanding the grayscale dynamic range and enhancing contrast. Adaptive histogram equalization (CLAHE) is used to divide the image into multiple sub-blocks, and histogram equalization is performed on each sub-block to preserve local contrast. Ambient light intensity data and image grayscale statistics are used to estimate the illumination distribution in the image. For areas with uneven illumination, a region growing-based method is used for illumination compensation. For shadowed areas, a trained shadow detection model is used to identify shadow boundaries. Pixel information from surrounding non-shadow areas is then used for interpolation and filling to restore the fish details in the shadows.

[0029] The image generation and completion submodule uses a GAN to generate new fish images and completes any missing information in the images, ultimately outputting complete, realistic fish images. Specifically, image generation is based on a generative adversarial network (GAN), which consists of a generator and a discriminator. The generator receives random noise or low-dimensional feature vectors to generate images, while the discriminator distinguishes between generated images and real images. The two train against each other, making the generator's images increasingly realistic. The GAN is trained on 100,000 labeled fish images. The generator employs a five-layer deconvolutional architecture, while the discriminator uses a PatchGAN. The loss function is a Wasserstein distance with a gradient penalty. The customized implementation of fish school image generation combines random noise with fish school feature vectors extracted from point cloud data, such as fish school shape and density distribution, as conditional inputs to the generator, guiding the generation of images that conform to actual fish school conditions. Deconvolution layers are used to gradually expand the size of feature maps to generate high-resolution images. After each deconvolution layer, skip connections are introduced to integrate low-level detail features, making the generated images more realistic. In addition to distinguishing whether an image is real or fake, a discriminant branch for fish school characteristics is added to ensure that the generated images are not only visually realistic but also conform to the biological characteristics and distribution patterns of fish schools. For missing areas in the image, such as information loss caused by occlusion and noise, an image completion algorithm based on the attention mechanism is proposed. First, the global features of the image are extracted using a convolutional neural network, and then the attention weights of the missing area and the surrounding area are calculated. The formula is: ; according to , extract information from the surrounding area and fill and repair the missing area. Represents the missing area pixels Pixels in the surrounding area The attention weight of the surrounding area pixels For the missing pixels The greater the weight, the more important The information of the restored pixel The more critical it is, the more the pixel information will be used when filling in the missing area later. :in and They are and The corresponding eigenvector; is the feature similarity function, which is cosine similarity; is an exponential function; Represents the set of pixels in the surrounding area; is an element in the set; The pixels in the missing area are used for calculation and all pixels in the surrounding area The sum of the similarities after exponential operation is used to normalize the numerator to ensure the calculated attention weight The weights are between 0 and 1, and the sum of all weights is 1. The weights obtained in this way have probabilistic meaning and can reasonably distribute the contribution ratio of the pixels in the surrounding area to the pixels in the missing area.

[0030] The image post-processing and feature extraction submodule is used to sharpen image edges, extract image features and perform dimensionality reduction processing, and output images with clear edges and feature vectors after dimensionality reduction. Specifically, the Laplace operator or Gaussian difference filter is applied to sharpen the image, highlighting the edge contours of the fish school, making the boundaries of the fish school in the image clearer, facilitating subsequent identification and analysis. The key points and descriptors in the image are extracted using one of the scale-invariant feature transform and accelerated robust feature algorithm for tasks such as image matching and target recognition. For high-dimensional feature vectors, principal component analysis is used for dimensionality reduction to retain the main feature information while reducing the amount of data.

[0031] Example 5: A multi-wave detection and imaging system for fish schools, which extracts features, splices and weightedly fuses the above-mentioned point clouds, images and environmental data to provide comprehensive and high-quality data input for subsequent decision-making models. The data specifically includes the three-dimensional fish school point cloud coordinates, energy intensity, and motion vectors obtained from the three-dimensional point cloud generation module, as well as the fish school images and image feature vectors output by the image enhancement and generation module; access to environmental parameters collected by water sensors such as temperature, salinity and turbidity sensors, dissolved oxygen sensors, weather stations and other equipment, including water temperature, salinity, dissolved oxygen content, light intensity, wind speed and direction; and calling historical fish school monitoring data stored in the database. , breeding operation records, fishery production and other information, where the coefficient allocation of weighted fusion is carried out according to data characteristics. Due to different advantages of different types of data in describing fish and environmental conditions, weights can be preliminarily set according to data characteristics. For example, point cloud data has high accuracy in three-dimensional spatial positioning and plays a key role in judging the position of fish schools, so it can be given a higher weight; image data performs well in identifying the form and color characteristics of fish schools, and plays an outstanding role in judging the health status of fish schools, so it should also be given a higher weight; and environmental monitoring data such as water temperature and dissolved oxygen content are crucial for analyzing the impact of the environment on the behavior of fish schools, and they also need to be reasonably weighted.

[0032] The decision model submodule constructs and trains appropriate rule-based and machine learning decision models, generates preliminary decision plans based on the fused data, and builds a fish status assessment indicator system based on the above content, covering fish density, distribution uniformity, movement activity, and growth and health indicators. Using the fused data, based on pre-established mathematical models such as statistical density estimation models and machine learning-based growth assessment models, the current state of the fish school is quantitatively scored, providing a clear understanding of the real-time status of the fish school. Develop rule-based decision-making strategies for clear, simple fishery production scenarios. For example, if the dissolved oxygen level falls below a set threshold, immediately activate the aerator; if the water temperature exceeds the appropriate range, trigger heating or cooling equipment control instructions. Rules are formulated based on fishery farming standards, expert experience, and common patterns in historical data. For complex and ever-changing scenarios that are difficult to describe with simple rules, we use machine learning algorithms to build decision-making models. For example, for feeding decisions, we collect a large amount of historical feeding data as training samples and use regression and reinforcement learning models. During model training, we continuously adjust parameters to optimize decision-making accuracy, enabling intelligent prediction of the optimal feeding amount and timing based on the current state of the fish and environmental conditions.

[0033] The optimization and evaluation submodule is used to perform multi-objective optimization on the initial decision plan, assess decision risks, and generate the final optimized and low-risk decision plan. Fishery production often needs to take into account multiple objectives, such as increasing fish production, reducing breeding costs, and protecting the water environment. A multi-objective optimization algorithm is used to assign weights to different objectives, converting the multi-objective problem into a single-objective optimization problem. For example, when pursuing high yields, the weight of the yield target can be appropriately increased; when focusing on environmental protection, the weight of environmental indicators can be increased, thereby obtaining the overall optimal decision plan. A risk assessment mechanism is introduced to account for uncertainties in decision-making, such as the impact of equipment failure and sudden severe weather on fishery production. Utilizing historical data and probabilistic statistical methods, a quantitative risk assessment is conducted for each decision, calculating the probability of risk occurrence and the potential losses. For high-risk decisions, corresponding response plans are developed, such as activating backup equipment and emergency fishing plans, to ensure the stable operation of fishery production.

[0034] The execution control submodule converts decision plans into control instructions and sends them to the corresponding execution devices. It also receives real-time feedback on the execution results, adjusts and re-executes the decision, and converts the optimized decision results into specific control instructions, which are then sent to the corresponding execution devices via the communication interface. The instruction format complies with the device communication protocol specifications to ensure accurate device recognition and execution. For example, the instructions sent to the intelligent feeding machine include the feeding amount, feeding time, and feeding area coordinates.

[0035] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-wave fish detection and imaging system, comprising a metamaterial acoustic emission module, an adaptive signal receiving module, a three-dimensional point cloud generation module, an image enhancement and generation module, and an intelligent decision-making control module, characterized by: The metamaterial acoustic emission module is used to emit adaptively adjustable acoustic wave signals to the water body; The adaptive signal receiving module is used to efficiently receive sonar echo signals, suppress environmental interference and enhance fish school characteristics; The three-dimensional point cloud generation module is used to convert the sonar echo signal into fish school point cloud data in three-dimensional space, providing a basis for imaging; The image enhancement and generation module is used to convert the point cloud data generated by the three-dimensional point cloud generation module into a high-resolution visual image to improve the identification of fish distribution; The intelligent decision-making control module realizes fish species identification, quantity estimation and automatic control of breeding equipment based on image analysis.

2. The multi-wave detection and imaging system for fish schools according to claim 1, characterized in that: The metamaterial acoustic emission module includes a metamaterial composite transducer array, an environmental parameter sensing unit, and an environmental parameter data preprocessing unit; The metamaterial composite transducer array is used to transmit acoustic wave signals and achieve long-distance detection and energy focusing; The environmental parameter sensing unit includes a turbidity sensor, a temperature sensor and a salinity sensor, and the data acquisition frequency of the environmental parameter sensing unit is ten times per second; The environmental parameter data preprocessing unit is used to perform outlier filtering and data normalization processing on the data collected by the environmental parameter sensing unit.

3. The multi-wave detection and imaging system for fish schools according to claim 2, characterized in that: An environment-acoustic attenuation model is established based on the data processed by the environmental parameter data preprocessing unit, and the formula is: ； in The water sound wave attenuation coefficient, expressed in dB / m, is used to directly determine the energy in sound wave propagation; Turbidity NTU, which indicates the concentration of suspended particles in water. Every increase of 1 NTU will increase the attenuation by 1.5 dB / m. is salinity PSU, which reflects the electrolyte content of water. Every increase of 1 PSU leads to an increase of 0.8 dB / m in attenuation. is the temperature °C. The increase in temperature will reduce the viscosity of water. The attenuation will decrease by 0.3 for every 1 °C increase in temperature. when When the metamaterial composite transducer array uses the 150-200kHz low frequency band to balance the detection distance and energy consumption; when When , the metamaterial composite transducer array switches to the 200-250kHz mid-frequency band to balance penetration and resolution; when , metamaterial composite transducer array enables high frequency band of 250-300kHz and energy focusing; Based on The power compensation of the metamaterial composite transducer array is calculated to ensure that the acoustic wave energy density at the target depth is maintained above 10μW / cm². The formula is: ； in is the power compensation of the metamaterial composite transducer array; is the basic power, which is 200W here, and 0.1 is the power adjustment coefficient based on experience.

4. The multi-wave detection and imaging system for fish schools according to claim 1, characterized in that: The adaptive signal receiving module specifically includes the following steps: The water-borne sonar echo signal is received via a sonar transducer and amplified with low noise using a superconducting quantum interference device preamplifier with a liquid helium sealed chamber, with a gain controlled at 20-40dB. The signal is sampled using a 16-bit ADC with a sampling rate of 1MHz. After conversion to a digital signal, the signal is subjected to a short-time Fourier transform using an FPGA preprocessing unit. This generates a time-frequency spectrum with a 512-point time window and a 75% overlap rate, which is used to extract the time-frequency distribution characteristics of the echo. The digital signal energy mean and variance are simultaneously calculated to provide a preliminary assessment of the echo intensity and noise level. The preprocessed sonar echo signal is divided into 512-point frames according to the time series, expanded to 128-channel input, and formed into an input tensor of dimension (512, 128); Extract convolutional features and introduce attention mechanism; Phase conjugate weight generation: Then pass it through the fully connected layer to calculate the 128×128 phase conjugate weight matrix , generate conjugate matrices through complex operations , for the input signal matrix Perform element-wise dot product: ； The phase inversion of the echo signal is achieved to compensate for the phase distortion caused by multipath propagation. The signal after phase conjugation processing is then output after denormalization to preliminarily suppress the regular interference of water surface and bottom reflections. Multipath interference classification and mask generation: The processed signal is subjected to short-time Fourier transform again to generate a time-frequency spectrum with a resolution of 512×256, which is input into a pre-trained CNN classifier. The CNN classifies each pixel of the time-frequency spectrum into the time-frequency region of fish, surface reflection, bottom reflection, and noise, and generates a three-dimensional mask matrix of time × frequency × category. The background noise mean is calculated based on the classification results. and standard deviation , calculate the adaptive threshold: ； Dynamically adjusted by the estimated fish density, ranging from 1.5 to 3.0; Interference signal suppression is performed by applying a threshold mask to suppress signal energy in the time-frequency regions classified as surface and bottom reflections; and applying wavelet threshold denoising to the noise regions. Real-time estimation of noise power spectrum: The processed signal is input into the noise feature extraction network, and the noise power spectrum of the current frame is output. Sum signal power ; Frequency domain enhancement: Perform fast Fourier transform on the signal and calculate the dynamic Wiener filter coefficient according to the formula. This is used to focus on enhancing the energy of the fish characteristic frequency band of 100-200kHz and suppress the noise frequency band. The formula is: ； Time domain enhancement: The frequency-domain enhanced signal is converted back to the time domain via inverse fast Fourier transform and input into a bidirectional LSTM network to capture the temporal correlation of fish echoes and further suppress burst noise. The final output is a pure echo signal with a signal-to-noise ratio improved by 35-40dB, retaining the phase and amplitude information of the fish echo.

5. The multi-wave detection and imaging system for fish schools according to claim 4, characterized in that: The signal quality after processing by the adaptive signal receiving module is analyzed: if the fish school echo energy is lower than the self-set threshold, the transmission power is increased by 10-20% and the frequency is adjusted by ±5kHz, and the signal is fed back to the metamaterial acoustic transmission module for subsequent transmission of the acoustic wave signal.

6. The multi-wave detection and imaging system for fish schools according to claim 1, characterized in that: The three-dimensional point cloud generation module performs bandpass filtering on the echo signal output by the adaptive signal receiving module, retaining only the 50-300kHz effective frequency band and eliminating environmental low-frequency noise and high-frequency electronic interference; based on the phase conjugation result of the adaptive signal receiving module, the multi-channel echo signal is calibrated for phase consistency to eliminate the phase error caused by the transducer array installation deviation.

7. The multi-wave detection and imaging system for fish schools according to claim 6, characterized in that: The three-dimensional point cloud generation module includes a time-frequency feature processing submodule, a point cloud generation submodule, a point cloud optimization submodule, and a motion modeling submodule. The time-frequency feature processing submodule uses a deep learning model to perform pixel-level classification on the time-frequency spectrum, identifies four types of features: fish echo, water surface reflection, bottom reflection, and environmental noise, generates a probability mask map, and automatically filters out non-fish echo points. The time-frequency feature processing submodule is electrically connected to the point cloud generation submodule. The point cloud generation submodule is used to calculate the three-dimensional coordinates of the valid echo points after the probability mask screening through the beamforming algorithm, wherein the distance is calculated based on the echo delay and the water sound speed, and needs to be corrected in real time in combination with the temperature, salinity and turbidity data; the azimuth and pitch angles are determined by the beam pointing of the phased array transducer; based on this, the echo energy intensity of each point is synchronously recorded to form the initial point cloud data; The time-frequency feature processing submodule is electrically connected to the point cloud optimization submodule, and the point cloud optimization submodule is used to calculate the local density of each point, automatically identify and eliminate outliers with a density significantly lower than that of the neighborhood, and retain the real fish echo points; Adaptive DBSCAN clustering algorithm is used to dynamically adjust according to the real-time point cloud density: the neighborhood radius is automatically reduced in high-density fish areas to avoid the accidental merging of multiple fish schools; the neighborhood radius is expanded in low-density areas to ensure that scattered fish schools are not segmented; finally, the point cloud is divided into multiple fish clusters, each cluster corresponds to the spatial distribution of a group of fish. The adaptive DBSCAN clustering algorithm formula includes the dynamic neighborhood radius formula and the dynamic minimum point number formula. The dynamic neighborhood radius formula is: ； in is the neighborhood radius of the current cluster, which represents the distance threshold for judging whether two points are neighbors in the point cloud space; is the basic neighborhood radius, which is pre-set by the sensor resolution and water environment; The real-time fish density estimation is calculated by point cloud density using the formula: ； in To detect the volume of water; is the number of valid fish echo points in the point cloud; The conversion coefficient is determined by calibration test, specifically: place a 10*10*5m pool Fish of known size, collect point clouds and count points at the same time , calculate the conversion coefficient according to the formula: When the fish density increases, the neighborhood radius is expanded to avoid high-density fish schools being divided into multiple clusters; when the fish density decreases, the neighborhood radius is reduced to distinguish scattered fish schools; The dynamic minimum point formula is: ； in The minimum neighborhood point threshold for determining the core point; 5 is the minimum number of basic points to ensure that noise is not mistaken for core points in low-density scenes; is a floor function to ensure is an integer; In low-density scenes, the core point determination threshold is lowered to avoid missing scattered fish schools; in high-density scenes, the core point requirement is increased to avoid mistaken merging of multiple fish schools. The motion modeling submodule is used to perform spatial registration on continuous frame point clouds, and find the optimal transformation relationship between frames through an iterative nearest point algorithm to achieve alignment of point clouds in a unified coordinate system, laying the foundation for motion analysis; the three-dimensional point cloud is projected onto a two-dimensional plane, and the optical flow algorithm is used to calculate the point cloud motion vector between adjacent frames, and the horizontal and vertical velocity components of each point are obtained. The three-dimensional velocity is calculated in combination with the distance change rate; the process noise parameters of the Kalman filter are dynamically adjusted based on the motion state of the fish school, specifically: the noise covariance of fast-swimming fish schools is automatically increased to improve the tracking response speed; the noise covariance of stationary or slowly moving fish schools is reduced to improve the trajectory smoothness; based on this, a smooth and continuous fish school motion trajectory is generated.

8. The multi-wave detection and imaging system for fish schools according to claim 7, characterized in that: Converting the point cloud in the metamaterial composite transducer array coordinate system to the world coordinate system, and unifying the point clouds at different observation locations based on GPS and attitude sensor data; Radial basis function interpolation is used to complete the sparse point cloud area to ensure the complete outline of the fish school; at the same time, the moving least squares method is applied to smooth the surface to reduce noise interference while retaining the geometric features.

9. The multi-wave detection and imaging system for fish schools according to claim 1, characterized in that: The image enhancement and generation module includes a point cloud projection imaging submodule, an image enhancement submodule, an image generation and completion submodule, and an image post-processing and feature extraction submodule; The point cloud projection imaging submodule is used to complete the projection transformation of the three-dimensional point cloud into a two-dimensional image, generate an initial fish school distribution image, and output a preliminary two-dimensional fish school image; The image enhancement submodule is used to improve image contrast, compensate for illumination, remove noise and shadows, enhance the display of fish school details, and output an enhanced high-quality two-dimensional fish school image; The image generation and completion submodule generates new fish school images based on the generative adversarial network and completes and repairs the images with missing information to output complete and realistic fish school images; The image post-processing and feature extraction submodule is used to sharpen image edges, extract image features and perform dimensionality reduction processing, and is used to output an image with clear edges and a feature vector after dimensionality reduction.

10. The multi-wave detection and imaging system for fish schools according to claim 1, characterized in that: The intelligent decision control module includes a data fusion submodule, a decision model submodule, an optimization and evaluation submodule, and an execution control submodule; The data fusion submodule is used to extract features, stitch and weightedly fuse the above-mentioned point cloud, image and environmental data, providing comprehensive and high-quality data input for subsequent decision-making models; The decision model submodule generates a preliminary decision plan based on the fused data by constructing and training an appropriate rule-based and machine learning decision model. The machine learning algorithm library includes random forest, LSTM time series prediction, and YOLOv5 target detection. The optimization and evaluation submodule is used to perform multi-objective optimization on the preliminary decision plan, evaluate the decision risk, and generate the final optimized and low-risk decision plan; The execution control submodule is used to convert the decision plan into control instructions and send them to the corresponding execution device, and receive feedback on the execution effect in real time to make decision adjustments and re-execution.

Images