Net breaking leakage detection system for mariculture net cage

Abstract

The invention relates to the technical field of mariculture detection, in particular to a net breakage leakage detection system for a mariculture net cage, and the system comprises a plurality of fixed underwater acoustic device units which are used for obtaining passive acoustic signals; the underwater detection robot is used for transmitting high-frequency sonar pulses and collecting reflection echo signals; the control unit is used for sending a driving instruction to the underwater detection robot and receiving a reflection echo signal and a passive acoustic signal; the signal processing unit is used for extracting time delay and frequency spectrum characteristics from the reflection echo signal and determining position information of a damaged noise source according to the passive acoustic signal; the analysis module is used for extracting a joint acoustic feature vector from the spectrum features through a convolutional neural network, generating a node feature vector representing acoustic association between the net cages through a graph neural network, fusing the joint acoustic feature vector and the node feature vector and then carrying out damage detection; and the position optimization unit is used for carrying out iterative optimization on the position information of the damage noise source and outputting a final net cage damage position.

Description

Technical Field

[0001] The present invention relates to the technical field of marine aquaculture detection, and particularly to a system for detecting broken nets and leaking water in a seawater aquaculture cage. Background Art

[0002] China is rich in fishery resources, with the total output of aquatic products and the output value of aquaculture exceeding one trillion yuan. Deep-sea aquaculture equipment and intelligent fishery are the key points of the current fishery development. Among them, seawater aquaculture includes port ponds, ponds, net enclosures, water tanks, raft-type and cage aquaculture, etc. Cage aquaculture is an important part of deep-sea aquaculture, and underwater cage aquaculture has gradually become the main development direction of China's seawater aquaculture industry.

[0003] The existing cage aquaculture includes two types: ordinary cage aquaculture and deep-sea gravity cage aquaculture. An ordinary cage is a farming equipment that uses a net structure suspended by a cage. Its construction cost is relatively low compared to deep-sea gravity cages, and its maintenance is relatively simple and flexible. It is often deployed in combination with deep-sea gravity cages and is mainly used for hatching fry. However, its structural strength is relatively low, and there is a risk of escape. A deep-sea gravity cage is a farming equipment that uses a rigid structural framework and a net structure for enclosure. It has the characteristics of high structural strength and high output, but its construction cost and technical requirements are relatively high.

[0004] According to statistics, about two-thirds of the losses in cage aquaculture are caused by damaged netting. The damage of the netting on the cage not only causes unpredictable economic losses but also causes serious ecological environment problems. At present, in order to reduce the losses caused by damaged netting, the method of regularly changing the net or having divers regularly check the cage is often used to check the integrity of the netting. However, this method not only has low work efficiency but also has great potential safety hazards. Therefore, a more reliable and safer cage detection solution is needed to solve this key problem that affects the healthy development of underwater cage aquaculture. Summary of the Invention

[0005] In view of the above-mentioned prior art, the present invention aims to provide a system for detecting broken nets and leaking water in a seawater aquaculture cage, mainly to solve the technical problems existing in the above background art.

[0006] To achieve the above object, the technical solution of the embodiment of the present invention is realized as follows: A system for detecting broken nets and leaking water in a seawater aquaculture cage, the system comprising: A plurality of fixed hydrophone units for acquiring passive acoustic signals, the passive acoustic signals including environmental noise signals and turbulent noise signals from different sources; An underwater detection robot equipped with an active sonar for transmitting high-frequency sonar pulses and collecting echo signals; A control unit for sending driving instructions to the underwater detection robot and for receiving the echo signals and passive acoustic signals; A signal processing unit, configured to extract time delay and spectral features from the reflected echo signal and determine the position information of the damage noise source according to the passive acoustic signal; An analysis module, configured with a convolutional neural network and a graph neural network, where the convolutional neural network is used to extract a joint acoustic feature vector from the spectral features, and the graph neural network is used to generate a node feature vector representing the acoustic correlation between the cages, and the two are fused for damage detection; A position optimization unit, configured to iteratively optimize the position information of the damage noise source and output the final damaged position of the cage.

[0007] Optionally, extracting the time delay and spectral features from the reflected echo signal specifically includes: Preprocessing the reflected echo signal, and the preprocessing includes band-pass filtering and noise reduction processing; Performing cross-correlation analysis on the transmitted signal and the preprocessed reflected echo signal to obtain a time delay value; Performing short-time Fourier transform on the preprocessed reflected echo signal to generate a time-frequency distribution matrix, and extracting the spectral centroid and frequency band energy ratio from the time-frequency distribution matrix as spectral features.

[0008] Optionally, determining the position information of the damage noise source according to the passive acoustic signal specifically includes: M hydrophone units acquire M passive acoustic signals, perform band-stop filtering on the M passive acoustic signals to suppress the ship noise frequency band and the biological noise frequency band, and obtain a target acoustic signal including a specific frequency band; Judge whether there is damage noise in the target acoustic signal. If so, perform cross-correlation processing on multiple passive acoustic signals to obtain the time difference of arrival, establish a TDOA positioning equation set based on multiple time differences of arrival, and use the least squares method to solve the TDOA positioning equation set to obtain the position coordinates of the damage noise source.

[0009] Optionally, when the position coordinates of the damage noise source are obtained, select the cage closest to the position coordinates of the damage noise source, and the control unit of the underwater detection robot issues a driving instruction to control the underwater detection robot to move to the cage to emit high-frequency sonar pulses and collect reflected echo signals.

[0010] Optionally, judging whether there is damage noise in the target acoustic signal specifically includes: calculating the signal power spectral density in the target acoustic signal, if the signal power spectral density exceeds the threshold , it means that there is damage noise in the target acoustic signal.

[0011] Optionally, perform cross-correlation processing on multiple passive acoustic signals to obtain time differences of arrival (TDOAs), and establish a TDOA positioning equation set based on the multiple TDOAs. Specifically, it includes: taking the hydrophone unit that first receives the target acoustic signal as the reference hydrophone unit, performing cross-correlation processing on the target acoustic signal received by the reference hydrophone unit and the target acoustic signals received by the other M - 1 hydrophone units respectively, obtaining the TDOAs of the target acoustic signal of the damaged noise source between the other M - 1 hydrophone units and the reference hydrophone unit, and forming a set of positioning information groups based on the multiple TDOAs, with a total of M - 1 sets of positioning information groups formed.

[0012] Optionally, extract a joint acoustic feature vector from the spectral features. Specifically, it includes: establishing and training a CNN model, inputting the time-frequency distribution matrix into the pre-trained CNN model to obtain CNN high-order features, and splicing the delay feature, spectral centroid feature, frequency band energy ratio feature, and CNN high-order features into a joint acoustic feature vector.

[0013] Optionally, generate a node feature vector representing the acoustic correlation between the cages. Specifically, it includes: establishing and training a GNN model. The pre-established GNN model includes nodes and edges, where each node represents a farming cage and the edge represents the acoustic propagation path loss situation between the nodes. By encoding and aggregating the features of the nodes and edges, a node feature vector for each node is generated.

[0014] Optionally, select the node feature vector corresponding to the cage closest to the position coordinates of the damaged noise source and perform dynamic weighted fusion based on the attention mechanism with the joint acoustic feature vector to obtain a cross-modal feature vector. Input the cross-modal feature into an SVM classifier, and output the probability value of damaged / undamaged, where the activation function of the SVM classifier is the Softmax function.

[0015] Optionally, if the SVM classifier outputs a high probability of cage damage, use the particle filter algorithm to further optimize the position coordinates of the damaged noise source and output the damaged position information of the target cage.

[0016] The beneficial effects of the present invention are as follows: By continuously collecting passive acoustic signals through hydrophone units and combining the high-frequency sonar pulse echoes directionally emitted by the underwater detection robot, a dual detection mechanism of passive noise positioning and active sonar scanning is realized. The passive acoustic signals quickly lock the potential damaged area through the TDOA algorithm, and the time-frequency analysis of the active sonar accurately extracts local features. The collaborative work of the two breaks through the limitations of traditional single detection means; By introducing the multimodal feature fusion of convolutional neural network and graph neural network, the robustness of damage recognition is further enhanced. The high-order acoustic features extracted by CNN from the time-frequency distribution matrix can capture microscopic patterns such as spectral distortion and texture abnormality caused by damage; the correlation of acoustic wave propagation path loss between cages modeled by GNN reflects the propagation characteristics of damage signals from the topological level. After the two types of features are dynamically weighted and fused by the attention mechanism, a cross-modal feature vector is formed, which solves the problem of the disconnection of time-domain, frequency-domain, and spatial-domain features in traditional methods; The particle filter algorithm is used to iteratively optimize the initial positioning result, and the damage location output with sub-meter accuracy is realized. By establishing a state model to describe the spatial constraints of the damage point, and combining the time-delay observation data of multiple hydrophones and the spectral features of active sonar, the particle filter dynamically adjusts the particle weights and gradually converges to the true damage coordinates. It can still remain stable in the presence of multipath effects or environmental noise interference, and the positioning error is reduced from 3-5 meters of the initial TDOA to within 0.5 meters; This application changes the high-risk operation mode that traditionally relies on manual diving inspections. It can complete damage positioning and verification without manual intervention, eliminate the safety hazards of divers operating in deep water, turbid or dangerous waters, and realize continuous monitoring of large-scale aquaculture areas, providing reliable technical support for the sustainable development of deep-sea aquaculture. Description of the Drawings

[0017] Figure 1 It is a schematic diagram of the module connection of a broken net leakage detection system for a seawater aquaculture cage in an embodiment of this application. Detailed Embodiments

[0018] The technical solutions of the present invention will be further elaborated in detail below in conjunction with the drawings in the specification and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this invention belongs. The terms used in the specification of this invention herein are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. In the following description, the expression "some embodiments" is mentioned, which describes a subset of all possible embodiments. However, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0019] In the following description, a large number of specific details are given to provide a more thorough understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known to the public are not described.

[0020] It should be understood that the present invention can be implemented in different forms and should not be construed as limited to the embodiments presented herein. On the contrary, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the present invention to those skilled in the art. And the purpose of the terms used herein is only to describe specific embodiments and not to limit the present invention. As used herein, the singular forms "a", "an" and "the" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms "comprising" and / or "including", when used in this specification, identify the presence of the stated features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups. As used herein, the term "and / or" includes any and all combinations of the related listed items.

[0021] It should be further noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "inner", "outer", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation.

[0022] To thoroughly understand the present invention, detailed structures will be presented in the following description to illustrate the technical solutions proposed by the present invention. The alternative embodiments of the present invention are described in detail as follows. However, in addition to these detailed descriptions, the present invention can also have other implementations.

[0023] Please refer to the attached Figure 1 , this application provides a system for detecting net breakage and leakage in a seawater aquaculture cage, and the system includes: A plurality of fixed hydrophone units for acquiring passive acoustic signals, where the passive acoustic signals include environmental noise signals and turbulent noise signals from different sources; An underwater detection robot equipped with an active sonar for transmitting high-frequency sonar pulses and collecting reflected echo signals; A control unit for sending driving instructions to the underwater detection robot and for receiving the reflected echo signals and passive acoustic signals; A signal processing unit configured to extract time delay and spectral features from the reflected echo signals and determine the location information of the breakage noise source based on the passive acoustic signals; An analysis module, configured with a convolutional neural network and a graph neural network, where the convolutional neural network is used to extract joint acoustic feature vectors from spectral features, and the graph neural network is used to generate node feature vectors representing the acoustic correlations between cages. After fusing the two, damage detection is performed; A position optimization unit, configured to iteratively optimize the position information of the damaged noise source and output the final damaged position of the cage.

[0024] Specifically, passive acoustic signals are continuously collected by a fixed hydrophone unit deployed under the deep-sea aquaculture cage. The fixed hydrophone unit can be suspended and fixed under the deep-sea aquaculture cage or fixed on the seabed. The passive acoustic signals it collects are input into the signal processing unit for analysis to determine whether there is a damaged source noise signal. When there is a damaged source noise signal, the position information of the damaged noise source is solved and mapped within the geometric space range of the target cage. The control unit drives the underwater detection robot to approach the target cage according to the positioning result, emits pulse signals through a high-frequency active sonar and collects the reflected echoes. The signal processing unit extracts time-frequency features from the echo signals, and at the same time obtains the time delay value of the reflection path based on cross-correlation analysis. The analysis module uses a convolutional neural network to perform multi-level feature extraction on the time-frequency diagram, captures the local spectral distortion and texture anomalies caused by damage to form acoustic feature vectors. At the same time, a graph neural network is used to construct an acoustic correlation model between cage nodes, quantifies the acoustic propagation loss as edge weights, aggregates the acoustic state features of adjacent cages through graph convolution, and generates node feature vectors representing spatial relationships. The two types of features are dynamically weighted and fused by an attention mechanism and then input into a classification model to comprehensively judge the probability of damage occurrence based on time domain, frequency domain, and spatial domain information. After determining that the cage is damaged, the position optimization unit iteratively optimizes the position information of the initial damaged noise source based on the particle filter algorithm, describes the spatial constraints of the damaged point through the state equation, fuses multi-source acoustic data through the observation equation, and gradually narrows the range of the damaged position in combination with the resampling strategy until it converges to the true damaged coordinates, and finally outputs the precise damaged position of the target cage with sub-meter accuracy. This system uses passive noise positioning to guide active sonar for directional scanning, combines data-driven and physical models to enhance the detection robustness, and realizes the full-process automated detection from wide-area anomaly discovery to local precise diagnosis.

[0025] It should be noted that the selection of the fixed hydrophone unit and the underwater detection robot, as well as the deployment method of the fixed hydrophone unit underwater, are conventional technical means in the art, and are not described in detail in this embodiment.

[0026] In some embodiments, extracting the time delay and spectral features from the reflected echo signal specifically includes: For the reflected echo signal Preprocessing is performed, and the preprocessing includes band-pass filtering and noise reduction. Specifically, for the original reflected echo signal, the digital band-pass filter retains the signal in the effective frequency band of 20 kHz to 200 kHz, eliminates low-frequency ship noise and high-frequency interference, and suppresses random noise through the wavelet threshold denoising algorithm to improve the signal-to-noise ratio. The expression of the wavelet threshold denoising algorithm is as follows:

[0027] where WT represents wavelet transform, is the threshold parameter. Cross-correlation analysis is performed on the transmitted signal and the preprocessed reflected echo signal to obtain the time delay value. The calculation formula for cross-correlation analysis is:

[0028] is the transmitted signal, is the preprocessed signal. The time delay value is determined by finding the peak position of the cross-correlation function:

[0029] This value corresponds to the round-trip time difference of sound waves and is converted into a physical distance , and then the short-time Fourier transform is performed on the preprocessed reflected echo signal to generate a time-frequency distribution matrix. The spectral centroid and frequency band energy ratio are extracted from the time-frequency distribution matrix as spectral features. Among them, performing the short-time Fourier transform on the preprocessed reflected echo signal and extracting the spectral centroid and frequency band energy ratio from the time-frequency distribution matrix are conventional technical means in the art, and will not be described in detail in this embodiment.

[0030] The breakage noise is essentially turbulent noise, which is caused by pressure pulsations in the turbulent flow field. When the turbulent flow passes through an obstacle or boundary, pressure pulsations will generate sound waves. Turbulence is the irregular and random motion in a fluid, and its instability causes the fluid molecules to collide and vibrate with each other, thus generating noise. When the aquaculture net cage is damaged, when the water flow passes through the break, the flow velocity and direction will change, resulting in unstable flow, and then forming turbulence. This turbulence will cause pressure pulsations, thus generating turbulent noise, and these noises may be mixed in the noises of other environments or the vortex noises generated by other sources.

[0031] Therefore, in some embodiments, determining the position information of the breakage noise source according to the passive acoustic signal specifically includes: M hydrophone units acquire M passive acoustic signals, perform band-stop filtering on the M passive acoustic signals to suppress the ship noise frequency band and the biological noise frequency band, and obtain the target acoustic signal containing a specific frequency band; Determine whether there is a breakage noise in the target acoustic signal. If there is, perform cross-correlation processing on multiple passive acoustic signals to obtain the time difference of arrival. Based on multiple time differences of arrival, establish a TDOA positioning equation set, and use the least squares method to solve the TDOA positioning equation set to obtain the position coordinates of the breakage noise source.

[0032] Further, perform cross-correlation processing on multiple passive acoustic signals to obtain the time difference of arrival. Based on multiple time differences of arrival, establish a TDOA positioning equation set, which specifically includes: taking the hydrophone unit that first receives the target acoustic signal as the reference hydrophone unit, performing cross-correlation processing on the target acoustic signal received by the reference hydrophone unit and the target acoustic signals received by the other M - 1 hydrophone units respectively, obtaining the time difference of arrival of the target acoustic signal of the breakage noise source between the other M - 1 hydrophone units and the reference hydrophone unit, and forming a set of positioning information groups based on multiple time differences of arrival, with a total of M - 1 sets of positioning information groups.

[0033] Specifically, perform band-stop filtering on the original signals collected by M hydrophone units. Design the stopband range as the ship noise frequency band and the biological noise frequency band, and obtain the target acoustic signal after suppressing interference. If it is determined that there is a breakage noise in the target acoustic signal, select the reference hydrophone unit M = 1, and calculate the cross-correlation function between the remaining hydrophone units and the reference hydrophone unit:

[0034] Improve the time delay resolution through PHAT weighting:

[0035] Extract the peak position to obtain the time delay difference:

[0036] Construct a TDOA positioning equation set based on the time delay difference: Assume the noise source coordinates are and the reference hydrophone unit coordinates are Then the equation between the reference hydrophone unit and one other hydrophone unit is expressed as:

[0037] where is the coordinate of one other hydrophone unit, and so on. Perform cross-correlation processing on the target acoustic signal received by the reference hydrophone unit and the target acoustic signals received by the other M - 1 hydrophone units respectively to obtain multiple equations, form a set of positioning information groups with multiple equations, use the L - M algorithm to transform the set of positioning information groups into a non-linear least squares problem and solve it, minimize the objective function, and obtain the initial position coordinates of the breakage noise source.

[0038] Furthermore, the frequency band of turbulent noise is closely related to the characteristics of turbulence, including the intensity, scale of turbulence, and the motion characteristics of the fluid, etc. Different turbulent structures will generate noises of different frequencies. For example, small-scale turbulent vortices will generate high-frequency noises, while large-scale turbulent vortices will generate low-frequency noises. Therefore, after the aquaculture cage is damaged, the turbulent noise generated by the water flow passing through the break may have a specific frequency band and signal power spectral density. Therefore, to determine whether there is breakage noise in the target acoustic signal, it specifically includes: calculating the signal power spectral density in the target acoustic signal, and its calculation formula is:

[0039] wherein, is the Fourier transform of the target acoustic signal is the acquisition period. If the signal power spectral density exceeds the threshold , it indicates that there is breakage noise in the target acoustic signal.

[0040] When obtaining the position coordinates of the breakage noise source, select the cage closest to the position coordinates of the breakage noise source, and the control unit of the underwater detection robot issues a driving instruction to control the underwater detection robot to move to the cage to emit high-frequency sonar pulses and collect the reflected echo signals.

[0041] In some embodiments, after obtaining the breakage noise source coordinates and the echo signal, and obtaining the time delay and time-frequency distribution matrix from the echo signal, the convolutional neural network in the analysis module is used to extract the joint acoustic feature vector from the spectral features. The process of extracting the joint acoustic feature vector from the spectral features aims to fuse multi-dimensional acoustic information and improve the robustness of breakage detection.

[0042] Furthermore, extracting the joint acoustic feature vector from the spectral features specifically includes: establishing and training a CNN model, inputting the time-frequency distribution matrix into the pre-trained CNN model to obtain the CNN high-order features, and splicing the time delay features, spectral centroid features, frequency band energy ratio features, and CNN high-order features into a joint acoustic feature vector.

[0043] The core principle of extracting the joint acoustic feature vector is to extract high-order abstract features from the time-frequency distribution matrix through a convolutional neural network and splice them with the time delay features, spectral centroid features, and frequency band energy ratio features to form a cross-domain joint representation. Specifically, the time-frequency distribution matrix is generated by short-time Fourier transform, representing the energy distribution of the signal in the time window t and frequency f. Inputting the time-frequency distribution matrix into the pre-trained CNN model, the model extracts spatial features through multi-layer convolution and pooling operations, and finally outputs a high-order feature vector , where D is the feature dimension, and the final time-delay feature, spectral centroid feature, frequency band energy ratio feature, and high-order feature vector are fused into a joint acoustic feature vector through a concatenation operation , that is , which fully utilizes the information complementarity in the time domain (time delay), frequency domain (centroid and energy ratio), and spatial domain (CNN features), and enhances the model's ability to distinguish damaged features. For example, the time-delay feature can locate the spatial distance of the damage point, the spectral centroid reflects the main frequency characteristics of the noise, and the CNN features capture local abnormal patterns in the time-frequency diagram. Through joint optimization, the system can more accurately identify abnormal signals in the presence of minor damage or complex interference environments, providing highly discriminative input features for subsequent classification and localization. Furthermore, a node feature vector representing the acoustic correlation between cages is generated, specifically including: establishing and training a GNN model. The pre-established GNN model includes nodes and edges, where each node represents a farming cage, and the edge represents the sound wave propagation path loss between nodes. By encoding and aggregating the features of nodes and edges, a node feature vector for each node is generated.

[0044] The process of generating the node feature vector representing the acoustic correlation between cages relies on the joint modeling of the topology structure of the farming cages and the sound wave propagation characteristics by the graph neural network (GNN). Its core is to abstract each cage as a node in the graph, quantify the sound wave propagation path loss between nodes as the weight of the edge, and gradually generate a node feature vector reflecting the global acoustic correlation through message passing and feature aggregation in multiple graph convolutional layers. After multiple rounds of iteration, the node features finally encode the sound wave propagation path loss, topological correlation, and local acoustic state between cages. For example, an edge with a large path loss will suppress the intensity of feature transmission between adjacent nodes, while high-frequency resonance characteristics may spread to neighboring nodes through multiple aggregations, thus reflecting the cooperative acoustic behavior of the cage group in the feature vector. Finally, the node feature vector not only contains local acoustic characteristics but also can reflect abnormal propagation paths in the global network, thereby enhancing the robustness of damage detection. For example, when a cage is damaged, its abnormal acoustic signal is weighted and transmitted through the path loss of adjacent cages, and finally forms a significant difference in the node features, facilitating the classifier to identify.

[0045] In an optional implementation, the node feature vector corresponding to the cage closest to the position coordinates of the damaged noise source is dynamically weighted and fused with the joint acoustic feature vector based on the attention mechanism to obtain a cross-modal feature vector. The cross-modal feature is input into an SVM classifier, and the probability value of damaged / undamaged is output, where the activation function of the SVM classifier is the Softmax function.

[0046] Specifically, the core of the attention mechanism is to adaptively allocate weights to highlight key information by calculating the correlation between two feature vectors. Specifically, two feature vectors are first mapped into a query vector and a key vector respectively, and the attention scores are calculated through the dot product. The softmax function normalizes the scores into a weight distribution. Subsequently, the value vector is weighted and summed with the attention weights to obtain the fused cross-modal feature vector , where , where are the attention weights, is the value vector, is the node feature vector. This process dynamically balances the contributions of local acoustic features (such as spectral distortion) and global topological correlations (such as path loss). For example, when the damaged noise feature is significant, the attention weight tends to the joint acoustic feature; while when the acoustic correlation between cages is complex, the weight of the node feature is enhanced.

[0047] Cross-modal feature vector Subsequently, it is input into a support vector machine (SVM) classifier for predicting the probability of damage. The SVM divides the feature space through a hyperplane, and its output is the output probability value. The Softmax function maps the decision value into a probability distribution, and the SVM parameters are optimized by maximizing the margin loss function, making the classification boundary as far away from the two types of samples as possible. For example, when the cross-modal feature vector contains high-frequency energy anomalies, such as a sudden increase in the frequency band energy ratio and a sudden change in path loss, that is, when the node feature is abnormal, the probability of damage output by Softmax increases significantly.

[0048] The attention mechanism strengthens the key modal information through adaptive weight allocation, while the SVM combines Softmax to map high-dimensional features into an interpretable probability output. The two work together to achieve an efficient conversion from multi-source data to decision results. In a complex marine noise background, if the weight of the node feature of a certain cage decreases due to abnormal path loss, but the joint acoustic feature shows obvious time delay and spectral distortion, the attention mechanism can still ensure the accurate identification of the damaged signal through dynamic weighting, thereby improving the environmental adaptability and detection accuracy of the system In some embodiments, if the SVM classifier outputs a high probability of cage damage, the particle filter algorithm is used to further optimize the position coordinates of the damaged noise source and output the damaged position information of the target cage.

[0049] Specifically, when the SVM classifier determines that the probability of the cage being damaged is high, the system iteratively optimizes the initial position of the damage noise source through the particle filter algorithm to improve the positioning accuracy. The particle filter is based on the Monte Carlo method, represents the posterior probability distribution through a set of weighted particles, and dynamically updates the particle state using the observed data, as follows: According to the initial position coordinates of the damage noise source obtained from the TDOA positioning equation, N particles are assigned to the possible state space, and the covariance matrix is set based on the positioning error. Each particle has the same initial weight. For each particle, the state at the next moment is predicted according to the motion model: Among them, the state space model describes the dynamic characteristics of the damage position. Assuming that the damage position is stationary in a short period of time, the state equation is:

[0050] Where is the coordinate at time step k, is the process noise. The observation model maps the state to the sensor data. The observed data includes the time delay measurement of multiple hydrophones, the spectral characteristics of the reflected echo of the active sonar, etc. For example, the sound wave propagates from the damage position to the actual time delay of the m-th hydrophone:

[0051] Adjust the particle weights using the observed data. For each particle calculate its matching degree with the observed data, generate a new particle set according to the weight distribution, avoid particle degeneracy, adopt the systematic resampling strategy, copy the high-weight particles according to the weight ratio, and eliminate the low-weight particles to ensure particle diversity. The optimized position is obtained through weighted averaging: , through multiple rounds of iteration, the particles gradually converge to the high-probability region, and finally the optimized position is output. In the complex ocean noise environment, the initial TDOA positioning may have a deviation of several meters due to the multipath effect. After the particle filter fuses multi-sensor data, the positioning error can be reduced to the sub-meter level. This process effectively combines the SVM probability trigger and the acoustic wave propagation equation, taking into account both efficiency and accuracy, and provides a reliable spatial positioning result for the cage damage detection.

[0052] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should be covered within the protection scope of the present invention. The protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A system for breaking and leak detection of marine aquaculture cages, characterized in that: The system comprises: A plurality of fixed hydroacoustic units, used for acquiring passive acoustic signals, wherein the passive acoustic signals include environmental noise signals and turbulent noise signals from different sources; An underwater inspection robot equipped with active sonar, used to emit high-frequency sonar pulses and collect reflected echo signals; A control unit, used for issuing a driving instruction to the underwater detection robot, and for receiving the reflected echo signal and the passive acoustic signal; A signal processing unit configured to extract time delay and spectrum characteristics from the reflected echo signal and determine the location information of the damage noise source according to the passive acoustic signal; An analysis module, which is configured with a convolutional neural network and a graph neural network. The convolutional neural network is used to extract a joint acoustic feature vector from the frequency spectrum features, and the graph neural network is used to generate a node feature vector representing the acoustic association between the cages. The two are combined to perform damage detection; The position optimization unit is configured to iteratively optimize the position information of the damage noise source and output the final cage damage position.

2. A system for breaking and leak detection of marine aquaculture cages according to claim 1, characterized in that: Extract the time delay and spectrum characteristics from the reflected echo signal, including: Preprocessing the reflected echo signal, wherein the preprocessing includes bandpass filtering and noise reduction processing; Perform cross-correlation analysis on the transmitted signal and the pre-processed reflected echo signal to obtain the delay value; The preprocessed reflected echo signal is subjected to short-time Fourier transform to generate a time-frequency distribution matrix, and the spectrum center of gravity and the frequency band energy ratio are extracted from the time-frequency distribution matrix as spectrum features.

3. A system for breaking and leak detection of seawater aquaculture cages according to claim 2, characterized in that: Determining the location information of the damage noise source according to the passive acoustic signal specifically includes: M hydrophone units acquire M passive acoustic signals, perform band-stop filtering on the M passive acoustic signals, suppress the frequency band of ship noise and biological noise, and obtain a target acoustic signal containing a specific frequency band; Determine whether there is damage noise in the target acoustic signal. If so, cross-correlate multiple passive acoustic signals to obtain arrival time differences, establish a TDOA positioning equation group based on multiple arrival time differences, and use the least squares method to solve the TDOA positioning equation group to obtain the position coordinates of the damage noise source.

4. A system for breaking and leak detection of seawater aquaculture cages according to claim 3, characterized in that: When the position coordinates of the damaged noise source are obtained, the net cage closest to the position coordinates of the damaged noise source is selected, and the control unit underwater detection robot issues a driving instruction to control the underwater detection robot to move to the net cage to emit high-frequency sonar pulses and collect reflected echo signals.

5. The system for breaking and leak detection of seawater aquaculture cages according to claim 3, characterized in that: Determining whether there is damage noise in the target acoustic signal specifically includes: calculating the signal power spectrum density in the target acoustic signal, and if the signal power spectrum density exceeds a threshold , it means that there is corruption noise in the target acoustic signal.

6. A system for breaking and leak detection of seawater aquaculture cages according to claim 5, characterized in that: After cross-correlating multiple passive acoustic signals, an arrival time difference is obtained, and a TDOA positioning equation group is established based on multiple arrival time differences, specifically including: taking the hydroacoustic unit that receives the target acoustic signal earliest as the reference hydroacoustic unit, cross-correlating the target acoustic signal received by the reference hydroacoustic unit with the target acoustic signals received by other M-1 hydroacoustic units, respectively, to obtain the arrival time difference between the target acoustic signal of the damage noise source reaching the other M-1 hydroacoustic units and the reference hydroacoustic unit, and forming a group of positioning information groups based on multiple arrival time differences, forming a total of M-1 positioning information groups.

7. A system for breaking and leak detection of seawater aquaculture cages according to claim 6, characterized in that: A joint acoustic feature vector is extracted from the spectral features, specifically including: establishing and training a CNN model, inputting the time-frequency distribution matrix into the pre-trained CNN model to obtain CNN high-order features, and concatenating time delay features, spectral center of gravity features, frequency band energy ratio features, and CNN high-order features into a joint acoustic feature vector.

8. A system for breaking and leak detection of seawater aquaculture cages according to claim 7, characterized in that: Generate a node feature vector that characterizes the acoustic correlation between cages, specifically including: establish and train a GNN model, the pre-established GNN model includes nodes and edges, wherein each node represents a breeding cage, and the edge represents the sound wave propagation path loss between nodes, and generate a node feature vector for each node by encoding and aggregating the features of the nodes and edges.

9. A system for breaking and leak detection of marine aquaculture cages according to claim 8, characterized in that: The node feature vector corresponding to the box closest to the position coordinates of the damage noise source is selected and dynamically weighted fused with the joint acoustic feature vector based on the attention mechanism to obtain the cross-modal feature vector. The cross-modal feature is input into the SVM classifier and the probability value of damage / non-damage is output, where the activation function of the SVM classifier is the Softmax function.

10. A system for breaking and leak detection of marine aquaculture cages according to claim 9, characterized in that: If the SVM classifier outputs that the probability of the cage being damaged is high, a particle filter algorithm is used to further optimize the position coordinates of the damaged noise source, and output the damaged position information of the target cage.

Images