Aquaculture net cage damage monitoring method based on multi-sensor information fusion

Abstract

The invention provides an aquaculture net cage damage monitoring method based on multi-sensor information fusion. The aquaculture net cage damage monitoring method comprises the following steps that 1, an equivalent numerical model of a net cage is established according to the size of a gravity deepwater type net cage; step 2, dividing the net cage into eight areas; step 3, carrying out analog simulation on tension and acceleration signals under the conditions of netting completeness and damage by utilizing the equivalent numerical model of the culture net cage; step 4, establishing a neural network damage monitoring model based on a CNN neural network, and training the neural network damage monitoring model by using the obtained simulation data set; step 5, obtaining basic probability assignment mA and mB of each group of sensors through a neural network damage monitoring model; and 6, synthesizing the BPA elementary probability assignment mA and mB of any group of two sensors into a new evidence C by using a Dempster synthesis rule, performing intra-region fusion, and judging the state of the netting according to the elementary probability assignment m and the region n of the final comprehensive evidence C.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-sensor information fusion-based net cage damage monitoring method, belonging to the field of ocean engineering and information technology integration. BACKGROUND

[0002] With the increasing global demand for aquatic products, the aquaculture industry has developed rapidly. However, the problem of net cage damage, especially in marine or large water body environments, can lead to the escape of farmed fish, causing significant environmental and economic losses. Therefore, timely and accurate detection of net cage damage has become an urgent problem to be solved.

[0003] Traditional net cage detection methods mainly rely on manual divers or underwater cameras for visual inspection. These methods have many problems, such as high labor costs, low detection efficiency, and difficulty in achieving real-time monitoring. Moreover, the underwater environment is complex, with low visibility, sea waves and water flow affecting factors, further increasing the limitations of traditional methods.

[0004] With the development of artificial intelligence and Internet of Things technology, multi-sensor information fusion-based damage detection technology can achieve high intelligence and automation. Through machine learning algorithms, the monitoring system can learn and optimize the detection model from a large amount of historical data, thereby achieving adaptive adjustment during detection and further improving the accuracy and efficiency of detection.

[0005] The existing Chinese public patent CN02145476.0 and CN202111053606.8 buried wire monitoring method is to embed metal wires in the netting. When the netting is damaged, the metal wires will be broken and form a loop with seawater and underwater electrodes. At this time, the transmitter will receive the current signal, thereby judging the netting damage.

[0006] The literature "Design of deep water net cage netting monitoring system based on ultra-short baseline" uses sonar to monitor the sound wave image outside the netting. First, a sonar warning zone is constructed around the netting. When the netting is damaged, the fish will escape from the net cage. At this time, the sonar will detect the sound wave image inside and outside the net cage, which will have a significant difference, thereby judging that the netting has been damaged.

[0007] Chinese public patents CN201911332867.6 and CN201910316890.X image detection method is to install underwater cameras on ROV, bionic fish and other underwater robots to collect netting images, and use image recognition algorithms to detect whether the netting has been damaged.

[0008] The machine learning monitoring method of Chinese patent CN202111555038.1 is to collect data of the netting by a sensor, use the data as a training set and train it by an intelligent algorithm, obtain a damage monitoring model after training, and the sensor will collect data of the netting in real time. When the netting is damaged, the damage monitoring model will identify abnormal signals and judge whether the netting is damaged.

[0009] (1) Artificial detection method: artificial can only dive regularly to monitor the state of the net cage, cannot realize real-time monitoring, and the monitoring personnel diving into the sea will have certain danger.

[0010] (2) Buried wire monitoring method: metal wires need to be embedded in the netting, which will increase the load of the netting, reduce the service life of the netting system, and increase the loss of the netting system.

[0011] (3) Sonar monitoring method: when wild fish swim outside the netting, the sonar monitoring method will have false detection, and the sonar monitoring method can only judge that the netting is damaged, but cannot judge the specific position of the damaged netting.

[0012] (4) Image detection method: the image detection method has high requirements for water quality. When the water body is turbid, the detection effect may not be ideal. The image detection method needs to continuously use underwater robots for inspection, and the cost is high.

[0013] Machine learning monitoring method: the current machine learning monitoring method only uses the tension sensor as data input, the data is too single, and the detection accuracy and anti-interference ability to the environment are not high. SUMMARY

[0014] Invention purpose: in view of the deficiencies in the prior art, the present application provides a breeding net cage damage monitoring method based on multi-sensor information fusion to solve at least one technical problem mentioned in the background art.

[0015] Technical scheme: the breeding net cage damage monitoring method based on multi-sensor information fusion comprises the following steps:

[0016] Step 1: establish an equivalent numerical model of the net cage according to the size of the gravity deep water net cage; the foundation includes netting, net joint, floating ring and sinker;

[0017] Step 2: divide the net cage into three layers, i.e. upper, middle and lower layers, and evenly distribute four points in each layer, a total of 12 points, define one tension sensor and one acceleration sensor as a group of sensors, install a group of sensors on the 12 points on the net cage, and collect the tension signal of the netting in real time, i.e. evidence A, and the acceleration signal, i.e. evidence B; 12 points divide the netting into 8 areas, i.e. areas 1 to 8;

[0018] Step 3: Set different irregular wave parameters according to the environment where the cage is located. Under different wave parameters, use the equivalent numerical model of the aquaculture cage to simulate the tension and acceleration signals of the net under intact and damaged conditions to obtain a simulation data set, and generate a simulation data waveform based on the data set;

[0019] Step 4: Establish a neural network damage monitoring model based on the CNN neural network and train the neural network damage monitoring model using the obtained simulation data set. The simulated tension, acceleration data, and wave parameters of the net are used as inputs, and the probability of each sensor identifying the net damage is used as output.

[0020] Step 5: Obtain the basic probability assignment m of each group of sensors through the neural network damage monitoring model A and m B ; The details are as follows:

[0021] Step 5.1, constructing a recognition framework, which includes three states of the net cage: normal state of the net, damaged state of the net, and state of foreign matter attached to the net. The basic propositions of the recognition framework are defined as normal state J, damaged state H, and foreign matter attached state I;

[0022] Step 5.2: Based on the data obtained from the tension sensor and acceleration sensor, the neural network damage monitoring model is used to obtain the preliminary diagnosis results, and the preliminary diagnosis results are assigned to the BPA basic probability assignment function, i.e., m A and m B ;

[0023] Step 6: Use Dempster's synthesis rule to assign the BPA basic probability of any group of two sensors to m A and m B Combined into new evidence C, the synthesis rules are as follows:

[0024]

[0025] Where k is the conflict coefficient, which indicates the degree of conflict between the evidence given by the two sensors; m C is m A and m B The fused basic probability assignment function is combined with the position of the tension sensor and acceleration sensor on the cage to obtain m C i, where i is the position on the cage where the tension sensor and acceleration sensor are located, and the values ​​are 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12;

[0026] According to the divided mesh area, the 12 m C i performs fusion within the region and obtains 8 final BPA basic probability assignments. According to the basic probability assignment m of the final comprehensive evidence C区域n , judge the state of the net, wherein n represents the area number, taking 1 to 8.

[0027] The step 1 is specifically: using an equivalent numerical model to represent the deep-sea culture net cage in Orcaflex marine engineering dynamic simulation tool; using a "Line" model to represent the net line, a "3Dbouy" model to represent the net joint, and a "6Dbouy" model to represent the float; the model bottom is simulated by 16 sinkers with 3D buoys, and is symmetrically distributed around the bottom of the net cage; the net cage model contains 64 "6Dbouy" models, 1280 "3Dbouy" models and 2656 "Line" in total; the whole net cage system is composed of a mooring system, a float system and a fishing net system.

[0028] The step 2 is specifically: the points measured by the combination of 12 tension sensors and acceleration sensors are marked as points 1 to 12, and areas 1 to 8 are respectively:

[0029] Area 1: points 1, 2, 5, 6; Area 2: points 2, 3, 6, 7; Area 3: points 3, 4, 7, 8; Area 4: points 1, 4, 5, 8; Area 5: points 5, 6, 9, 10; Area 6: points 6, 7, 10, 11; Area 7: points 7, 8, 11, 12; Area 8: points 5, 8, 9, 12.

[0030] The step 3 is specifically: the different sea wave parameters include wave height of 0.2m-1.6m and wave period of 1s-5s.

[0031] In the step 4, the method for establishing the neural network damage monitoring model comprises:

[0032] Based on the simplification and variation of the VGG16 network architecture, the continuous wavelet transform (CWT) is used to extract features from the hydrodynamic data, and the extracted features are provided as input to the network for further analysis;

[0033] The network contains 12 convolutional layers, each using a 3x3 convolutional kernel and adopting a "same" padding method to ensure that the size of the feature map remains unchanged;

[0034] The number of filters in the convolutional layer gradually increases from 64 to 128, 256, 512 and 4096 to extract more complex features; after each convolutional layer, a ReLU activation function is used to enhance the nonlinearity of the network, and a batch normalization layer is used to speed up the training process;

[0035] The pooling layer uses a max-pooling operation with a fixed step of 2 in each layer to reduce the spatial dimension of the feature map and reduce the computational load; the network finally classifies through a fully connected layer, and uses a Softmax function to output the probability of each class.

[0036] The number of convolutional layers and pooling layers is simplified, specifically as follows: in the first convolutional module, the original 2 convolutional layers are reduced to 1; in the second convolutional module, 2 convolutional layers are retained, and the number of channels is reduced from the original 128 to 64; in the third convolutional module, the original 3 convolutional layers are reduced to 2, and the number of channels is reduced from 256 to 128; in the fourth module, the number of channels is reduced from 512 to 256, and 3 convolutional layers are retained; the fifth module remains unchanged and contains 3 convolutional layers and 512 channels; the fully connected part uses a convolutional layer instead of the original first fully connected layer, and two subsequent fully connected layers are set for feature compression and classification output; the overall structure retains the deep feature extraction capability of VGG16 while reducing the number of parameters.

[0037] The step 5.2 is specifically:

[0038] For the evidence A and B of the tension sensor and the acceleration sensor at any position on the net cage, the basic probability assignment is m A and m B , the tension sensor monitoring result is:

[0039] m A ({normal state})=a1

[0040] m A ({damaged state})=a2

[0041] m A ({foreign object attachment state})=a3

[0042] Wherein, a1 represents the probability of the normal state of the position point, taking 0 to 1, a2 represents the probability of the damaged state of the position point, taking 0 to 1, and a3 represents the probability of the foreign object attachment state of the position point, taking 0 to 1;

[0043] The acceleration sensor monitoring result is:

[0044] m B ({normal state})=b1

[0045] m B ({damaged state})=b2

[0046] m B ({foreign object attachment state})=b3

[0047] Wherein, b1 represents the probability of the normal state of the position point, taking 0 to 1, b2 represents the probability of the damaged state of the position point, taking 0 to 1, and b3 represents the probability of the foreign object attachment state of the position point, taking 0 to 1.

[0048] In the step 6,

[0049] The conflict coefficient K represents the degree of conflict between the evidence given by two sensors, and is calculated as follows:

[0050]

[0051] The probability product of each team disjoint hypothesis combination is calculated and summed up:

[0052] K = m A ({normal state}) x m B ({damaged state}) + m A ({normal state}) x m B ({foreign matter attached state}) + m A ({damaged state}) x m B ({normal state}) + m A ({damaged state}) x m B ({foreign matter attached state}) + m A ({foreign matter attached state}) x m B ({normal state}) + m A ({foreign matter attached state}) x m B ({damaged state})

[0053] Then:

[0054] K = (a1 x b2) + (a1 x b3) + (a2 x b1) + (a2 x b3) + (a3 x b1) + (a3 x b2)

[0055] The BPA basic probability assignment m after synthesis is calculated C ,

[0056] For the normal state:

[0057]

[0058] For the damaged state:

[0059]

[0060] For the foreign matter attached state:

[0061]

[0062] Wherein, c1 represents the probability of the normal state of the position point after synthesis, taking 0 to 1, c2 represents the probability of the damaged state of the position point after synthesis, taking 0 to 1, and c3 represents the probability of the foreign matter attached state of the position point after synthesis, taking 0 to 1.

[0063] In step 6, the fusion in the region specifically includes:

[0064] The fusion results of the four point tension sensors and acceleration sensors in region 1 are preliminarily fused, and the fusion result is:

[0065] m C 1,m C 2,m C 5,m C 6

[0066] Wherein:

[0067] m i ({normal})=a i ,m i ({damage})=b i ,m i ({foreign matter attachment})=c i ,

[0068] The two BPA basic probability values of point 1 and point 2 are set as m1 and m2, and are combined into m 12 , which is defined as follows:

[0069] K=m A ({normal state})×m B ({damage state})+m A ({normal state})×m B ({foreign matter attachment state})+m A ({damage state})×m B ({normal state})+m A ({damage state})×m B ({foreign matter attachment state})+m A ({foreign matter attachment state})×m B ({normal state})+m A ({foreign matter attachment state})×m B ({damage state})

[0070] According to the monitoring results of the sensors at the four positions in region 1, a1, a2, a3, b1, b2, b3 are obtained, k is calculated according to the formula in step 7, and the synthesized BPA basic probability value m 12 is further calculated.

[0071] For normal state:

[0072]

[0073] For damage state:

[0074]

[0075] For foreign matter attachment state:

[0076]

[0077] The synthesis of m1 and m2 is completed to obtain m 12 Then, the following synthesis is continued

[0078] m 12 ⊕m5=m 125 ,m 125 ⊕m6=m 1256 ,

[0079] Finally, the BPA basic probability assignment of region 1 is obtained, that is:

[0080] m 区域1 =m 1256

[0081] The fusion result of region 1 is finally obtained:

[0082]

[0083] Wherein, m({normal state}) represents the probability of region 1 being normal on the net, taking 0 to 1; m({damage state}) represents the probability of region 1 being damaged on the net, taking 0 to 1; m({foreign matter attachment state}) represents the probability of region 1 being foreign matter attachment state on the net, taking 0 to 1;

[0084] The BPA basic probability assignments m 区域2 , m 区域3 , m 区域4 , m 区域5 , m 区域6 , m 区域7 , m 区域8 of regions 2 to 8 are obtained according to the above synthesis rules.

[0085] Beneficial effects:

[0086] (1) The aquaculture net cage damage monitoring technology based on multi-sensor information fusion can monitor the state of the aquaculture net cage in real time. The data of the tension sensor and the acceleration sensor can quickly and accurately determine the real-time state of the aquaculture net cage. Compared with a single sensor, multi-sensor fusion can comprehensively utilize the data of different sensors, reduce the errors and blind spots that may exist in a single sensor, and improve the accuracy and reliability of monitoring.

[0087] (2) The aquaculture net cage damage monitoring technology based on multi-sensor information fusion can simultaneously determine three states of the net, that is, the normal state, the damaged state and the biological attachment state. When the net is damaged, an alarm can be sent in real time, and the staff can repair it. When the net has biological attachment, the staff can clean the net according to the amount of biological attachment to avoid tearing of the net caused by biological attachment. BRIEF DESCRIPTION OF DRAWINGS

[0088] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced. Obviously, the accompanying drawings in the following description only represent some embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on the accompanying drawings can also fall within the scope of the present application.

[0089] Figure 1 Flow chart of the method of the present application.

[0090] Figure 2 Structure diagram of the numerical net of the present application.

[0091] Figure 3 Arrangement diagram of the sensor of the present application.

[0092] Figure 4 Improved VGG16 network architecture diagram of the present application.

[0093] Figure 5 Waveform diagram of the simulation acceleration data at the L-1 position of the net cage under the wave height of 0.2m and the wave period of 1s in the present application.

[0094] Figure 6 Waveform diagram of the simulation tension data at the L-1 position of the net cage under the wave height of 0.2m and the wave period of 1s in the present application.

[0095] Figure 7 Waveform diagram of the simulation tension data at the L-1 position of the net cage under the wave height of 0.2m and the wave period of 1s in the present application. DETAILED DESCRIPTION

[0096] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort fall within the scope of the present application.

[0097] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only used to facilitate the description of the present application and simplify the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application.

[0098] In the present application, unless otherwise explicitly specified and limited, the first feature is "on" or "under" the second feature can include that the first and second features are in direct contact, or can include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, the first feature "on", "above" and "over" the second feature includes that the first feature is directly above and obliquely above the second feature, or only indicates that the first feature is higher in horizontal height than the second feature. The first feature "under", "below" and "under" the second feature includes that the first feature is directly below and obliquely below the second feature, or only indicates that the first feature is lower in horizontal height than the second feature.

[0099] As shown in Figure 1 , the aquaculture net cage damage monitoring method based on multi-sensor information fusion comprises the following steps:

[0100] Step 1: Establish an equivalent numerical model of the net cage according to the size of the gravity deep water net cage; establish the foundation including the net line, net joint, float and sinker; the step 1 is specifically: use the equivalent numerical model to represent the deep sea aquaculture net cage in the Orcaflex marine engineering dynamic simulation tool; use the "Line" model to represent the net line, the "3Dbouy" model to represent the net joint, and the "6Dbouy" model to represent the float; the model bottom is simulated by 16 sinkers with weight 3D buoys, and is symmetrically distributed around the bottom of the net cage; the net cage model contains a total of 64 "6Dbouy" models, 1280 "3Dbouy" models and 2656 "Line"; the whole net cage system is composed of mooring system, float system and fishing net system. As shown in Figure 2 .

[0101] Step 2, as shown in Figure 3 , the net cage is evenly divided into upper, middle and lower three layers, and four points are evenly distributed in each layer, a total of 12 points, one tension sensor and one acceleration sensor are defined as a group of sensors, and a group of sensors are installed on the 12 points on the net cage to collect the tension signal of the net cover in real time, that is, evidence A, and the acceleration signal, that is, evidence B; the 12 points divide the net cover into 8 areas, namely areas 1 to 8; the step 2 is specifically: the point positions measured by the combination of 12 tension sensors and acceleration sensors are marked as points 1 to 12, and areas 1 to 8 are respectively:

[0102] Area 1: points 1, 2, 5, 6; Area 2: points 2, 3, 6, 7; Area 3: points 3, 4, 7, 8; Area 4: points 1, 4, 5, 8; Area 5: points 5, 6, 9, 10; Area 6: points 6, 7, 10, 11; Area 7: points 7, 8, 11, 12; Area 8: points 5, 8, 9, 12. Table 1 is the parameters of the numerical net cage.

[0103] Table 1 comparison between net cage numerical model and actual model

[0104]

[0105]

[0106] Step 3, according to the environment setting different irregular wave parameters of net cage, under different wave parameters, using the equivalent numerical model of net cage to simulate the tensile force and acceleration signal of intact and damaged netting conditions, obtain the simulation data set, and generate the simulation data waveform according to the data set;

[0107] The step 3 is specifically: the different sea wave parameters include wave height of 0.2m, 0.6m, 1m, 1.4m, 1.6m, and wave period of 1s, 2s, 3s, 4s, 5s, which are combined in pairs. The simulation data waveform at the damaged L-1 position of the net cage under the wave height of 0.2m and the wave period of 1s is as shown in Figure 5 and 6 , and Figure 7 is the simulation data waveform under the state of foreign matter attachment;

[0108] Step 4, as shown in Figure 4 , a neural network damage monitoring model is established based on CNN neural network, and the simulation data set obtained is used to train the neural network damage monitoring model; wherein the simulation tensile force, acceleration data and wave parameters of the netting are used as input, and the probability of each sensor identifying the damage of the netting is used as output;

[0109] In the step 4, the establishment method of the neural network damage monitoring model includes:

[0110] Based on the simplification and variation of VGG16 network architecture, the characteristics of the hydrodynamic data are extracted by using continuous wavelet transform CWT, and the extracted characteristics are provided as input to the network for further analysis;

[0111] The network contains 12 convolutional layers, each convolutional layer uses a 3x3 convolutional kernel, and adopts a "same" padding method to ensure that the size of the feature map remains unchanged;

[0112] The number of filters of the convolutional layer gradually increases from 64 to 128, 256, 512 and 4096 to extract more complex features; after each convolutional layer, a ReLU activation function is used to enhance the nonlinearity of the network, and a batch normalization layer is used to speed up the training process;

[0113] The pooling layer adopts the maximum pooling operation, and the step is fixed as 2 in each layer to reduce the spatial dimension of the feature map and reduce the calculation amount; the network finally classifies through the fully connected layer, and uses the Softmax function to output the probability of each class;

[0114] The number of convolutional and pooling layers has been simplified as follows: in the first convolutional module, the original two convolutional layers are reduced to one; in the second convolutional module, two convolutional layers are retained, and the number of channels is reduced from the original 128 to 64; in the third convolutional module, the original three convolutional layers are reduced to two, and the number of channels is reduced from 256 to 128; in the fourth module, the number of channels is reduced from 512 to 256, while three convolutional layers are retained; the fifth module remains unchanged, containing three convolutional layers and 512 channels; the fully connected portion uses a single convolutional layer with a channel replacing the original first fully connected layer, and two subsequent fully connected layers are used for feature compression and classification output. The overall structure reduces the number of parameters while retaining the deep feature extraction capabilities of VGG16, making it more suitable for the computational requirements of net damage detection tasks. Through this design, the network can extract effective features from hydrodynamic data, identify the potential location and extent of net damage, and provide an efficient and real-time detection solution. The obtained simulation data set is used to train the artificial neural network damage monitoring model. The simulated tension, acceleration data, and wave parameters of the net are used as input, and the probability of each sensor identifying the net damage is used as output.

[0115] Step 5: Obtain the basic probability assignment m for each group of sensors through DS evidence theory A and m B ; The details are as follows:

[0116] Step 5.1, build the recognition framework, which includes three states of the net cage: normal state of the net, damaged state of the net, and state of foreign matter attached to the net. The basic propositions of the recognition framework are defined as normal state J, damaged state H, and foreign matter attached state I; the three states

[0117] Step 5.2: Based on the data obtained from the tension sensor and acceleration sensor, the neural network damage monitoring model is used to obtain the preliminary diagnosis results, and the preliminary diagnosis results are assigned to the BPA basic probability assignment function, i.e., m A and m B ;

[0118] Step 6: Use Dempster's synthesis rule to assign the BPA basic probability of any group of two sensors to m A and m B Combined into new evidence C, the synthesis rules are as follows:

[0119]

[0120] Where k is the conflict coefficient, which indicates the degree of conflict between the evidence given by the two sensors; m C is m A and m BThe fused basic probability assignment function is combined with the position of the tension sensor and the acceleration sensor on the net cage to obtain m C i, where i is the position of the tension sensor and the acceleration sensor on the net cage, and takes values of 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, and 12; the 12 m C i obtained are fused in the region to obtain 8 final BPA basic probability assignments, and the state of the netting is judged according to the basic probability assignment m 区域n of the final comprehensive evidence C, where n represents the region number and takes values of 1 to 8.

[0121] The step 5.2 is specifically:

[0122] For the evidence A and B of the tension sensor and the acceleration sensor at any position on the net cage, the basic probability assignments are m A and m B , the tension sensor monitoring result is:

[0123] m A ({normal state})=a1

[0124] m A ({damaged state})=a2

[0125] m A ({foreign matter adhesion state})=a3

[0126] Wherein, a1 represents the probability of the normal state of the position point, taking values of 0 to 1, a2 represents the probability of the damaged state of the position point, taking values of 0 to 1, and a3 represents the probability of the foreign matter adhesion state of the position point, taking values of 0 to 1.

[0127] The acceleration sensor monitoring result is:

[0128] m B ({normal state})=b1

[0129] m B ({damaged state})=b2

[0130] m B ({foreign matter adhesion state})=b3

[0131] Wherein, b1 represents the probability of the normal state of the position point, taking values of 0 to 1, b2 represents the probability of the damaged state of the position point, taking values of 0 to 1, and b3 represents the probability of the foreign matter adhesion state of the position point, taking values of 0 to 1.

[0132] In the step 6,

[0133] The conflict coefficient K represents the conflict degree between the evidence given by the two sensors, and the calculation method is:

[0134]

[0135] The probability product of each team disjoint hypothetical combination is calculated and summed up:

[0136] K = m A ({normal state}) x m B ({damaged state}) + m A ({normal state}) x m B ({foreign matter attached state}) + m A ({damaged state}) x m B ({normal state}) + m A ({damaged state}) x m B ({foreign matter attached state}) + m A ({foreign matter attached state}) x m B ({normal state}) + m A ({foreign matter attached state}) x m B ({damaged state})

[0137] Then:

[0138] K = (a1 x b2) + (a1 x b3) + (a2 x b1) + (a2 x b3) + (a3 x b1) + (a3 x b2)

[0139] The BPA basic probability assignment m after synthesis is calculated C ,

[0140] For normal state:

[0141]

[0142] For damaged state:

[0143]

[0144] For foreign matter attached state:

[0145]

[0146] Wherein, c1 represents the probability of normal state of the position point after synthesis, taking 0 to 1, c2 represents the probability of damaged state of the position point after synthesis, taking 0 to 1, and c3 represents the probability of foreign matter attached state of the position point after synthesis, taking 0 to 1.

[0147] In this embodiment, the fusion in the region in step 6 specifically includes:

[0148] First, the detection results of the four positions in region 1 are obtained from the previous steps, as shown in the table.

[0149]

[0150]

[0151] The fusion results of the four point tensile sensors and acceleration sensors in region 1 are preliminarily fused, and the fusion result is:

[0152] m C 1,m C 2,m C 5,m C 6

[0153] Wherein: the BPA basic probability assignment m C of any point contains three states:

[0154] m i ({normal state}), m i ({damage state}), m i ({foreign matter attachment state})

[0155] The two BPA basic probability assignments of point 1 and point 2 are set as m1 and m2, and are combined into m 12 , which is defined as follows:

[0156] K=m A ({normal state})×m B ({damage state})+m A ({normal state})×m B ({foreign matter attachment state})+m A ({damage state})×m B ({normal state})+m A ({damage state})×m B ({foreign matter attachment state})+m A ({foreign matter attachment state})×m B ({normal state})+m A ({foreign matter attachment state})×m B ({damage state})

[0157] According to the monitoring results a1, a2, a3, b1, b2, b3 of the sensors at the four positions in region 1, k is calculated according to the formula in step 7, and the synthesized BPA basic probability assignment m 12 is further calculated.

[0158] K=(a1×b2)+(a1×b3)+(a2×b1)+(a2×b3)+(a3×b1)+(a3×b2)

[0159] = (0.6 x 0.2) + (0.6 x 0.1) + (0.3 x 0.7) + (0.3 x 0.1) + (0.1 x 0.7) + (0.1 x 0.2)

[0160] = 0.51

[0161] For normal state:

[0162]

[0163] For damaged state:

[0164]

[0165] For foreign matter attached state:

[0166]

[0167] The synthesis of m1 and m2 is completed to obtain m 12 Then, the synthesis is continued as follows

[0168] m 12 ⊕ m5 = m 125 , m 125 ⊕ m6 = m 1256 ,

[0169] Finally, the BPA basic probability assignment of region 1 is obtained, that is:

[0170] m 区域1 = m 1256

[0171] From the table, m5 = {0.5, 0.3, 0.2}

[0172] Then:

[0173] K = 0.8571 x 0.3 + 0.8571 x 0.2 + 0.1224 x 0.5 + 0.1224 x 0.2 + 0.0204 x 0.5 + 0.0204 x 0.3 = 0.5305

[0174] The BPA basic probability assignment after synthesis is calculated.

[0175] For normal state:

[0176]

[0177] For damaged state:

[0178]

[0179] For foreign matter attached state:

[0180]

[0181] From the table, m6={0.4, 0.5, 0.1}

[0182] Then:

[0183] K=0.9125*0.5+0.9125*0.1+0.0782*0.4+0.0782*0.1+0.0087*0.4+0.0087

[0184] *0.5=0.5946

[0185] The BPA basic probability assignment of the synthesized BPA is calculated.

[0186] For the normal state:

[0187]

[0188] For the damaged state:

[0189]

[0190] For the foreign matter adhesion state:

[0191]

[0192] The fusion result of the final region 1 is:

[0193]

[0194] Wherein m({normal state}) represents the probability of the normal state of region 1 on the net, taking 0 to 1; m({damaged state}) represents the probability of the damaged state of region 1 on the net, taking 0 to 1; m({foreign matter adhesion state}) represents the probability of the foreign matter adhesion state of region 1 on the net, taking 0 to 1.

[0195] The BPA basic probability assignments m of regions 2 to 8 区域2 , m 区域3 , m 区域4 , m 区域5 , m 区域6 , m 区域7 , m 区域8 are obtained according to the above synthesis rule.

[0196] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between various embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0197] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for monitoring damage of aquaculture cages based on multi-sensor information fusion, characterized in that: The following steps are involved: Step 1: Establish an equivalent numerical model of the gravity deepwater cage according to its dimensions; establish the foundation including the net line, net knot, float and sinker; Step 2: Evenly divide the net cage into three layers: upper, middle, and lower. Evenly distribute four points on each layer, for a total of 12 points. Define a sensor set consisting of one tension sensor and one acceleration sensor. Install these sensors at the 12 points on the net cage to collect the net's tension signal (Evidence A) and acceleration signal (Evidence B) in real time. The 12 points divide the net into eight regions, namely, Regions 1 to 8. Step 3: Set different irregular wave parameters according to the environment where the cage is located. Under different wave parameters, use the equivalent numerical model of the aquaculture cage to simulate the tension and acceleration signals of the net under intact and damaged conditions to obtain a simulation data set, and generate a simulation data waveform based on the data set; Step 4: Establish a neural network damage monitoring model based on the CNN neural network and train the neural network damage monitoring model using the obtained simulation data set. The simulated tension, acceleration data, and wave parameters of the net are used as inputs, and the probability of each sensor identifying the net damage is used as output. Step 5: Obtain the basic probability assignment m of each group of sensors through the neural network damage monitoring model A and m B ; The details are as follows: Step 5.1, constructing a recognition framework, which includes three states of the net cage: normal state of the net, damaged state of the net, and state of foreign matter attached to the net. The basic propositions of the recognition framework are defined as normal state J, damaged state H, and foreign matter attached state I; Step 5.2: Based on the data obtained from the tension sensor and acceleration sensor, the neural network damage monitoring model is used to obtain the preliminary diagnosis results, and the preliminary diagnosis results are assigned to the BPA basic probability assignment function, i.e., m A and m B ; Step 6: Use Dempster's synthesis rule to assign the BPA basic probability of any group of two sensors to m A and m B Combined into new evidence C, the synthesis rules are as follows: Where k is the conflict coefficient, which indicates the degree of conflict between the evidence given by the two sensors; m C is m A and m B The fused basic probability assignment function is combined with the position of the tension sensor and acceleration sensor on the cage to obtain m C i, where i is the position on the cage where the tension sensor and acceleration sensor are located, and the values ​​are 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12; According to the divided mesh area, the 12 m C i performs fusion within the region and obtains 8 final BPA basic probability assignments. According to the basic probability assignment m of the final comprehensive evidence C 区域n , judge the state of the net, where n represents the area number, ranging from 1 to 8.

2. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 1, characterized in that: Step 1 specifically includes: using an equivalent numerical model in the Orcaflex marine engineering dynamic simulation tool to represent the deep-sea aquaculture cage; using a "Line" model to represent the net line, a "3Dbouy" model to represent the net knot, and a "6Dbouy" model to represent the buoy. At the bottom of the model, 16 weighted 3D buoys are used to simulate 16 sinkers, and are symmetrically distributed around the bottom of the cage. The cage model contains a total of 64 "6Dbouy" models, 1,280 "3Dbouy" models, and 2,656 "Line" models. The entire cage system consists of a mooring system, a floating frame system, and a fishing net system.

3. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 1, characterized in that: The step 2 is specifically as follows: the points measured by the combination of 12 tension sensors and acceleration sensors are marked as points 1 to 12, and areas 1 to 8 are respectively: Area 1: points 1, 2, 5, 6; Area 2: points 2, 3, 6, 7; Area 3: points 3, 4, 7, 8; Area 4: points 1, 4, 5, 8; Area 5: points 5, 6, 9, 10; Area 6: points 6, 7, 10, 11; Area 7: points 7, 8, 11, 12; Area 8: points 5, 8, 9, 12.

4. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 1, characterized in that: The step 3 specifically includes: the different wave parameters include a wave height of 0.2m to 1.6m and a wave period of 1s to 5s.

5. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 1, characterized in that: In step 4, the method for establishing the neural network damage monitoring model includes: Based on the simplification and variation of the VGG16 network architecture, the continuous wavelet transform (CWT) is used to extract features from the hydrodynamic data, and the extracted features are provided as input to the network for further analysis; The network contains 12 convolutional layers, each of which uses a 3x3 convolution kernel and adopts the "same" padding method to ensure that the size of the feature map remains unchanged; The number of filters in the convolutional layer is gradually increased from 64 to 128, 256, 512, and 4096 to extract more complex features. After each convolutional layer, the ReLU activation function is used to enhance the nonlinearity of the network, and the batch normalization layer is used to accelerate the training process. The pooling layer uses the maximum pooling operation with a fixed step size of 2 per layer to reduce the spatial dimension of the feature map and reduce the amount of computation. The network is finally classified through a fully connected layer, and the Softmax function is used to output the probability of each category. The number of convolutional layers and pooling layers has been simplified as follows: in the first convolutional module, the original two convolutional layers are reduced to one; in the second convolutional module, two convolutional layers are retained, and the number of channels is reduced from the original 128 to 64; in the third convolutional module, the original three convolutional layers are reduced to two, and the number of channels is reduced from 256 to 128; in the fourth module, the number of channels is reduced from 512 to 256, and three convolutional layers are retained; the fifth module remains unchanged, containing three convolutional layers and 512 channels; the fully connected part uses one convolutional layer, and the channel replaces the original first fully connected layer, and two subsequent fully connected layers are set for feature compression and classification output; the overall structure retains the deep feature extraction capability of VGG16 while reducing the number of parameters.

6. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 1, characterized in that: The step 5.2 is specifically as follows: For evidence A and B of the tension sensor and acceleration sensor at any position on the cage, the basic probability is assigned as m A and m B , the monitoring results of the tension sensor are: m A ({normal state}) = a1 m A ({damaged state}) = a2 m A ({Foreign matter adhesion state}) = a3 Where a1 represents the probability of the position point being in a normal state, ranging from 0 to 1; a2 represents the probability of the position point being in a damaged state, ranging from 0 to 1; a3 represents the probability of the position point being in a foreign body attachment state, ranging from 0 to 1; Acceleration sensor monitoring results: m B ({normal state}) = b1 m B ({Damaged state}) = b2 m B ({Foreign matter adhesion status}) = b3 Among them, b1 represents the probability of the normal state of the position point, which is between 0 and 1; b2 represents the probability of the damaged state of the position point, which is between 0 and 1; b3 represents the probability of the foreign matter attached state of the position point, which is between 0 and 1.

7. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 1, characterized in that: In step 6, The conflict coefficient K represents the degree of conflict between the evidence given by two sensors and is calculated as: Calculate the product of the probabilities of each set of disjoint hypothetical combinations and sum them: K=m A ({normal state})×m B ({Damaged State})+m A ({normal state})×m B ({foreign matter attachment status})+m A ({Damaged state})×m B ({normal state})+m A ({Damaged state})×m B ({foreign matter attachment status})+m A ({Foreign matter attachment status})×m B ({normal state})+m A ({Foreign matter attachment status})×m B ({Broken State}) but: K=(a1×b2)+(a1×b3)+(a2×b1)+(a2×b3)+(a3×b1)+(a3×b2) Calculate the synthesized BPA basic probability assignment m C , For normal status: For the damaged state: For foreign matter attachment: Among them, c1 represents the probability of the normal state of the position point after synthesis, ranging from 0 to 1, c2 represents the probability of the damaged state of the position point after synthesis, ranging from 0 to 1, and c3 represents the probability of the foreign matter attached state of the position point after synthesis, ranging from 0 to 1.

8. The method for monitoring damage of aquaculture cages based on multi-sensor information fusion according to claim 6, characterized in that: In step 6, the integration within the region specifically includes: The fusion results of the tension sensor and acceleration sensor at the four points in area 1 are preliminarily fused, and the fusion results are: m C 1,m C 2,m C 5,m C 6 in: m i ({Normal})=a i ,m i ({broken}) = b i ,m i ({foreign matter attached}) = c i , Assign the two BPA basic probabilities of point 1 and point 2 to m1 and m2, and combine them into m 12 , defined as follows: K=m A ({normal state})×m B ({Damaged State})+m A ({normal state})×m B ({foreign matter attachment status})+m A ({Damaged state})×m B ({normal state})+m A ({Damaged state})×m B ({foreign matter attachment status})+m A ({Foreign matter attachment status})×m B ({normal state})+m A ({Foreign matter attachment status})×m B ({Broken State}) Based on the a1, a2, a3, b1, b2, and b3 obtained from the sensor monitoring results at the four locations in area 1, k is calculated according to the formula in step 7, and the synthesized BPA basic probability assignment m is further calculated. 12 ; For normal status: For the damaged state: For foreign matter attachment: Complete the synthesis of m1 and m2 to obtain m 12 Then, continue with the following synthesis Finally, the BPA basic probability assignment of area 1 is obtained, namely: m 区域1 =m 1256 The final fusion result of region 1 is: Where m({normal state}) represents the probability that area 1 on the net is normal, ranging from 0 to 1; m({damaged state}) represents the probability that area 1 on the net is damaged, ranging from 0 to 1; m({foreign matter attached state}) represents the probability that area 1 on the net is attached with foreign matter, ranging from 0 to 1; BPA basic probability assignment m for areas 2 to 8 区域2 、m 区域3 、m 区域4 、m 区域5 、m 区域6 、m 区域7 、m 区域8 All are obtained according to the above synthesis rules.

Images