Intelligent fish feeding method and system based on energy model and behavior feedback

Abstract

The application discloses an intelligent fish feeding method and system based on an energy model and behavior feedback, which comprises the following steps: obtaining the required feed quality to be provided on the current day based on a biological energy model and actual breeding conditions; constructing a graph G representing overall motion characteristics based on a real-time video picture of fish feeding captured by shooting; inputting the graph G into a GCN fish appetite grading model to divide the feeding desire of the fish into two levels, i.e., having feeding desire and not having feeding desire; and performing a feeding decision process, wherein half of the feeding amount is fed according to the required feed quality to be provided on the current day, and if the fish school shows feeding desire, the system continues to feed half of the remaining feed, and the feeding is cyclically performed in this way until the fish school no longer shows feeding desire. The method and system can avoid the uncertainty of the feeding amount caused by human experience and subjective factors, and can improve breeding benefits and reduce production costs.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of recirculating aquaculture, and relates to a solution for fish feeding demand and feeding decision-making in recirculating aquaculture, in particular to an intelligent fish feeding method and system based on energy model and behavior feedback. BACKGROUND

[0002] Recirculating aquaculture is a high-efficiency water resource utilization mode, which has the advantages of saving water resources, reducing environmental pollution and improving production efficiency. However, in the process of recirculating aquaculture, feeding decision-making has always been a challenging problem. The traditional feeding decision-making method relying on human experience has the problems of uncertainty and subjectivity, which cannot meet the requirements of accurately controlling the feeding demand of farmed fish. Therefore, accurately calculating the energy demand of fish and precisely controlling the feeding amount are crucial for aquaculture management.

[0003] Based on the above background, the present application aims to provide an intelligent feeding decision-making system for recirculating aquaculture fish based on energy model and fish behavior feedback. By proposing a new fish energy model, combined with an improved graph convolutional neural network model of kinetic energy model, the system can accurately calculate the energy demand of fish by combining water temperature, dissolved oxygen and fish weight in the aquaculture water body. Through collecting video data of fish feeding behavior, the system can accurately classify the feeding desire according to the feeding behavior characteristics of fish, and then obtain accurate feedback of feeding demand and realize precise feeding decision-making. Thus, while meeting the energy supply of farmed fish, the production efficiency of recirculating aquaculture is improved.

[0004] Compared with the traditional feeding decision-making method relying on human experience, the system has the advantages of accurate energy calculation and intelligent analysis based on deep learning. The energy demand of fish can be calculated more accurately, avoiding the uncertainty of feeding amount caused by human experience and subjective factors. At the same time, the deep learning algorithm based on graph convolutional neural network model can extract rich spatial features from limited time data, realizing accurate classification of fish feeding desire. This makes the feeding amount can be precisely controlled according to the real-time feeding demand of fish, improving the breeding efficiency and reducing the production cost.

[0005] In summary, the present application provides an intelligent feeding decision-making system for recirculating aquaculture fish based on energy model and fish behavior feedback. By combining biological energy model and machine vision and deep learning technology, the system can accurately calculate the energy demand of fish, and use graph convolutional neural network model to realize accurate classification of feeding desire. This will provide a more scientific and intelligent feeding decision-making solution for the recirculating aquaculture industry, improve production efficiency and realize sustainable development. SUMMARY

[0006] The present application aims to provide a fish intelligent feeding method and system based on energy model and behavior feedback, which can calculate the energy required for fish growth according to the water temperature of the breeding water body and the fish weight information, and complete the decision of the feeding amount through the feeding behavior feedback of the fish, thereby providing good technical support for the intelligent feeding operation of the recirculating aquaculture.

[0007] The technical scheme adopted by the present application is as follows:

[0008] In the first aspect, the present application provides a fish intelligent feeding method based on energy model and behavior feedback, which comprises the following steps:

[0009] Based on the biological energy model, the maximum energy required for feeding is determined according to the actual water temperature of the breeding water body, the water saturation dissolved oxygen content under the temperature, the actual water dissolved oxygen content and the fish body weight, so as to obtain the required feeding quality provided on the same day;

[0010] The motion features between the continuous video frames are extracted and quantified based on the Gunnar Farneback dense optical flow algorithm for the real-time video pictures of the fish feeding, and a graph G representing the overall motion features is constructed;

[0011] The graph G is input into the GCN fish appetite grading model, and the feeding desire of the fish is divided into two levels of having feeding desire and not having feeding desire; the feeding decision process is as follows: according to the required feeding quality provided on the same day obtained in the foregoing, half of the feeding amount is fed, and if the fish group shows feeding desire, the system continues to feed half of the remaining feed, and the process is repeated in this way until the fish group no longer shows feeding desire.

[0012] In the above technical scheme, further, the constructed energy biology model is:

[0013] T≤Topt,

[0014]

[0015] T> Topt

[0016]

[0017] Wherein, Q q is the maximum energy required for feeding on the same day, a is the dissolved oxygen coefficient, h1=0.077, h2=0.076, m=0.708, Topt is the optimum growth temperature of fish, Tmax is the maximum feeding temperature of fish, T is the actual water temperature of the breeding water body, and w is the mass of the fish group on the same day, unit: g.

[0018] Further, the features of the real-time video of the fish feeding are extracted, which specifically includes:

[0019] Quantifying the motion characteristics between consecutive video frames, using Gunnar Farneback dense optical flow algorithm to obtain the optical flow value F between two consecutive frames of images, normalizing the optical flow amplitude F, using (x, y) to represent the coordinates of the target pixel point in the current frame, N represents the total number of motion vectors in the current video frame, that is, the total number of target pixel points, and the average change speed v of the target region is calculated E , Divide 0-tl / s into 25 intervals, where tl is the initial measured average length of the micropterus salmoides, s is the second, calculate the number of motion vectors n(j) in the target region falling into the jth speed interval, and obtain the motion vector probability m(j) falling into the specified interval, The irregularity C of the target region is calculated using the change characteristics of the target region E , Then calculate the spatial motion characteristics E K between video frames K = ln(T x a) x C E x v E 2 , where T is the actual water temperature of the aquaculture water body, and a is the dissolved oxygen coefficient of the aquaculture water body.

[0020] Further, the method for constructing the graph G is: according to the spatial motion characteristics E K between video frames, constructing an adjacency matrix A (A e R O×O ) and a feature matrix X (X e R O×F ); wherein the adjacency matrix A is used to represent the connection between video frames, that is, the time correlation, the adjacency matrix only contains two values 0 and 1, when there is a connection between two video frames, the value is 1, otherwise it is 0, the feature matrix X, the spatial motion characteristics E k and the time characteristics of the fish feeding behavior extracted from the video frames are used as the attribute characteristics of the nodes in the graph, indicating that X e R O×F , wherein F is the dimension of the node attribute characteristics, here F = 2, which are the time characteristics and the spatial motion characteristics, respectively, and O represents the number of frames; by combining the adjacency matrix A and the feature matrix X, the graph G = (A, X) representing the overall motion characteristics is constructed.

[0021] Further, the GCN fish appetite grading model is constructed, and the model structure includes graph convolution layer, pooling layer, readout layer, MLP layer and softmax layer, which includes 7 graph convolution layers, the output of each layer is connected in series, the node and graph features are updated and aggregated in the pooling layer through the self-attention mechanism, then transmitted to the MLP layer through the readout layer, and finally the fish feeding desire level is determined by using the softmax layer.

[0022] In a second aspect, the embodiments of the present application provide a fish intelligent feeding system based on an energy model and behavior feedback, which can implement any of the above methods.

[0023] The system comprises a circulating water treatment system, a variable frequency water pump, a circulating water aquaculture pond, a computer, a PLC controller, a high-definition waterproof camera, a feeder, a dissolved oxygen sensor, a temperature sensor and an LED light supplementing lamp.

[0024] The feeder is installed directly above the circulating water aquaculture pond, the discharge port is located at the side of the high-definition waterproof camera, the LED light supplementing lamp is arranged below the feeder, the feeder is connected with the output end of the PLC controller, the dissolved oxygen sensor and the temperature sensor are fixed inside the circulating water aquaculture pond, and the computer is connected with the PLC controller and the high-definition waterproof camera at the same time.

[0025] The present application has the following advantages:

[0026] The fish intelligent feeding decision method and system based on the energy model and the fish behavior feedback provided by the present application propose a fish energy intake model and a fish appetite grading model based on GCN, which can accurately calculate the daily food intake demand of the fish group, and a large number of experiments have proved the accuracy of the system. Meanwhile, the scheme has strong practicability, and thus becomes an important tool for fish feeding management, ensures that the fish obtains suitable feed supply, and avoids resource waste and water pollution caused by overfeeding. The GCN fish appetite grading model part involves analyzing the fish feeding video by using the graph convolution network (GCN) technology. Through feature extraction of the feeding video and fish behavior pattern recognition, the model can accurately judge the feeding desire state of the fish. This provides an important reference for intelligent feeding decision, and enables the system to accurately feed the fish according to the real-time appetite condition of the fish.

[0027] In summary, the present application can realize accurate and efficient feeding management, ensures that the fish obtains suitable feed supply, and maximally meets the feeding demand of the fish. While improving the feeding efficiency and economic benefits, the present application can also reduce resource waste and environmental pollution caused by overfeeding. The present application provides reliable decision support for breeders, and promotes the sustainable development of the industry. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a GCN fish appetite grading model structure diagram in the present application;

[0029] Figure 2 is a structure diagram of the fish intelligent feeding system based on the energy model and the behavior feedback in the present application.

[0030] In the figure: 1-circulating water treatment system, 2-variable frequency water pump, 3-circulating water aquaculture pond, 4-computer, 5-PLC controller, 6-high-definition waterproof camera, 7-feeding machine, 8-dissolved oxygen sensor, 9-temperature sensor, 10-LED light supplement. DETAILED DESCRIPTION

[0031] The application will be further described in detail below with reference to the accompanying drawings and specific examples.

[0032] Reference Figure 2 , is a specific example of the intelligent feeding decision system for fish in a circulating water aquaculture system based on an energy model and fish behavior feedback, including a circulating water treatment system 1, a variable frequency water pump 2, a circulating water aquaculture pond 3, a computer 4, a PLC controller 5, a high-definition waterproof camera 6, a feeding machine 7, a dissolved oxygen sensor 8, a temperature sensor 9, and an LED light supplement 10.

[0033] In this example, the feeding machine 7 is installed directly above the circulating water aquaculture pond 3, the discharge port is located on the side of the high-definition waterproof camera 6, the LED light supplement 10 is arranged below the feeding machine 7, and the feeding machine 7 is connected to the output end of the PLC controller 5. The dissolved oxygen sensor 8 and the temperature sensor 9 are fixed inside the circulating water aquaculture pond 3. The computer 4 is connected to the PLC controller 5 and the high-definition waterproof camera 6 at the same time.

[0034] The system can perform the following intelligent feeding method, including:

[0035] Based on the biological energy model, the actual water temperature, the water saturation dissolved oxygen content at this temperature, the actual water dissolved oxygen content, and the fish body weight, the maximum energy required for feeding on the same day is determined, and the mass of feed required to be provided on the same day is obtained;

[0036] Based on the Gunnar Farneback dense optical flow algorithm, the motion features between consecutive video frames are extracted and quantified, and a graph G representing the overall motion features is constructed.

[0037] The graph G is input into the GCN fish appetite classification model, which classifies the fish feeding desire into two levels: feeding desire and no feeding desire. The feeding decision process is as follows: according to the aforementioned required feed mass to be provided on the same day, half of the feeding amount is fed, and the appetite classification model is used to determine whether the fish show feeding desire. If the fish show feeding desire, the system continues to feed half of the remaining feed, and the process is repeated in this way until the fish no longer show feeding desire.

[0038] Taking the California bass in a circulating water aquaculture system as an example, specifically:

[0039] Biological energy model:

[0040] (1) Establish the bioenergetics model of Micropterus salmoides

[0041] T≤Topt,

[0042]

[0043] T> Topt

[0044]

[0045] wherein h1, h2, m are obtained by fitting, h1 = 0.077, h2 = 0.076, m = 0.708, Topt is the optimum temperature for the growth of Micropterus salmoides, unit: ℃, Tmax is the maximum feeding temperature of Micropterus salmoides, unit: ℃ (above which Micropterus salmoides does not feed), T is the actual water temperature of the water body, unit: ℃, w is the mass of the fish population on that day, unit: g.

[0046] (2) Collect field data at the Micropterus salmoides breeding environment points

[0047] Use temperature sensor 9 to collect the actual water temperature T of the water body, and use dissolved oxygen sensor 8 to collect the dissolved oxygen O of the water body a , and weigh the fish population w. Calculate the dissolved oxygen coefficient a:

[0048]

[0049] wherein O a is the actual dissolved oxygen concentration in the water body collected by the dissolved oxygen sensor, unit: mg / L b is the saturated dissolved oxygen at that water temperature, unit: mg / L

[0050] (3) According to the data obtained in step (2), substitute into the bioenergetics model of Micropterus salmoides established in step (1) to calculate the maximum feeding energy Q q required on that day.

[0051] (4) The maximum feeding energy Q q = the energy Q e required to be provided.

[0052] (5) Calculate the feed quality required to be provided on that day, the feed energy value is N, unit: KJ / g, the feed quality required to be provided is M, unit: g, and the calculation method of M is as follows:

[0053]

[0054] Using the above bioenergetics model, different energy requirements are calculated under different temperature and fish body weight conditions. Taking the fry stage as an example, the specific data are shown in Table 1.

[0055] Table 1: Bioenergetics model calculation table in the field experiment of Micropterus salmoides

[0056]

[0057] Feeding behavior feedback

[0058] (1) Feeding behavior feature extraction

[0059] The high-definition waterproof camera transmits the real-time video of the largemouth bass feeding to the computer, and the computer uses an improved kinetic model based on the Gunnar Farneback dense optical flow algorithm to extract and quantify the motion features between consecutive video frames. The model aims to represent the overall feeding activity of the fish group. The specific calculation method is as follows:

[0060] E k = ln(T x a) x C E x v E 2

[0061] Where C E and v E are the irregularity and average speed of the target area change, respectively, to reduce the influence of fish length change and difference in the test process on motion feature extraction, the normalized optical flow value is used to calculate the average change speed v E of the target area, E k is the spatial motion feature between video frames, representing the kinetic energy of the target fish group, where the larger the kinetic energy value, the more intense the water fluctuation, indicating that the fish group's feeding desire is stronger.

[0062] The following is the algorithm for feature extraction and graph construction

[0063] The whole algorithm includes two parts, the first stage is feature extraction, which processes the video frames collected by the high-definition waterproof camera 6, specifically: quantifying the motion features between consecutive video frames, using the Gunnar Farneback dense optical flow algorithm to get the optical flow value F between two consecutive frames, where (x, y) represents the coordinates of the target pixel point in the current frame, N represents the number of moving pixel points in the current video frame, i.e. the total number of target pixel points, and the average change speed v E of the target area is calculated. Since 1 second contains 25 frames in a video image, 0-tl / s is divided into 25 intervals, where tl is the initial measured average length of the largemouth bass, s is the second, the number of motion vectors n(j) falling into the jth speed interval in the target area is calculated, and the motion vector probability m(j) falling into the specified interval is obtained. The irregularity C E of the target area is calculated, and then E K is calculated, E K = ln(T x a) x CE xv E 2 , i.e. E K .

[0064] The second stage is graph construction, constructing an adjacency matrix A (A ∈ R K ) and a feature matrix X (X ∈ R o×O ) according to the video frames and the spatial motion features E O×F ; wherein the adjacency matrix A is used to represent the connection between video frames, i.e. the time correlation, the adjacency matrix only contains two values of 0 and 1, when there is a connection between two video frames, the value is 1, otherwise it is 0, the feature matrix X, the spatiotemporal features (including spatial motion features E k and temporal features) of the fish feeding behavior extracted from the video frames are used as the attribute features of the nodes in the graph, indicating that X ∈ R O×F , wherein F is the dimension of the node attribute feature, here F = 2, which is the time feature and the spatial motion feature respectively; O is the number of frames; by combining the adjacency matrix A and the feature matrix X, a graph G = (A, X) representing the overall motion features is constructed.

[0065] For example, a video composed of 5 frames, construct an adjacency matrix A, A ∈ R 5×5 , the matrix is composed of 0 and 1, there is no connection between the first frame and the fourth frame, i.e. 0; there is a connection between the first frame and the second frame, i.e. 1, thus the adjacency matrix is constructed. Then construct the feature matrix X, X ∈ R 5×2 , the matrix is composed of E k representing the spatial motion features and the values 0 and 1 representing the temporal features.

[0066] (2) Feeding desire evaluation

[0067] In this study, the feeding desire of the California bass is divided into two types: 1. "No feeding desire" (No), the fish group has no reaction to the scattered feed, the water surface is close to static state, and the fed feed can be observed to still remain on the water surface after feeding; 2. "Feeding desire" (Yes), the fish group has a reaction to the scattered feed, and there is a feeding behavior of the California bass.

[0068] (3) Feeding desire representation model design

[0069] Construct a fish appetite classification model based on GCN (such as Figure 1According to the graph G constructed in step (1), accurate fish feeding desire classification can be achieved using limited data. The model structure includes a graph convolution layer, a pooling layer, a readout layer, an MLP layer, and a softmax layer: including 7 graph convolution layers, the output of each layer is concatenated, the node and graph features are updated and aggregated in the pooling layer through the self-attention mechanism, then transmitted to the MLP layer through the readout layer, and finally the feeding desire level of the fish school is determined by the softmax layer.

[0070] The propagation rule of GCN can be summarized as:

[0071]

[0072] wherein hl represents the node of the lth layer; represents the convolution weight with input feature dimension F and output feature F'; is a matrix added with a self-attention mechanism; I N represents the unit matrix; and σ is an activation function, which can adopt a rectified linear unit (ReLU) as the activation function.

[0073] The self-attention score a is calculated as follows:

[0074]

[0075] wherein, represents that the input feature of the graph has N nodes and F dimensions; is the parameter of the self-attention pooling layer.

[0076] In the classification model based on GCN, the readout layer is used to aggregate the features of the nodes to form a fixed-size representation, and the output feature of the readout layer is as follows:

[0077]

[0078] wherein N represents the number of nodes, xi represents the ith node of the feature vector, and symbol ∥ represents concatenation. i

[0079] ​​During the training process of the deep learning model, Adam optimization, early stopping and hyperparameter selection strategies can be used to train a model with high robustness. If the validation loss does not improve for 60 consecutive training rounds, the model terminates training, and the maximum training rounds are 100k. The self-defined GCN model is developed on Pytorch 1.10.0 (Python3.7), and the initial parameters and training strategy settings based on the GCN method are as follows: learning rate 5x10 -4 , batch size 128, dropout rate 0.5, pooling rate 0.8, number of hidden neurons 256, and decay weight 1x10 -4 .

[0080] The fish feeding desire level of the fish school is obtained by the GCN fish feeding grading model, and is divided into two categories: no feeding desire and feeding desire.

[0081] In actual farming process, by collecting water temperature, dissolved oxygen concentration and fish body weight and other parameters of the water body in the recirculating aquaculture system, the proposed energy model can accurately calculate the daily feeding demand of the fish school, and by analyzing the feeding video of the fish school, the proposed GCN fish feeding grading model can divide the feeding desire of the fish into two levels: feeding desire and no feeding desire.

[0082] The feeding decision process of the system is as follows: the energy model calculates the daily feeding amount, and then the feeding machine feeds half of the feeding amount each time. Then, if the fish school shows feeding desire, the system continues to feed the remaining half of the feed. In this way, the cycle is repeated until the fish school no longer shows feeding desire.

[0083] The present application can accurately calculate the feeding energy demand of fish by designing a fish feeding energy model combined with an improved graph convolutional neural network model of kinetic energy model, and can accurately classify the feeding desire of fish according to the feeding behavior characteristics of fish by collecting video data of fish feeding behavior, and then obtain accurate feedback of feeding demand and realize precise feeding decision. Thus, while meeting the energy supply of farmed fish, the production efficiency of recirculating aquaculture is improved. Compared with the traditional feeding decision relying on artificial experience, the present application can avoid the uncertainty of feeding amount caused by human experience and subjective factors. At the same time, the deep learning algorithm based on the graph convolutional neural network model can extract rich spatial features from limited time data, and realize accurate classification of fish feeding desire. This makes the feeding amount can be accurately controlled according to the real-time feeding demand of fish, improves the breeding efficiency and reduces the production cost.

Claims

1. A fish intelligent feeding method based on an energy model and behavior feedback, characterized in that, The method comprises the following steps: Based on the bioenergy model, the actual water temperature, the water saturation dissolved oxygen content at the temperature, the actual water dissolved oxygen content, and the fish body weight are combined to determine the maximum energy required for feeding on the day, thereby obtaining the required feed quality provided on the day; Based on the Gunnar Farneback dense optical flow algorithm, the motion features between consecutive video frames are extracted and quantified, and a graph G representing the overall motion features is constructed; wherein the feature extraction of the photographed real-time fish feeding video specifically includes: Quantifying the motion characteristics between continuous video frames, using Gunnar Farneback dense optical flow algorithm to get the optical flow value F between two continuous frames, normalizing the optical flow value F, using (x, y) to represent the coordinates of the target pixel point in the current frame, N represents the total number of motion vectors in the current video frame, that is, the total number of target pixel points, calculating the average change speed v of the target area E , , dividing 0~tl / s into 25 intervals, where tl is the initial measured average length of fish, s is second, calculating the number of motion vectors n(j) falling into the jth speed interval in the target area, getting the motion vector probability m(j) falling into the interval, , using the change characteristics of the target area to calculate its irregularity C E , , then calculating the spatial motion characteristics E K , of the video frame, where T is the actual water temperature of the aquaculture water body, and α is the dissolved oxygen coefficient of the aquaculture water body. The method for constructing the graph G is specifically: constructing an adjacency matrix A and a feature matrix X according to video frames and spatial motion features E K ; wherein the adjacency matrix A is used to represent the connection between video frames, that is, the time correlation, the adjacency matrix only contains two values of 0 and 1, when there is a connection between two video frames, the value is 1, otherwise, the value is 0, the feature matrix X uses the spatial motion features E k and the time features extracted from the video frames as the attribute features of the nodes in the graph, representing , wherein F is the dimension of the node attribute feature, here F=2, which is the time feature and the spatial motion feature respectively, and O represents the number of frames; by combining the adjacency matrix A and the feature matrix X, a graph G=(A,X) representing the overall motion features is constructed.​​ The graph G is input into the GCN fish appetite grading model, and the fish feeding desire is divided into two levels: feeding desire and no feeding desire; the feeding decision process is as follows: according to the required feed quality provided on the day obtained in the foregoing, half of the feeding amount is fed, and if the fish group shows feeding desire, the system continues to feed half of the remaining feed, and the process is repeated in this way until the fish group no longer shows feeding desire.

2. The intelligent fish feeding method based on energy model and behavior feedback according to claim 1, characterized in that, The constructed energy biology model is: When T≤Topt, ； When T> Topt ； where Q q is the maximum energy required for fish to consume on the day, a is the dissolved oxygen coefficient, h1=0.077, h2=0.076, m=0.708, Topt is the optimum growth temperature of fish, Tmax is the maximum feeding temperature of fish, T is the actual water temperature of the water body for aquaculture, and w is the mass of the fish population on the day, in units of g. 3.The fish intelligent feeding method based on the energy-based model and behavior feedback according to claim 1, wherein, The GCN fish appetite grading model is constructed, and the model structure includes a graph convolution layer, a pooling layer, a readout layer, an MLP layer, and a softmax layer, wherein 7 graph convolution layers are included, the outputs of each layer are connected in series, the node and graph features are updated and aggregated in the pooling layer through a self-attention mechanism, then transmitted to the MLP layer through the readout layer, and finally the fish feeding desire level is determined by using the softmax layer.

4. The intelligent fish feeding system based on energy model and behavior feedback, characterized in that, The system can realize the method of any one of claims 1-3.

5. The intelligent fish feeding system based on energy model and behavior feedback according to claim 4, characterized in that, The system comprises a circulating water treatment system (1), a variable frequency water pump (2), a circulating water aquaculture pond (3), a computer (4), a PLC controller (5), a high-definition waterproof camera (6), a feeding machine (7), a dissolved oxygen sensor (8), a temperature sensor (9), and an LED light supplementing lamp (10); The feeding machine (7) is installed directly above the circulating water aquaculture pond (3), the discharge port is located on the side of the high-definition waterproof camera (6), the LED light supplementing lamp (10) is arranged below the feeding machine (7), the feeding machine (7) is connected with the output end of the PLC controller (5), the dissolved oxygen sensor (8) and the temperature sensor (9) are fixed inside the circulating water aquaculture pond (3), and the computer (4) is connected with the PLC controller (5) and the high-definition waterproof camera (6).

Images