Fish health condition monitoring device and method based on fish behavior analysis

Abstract

The invention discloses a fish health condition monitoring device based on fish behavior analysis, which relates to the aquaculture technology and comprises a water surface unit, an underwater unit, an underwater unit and a data processing center. The water surface unit, the underwater unit and the underwater unit are respectively installed on the water surface, the underwater and the underwater of the culture pond, the water surface unit, the underwater unit and the underwater unit respectively collect real-time data of the water surface, the underwater and the underwater of the culture pond, the data processing center receives the real-time data and processes the real-time data through a Stacking stacking learning model, and the data processing center processes the real-time data through a data processing module. The health condition of the fish school in the culture pond and the abnormal position are judged, and an abnormal alarm signal is generated when it is judged that the fish school at a certain position in the area makes an abnormal behavior. The invention further discloses a fish health condition monitoring method based on fish behavior analysis. The system can achieve the real-time detection and early warning of the health condition of the fish school, and effectively improves the anti-risk capability and economic benefits in the breeding process.

Description

Technical Field

[0001] The present invention relates to aquaculture technology, and more particularly to a device and method for monitoring the health status of fish based on fish behavior analysis. Background Art

[0002] Aquaculture is an important part of modern agriculture, and its production efficiency is closely related to the level of aquaculture management. As an important part of aquaculture management, the health status of fish will directly determine the economic benefits of this round of aquaculture. Farmed fish usually have behaviors such as frequent rubbing against the wall, frequent floating, high-density aggregation, rapid swimming and tail flicking, and reduced food intake, which reflects the possible stress factors in the aquaculture environment: such as parasites, fungi, bacterial infections, water quality stress, external damage, etc.

[0003] In traditional aquaculture, the determination of fish health status mainly relies on the experience and observation of breeders, which lacks scientificity and accuracy. Especially in today's high-density breeding environment, if early warnings are not found in time at the early stage of disease and environmental stress, it may cause irreparable and huge losses to the entire breeding activities. With the development of Internet of Things and artificial intelligence technology, a variety of intelligent detection equipment has been gradually applied to aquaculture. By real-time detection and analysis of various behaviors of fish, the risk resistance and economic benefits of the breeding process are improved. At present, there are some devices on the market for detecting fish behavior, but such devices are usually based on image information. Although image information contains richer data features, it has some problems that are difficult to avoid, such as insufficient clarity of underwater images, the accuracy of image information for high-density breeding environments needs to be improved, and it is difficult to achieve real-time monitoring under massive data dimensions. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a fish health status monitoring device and method based on fish behavior analysis in view of the deficiencies in the prior art.

[0005] The fish health status monitoring device based on fish behavior analysis described in the present invention includes a water surface unit, a submerged unit, a bottom unit and a data processing center; the water surface unit, the submerged unit and the bottom unit are respectively installed on the water surface, in the water and on the bottom of the breeding pond, the water surface unit and the bottom unit are communicatively connected with the submerged unit, and the submerged unit is communicatively connected with the data processing center.

[0006] Preferably, the underwater unit includes an early warning light, a buoyancy module, a second communication module, a single-chip microcomputer, a first PVC layer, a second flexible pressure sensor for collecting pressure data in water, an underwater water quality sensor for collecting water quality parameters in water, and a second capacitive accelerometer for collecting vibration data in water; the early warning light is installed on the top of the first PVC layer, the buoyancy module is installed in the first PVC layer, the second flexible pressure sensor is installed below the first PVC layer, the second communication module, the single-chip microcomputer, and the second capacitive accelerometer are all installed in the first PVC layer, the underwater water quality sensor is installed on the outside of the first PVC layer, the second flexible pressure sensor, the second communication module, the underwater water quality sensor, and the second capacitive accelerometer are all electrically connected to the single-chip microcomputer, and the single-chip microcomputer transmits data with the data processing center through the second communication module.

[0007] Preferably, the buoyancy module controls the proportion of air therein through a single chip microcomputer to flexibly adjust the height of the underwater unit in the water so that the warning light is located above the water surface and the second flexible pressure sensor is located below the water surface.

[0008] Preferably, the water surface unit includes a buoyancy foam board, a first communication module, a first capacitive accelerometer for collecting water surface pressure data, and a first flexible pressure sensor for collecting water surface vibration data; the first communication module, the first capacitive accelerometer and the first flexible pressure sensor are all installed on the buoyancy foam board, and the first communication module is connected to the single-chip microcomputer through the second communication module.

[0009] Preferably, the underwater unit includes a third communication module, a counterweight module, a second PVC layer, a third flexible pressure sensor for collecting underwater pressure data, an underwater water quality sensor for collecting underwater water quality parameters, and a third capacitive accelerometer for collecting underwater vibration data; the third communication module, the counterweight module, and the third capacitive accelerometer are all installed in the second PVC layer, the third flexible pressure sensor is installed on the second PVC layer, and the underwater water quality sensor is installed on the outside of the second PVC layer.

[0010] Preferably, the flexible pressure sensor includes an upper encapsulation silicone layer, a lower encapsulation silicone layer, a nanolayer, a cross-electrode positive electrode and a cross-electrode negative electrode; the cross-electrode positive electrode and the cross-electrode negative electrode are both mounted on the lower encapsulation silicone layer, and the cross-electrode positive electrode and the cross-electrode negative electrode are arranged in a non-contact manner, the nanolayer is mounted between the upper encapsulation silicone layer and the lower encapsulation silicone layer, and a sealed package is formed between the upper encapsulation silicone layer and the lower encapsulation silicone layer to package the nanolayer, the cross-electrode positive electrode and the cross-electrode negative electrode.

[0011] Preferably, the capacitive accelerometer includes a shell and a micro-spring, a movable capacitor negative plate, a fixed capacitor negative plate, and a counterweight located in the shell; micro-springs are installed at both ends of the shell, a counterweight is installed on the micro-spring, a connecting rod is fixed between the two counterweights, a movable capacitor negative plate is fixed in the connecting rod, and the fixed capacitor negative plate is fixed on the inner wall of the shell.

[0012] A fish health monitoring method based on the fish health monitoring device based on fish behavior analysis, wherein a data processing center of the fish health monitoring device receives real-time data sent from a water surface unit, an underwater unit, and an underwater unit, and the data processing center processes the real-time data through a Stacking learning model to obtain a plurality of probability distribution values ​​for describing fish behavior, takes the largest probability distribution value among the plurality of probability distribution values ​​as a target distribution value, compares the target distribution value with a preset fish abnormal behavior probability threshold to determine the abnormal behavior of the fish, and generates an abnormal alarm signal corresponding to the abnormal behavior.

[0013] Preferably, the Stacking learning model is composed of XGBoost, RF, SVM as base learners, and Softmax classifier as meta learner; the real-time data includes vibration data, pressure data and water quality parameters; The vibration data are respectively input into the XGBoost and RF for processing, and the processing results are input into the Softmax classifier; the pressure data and water quality parameters are simultaneously input into the SVM for processing, and the processing results are input into the Softmax classifier, and the Softmax classifier outputs a plurality of probability distribution values ​​according to the input data.

[0014] Preferably, the calculation formula of the probability distribution value is: ; In the above formula, Data input to the kth Softmax classifier The corresponding probability distribution value; Refers to the data input to the kth Softmax classifier Perform an exponential transformation.

[0015] Beneficial Effects The advantages of the present invention are: vibration information, pressure information and current water quality parameters caused by fish behavior are collected in real time through multiple capacitive accelerometers, flexible pressure sensors and electrochemical sensors, and the vibration information, pressure information and water quality parameters are classified through a classification model, and the health status of the fish school is detected and warned in real time according to the classification results of the classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the installation structure of the fish health status detection and early warning device of the present invention; Figure 2 It is a data transmission flow chart of the fish health status monitoring device of the present invention; Figure 3 It is a schematic diagram of the upper three-dimensional structure of the water surface unit of the present invention; Figure 4 Schematic diagram of the underwater unit structure of the present invention, wherein the left side is a front view of the underwater unit, and the right side is a top view of the underwater unit; Figure 5 Schematic diagram of the underwater unit structure of the present invention, wherein the left side is a front view of the underwater unit, and the right side is a schematic diagram of the upper three-dimensional structure of the underwater unit; Figure 6 It is a schematic diagram of the structure of the flexible pressure sensor of the present invention; Figure 7 It is a schematic diagram of the structure of the capacitive accelerometer of the present invention; Figure 8 This is a schematic diagram of the working process of the device for detecting and warning the health status of fish of the present invention; Fig. 9 A schematic diagram of the Stacking learning model architecture of the present invention; Fig.10 The present invention provides a comparison table of abnormal fish behaviors and corresponding warning lights. DETAILED DESCRIPTION

[0017] The present invention will be further described below in conjunction with the embodiments, but it does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0018] The fish health status monitoring device based on fish behavior analysis of the present invention is as follows: Figure 1As shown, it includes a surface unit 102, an underwater unit 101, and an underwater unit 103. Among them, the underwater unit 101 uses STM32 as the main control system, which is responsible for collecting data from the surface and the underwater unit 101, and sending it to the cloud data processing center, and also for executing control commands from the data processing center. In this embodiment, the data processing center can use a cloud server, referred to as the cloud. In addition, the main control system of the underwater unit 101 also controls the air content in the buoyancy module to adjust the depth of the underwater unit 101 in the pool. The underwater unit 101 and the surface unit 102 are connected together by a nylon rope 104. The surface unit 102 floats on the water surface through a buoyancy plate. It has a built-in accelerometer for collecting vibration data, a flexible pressure sensor for collecting pressure data, a light sensor for collecting ambient light intensity, and a temperature sensor for collecting ambient temperature. These data are transmitted to the main control system of the underwater unit 101 through a communication module. The underwater unit 103 is fixed to the bottom of the water by a counterweight. It has a built-in accelerometer for collecting vibration data, a flexible pressure sensor for collecting pressure data, a turbidity sensor for collecting bottom turbidity, a pH sensor for collecting pH value at the bottom of the pool, and an ammonia nitrogen sensor for collecting ammonia nitrogen value at the bottom of the pool. These data are also sent to the main control system of the underwater unit 101 through the communication module. Figure 2 shown.

[0019] like Figure 3 As shown, the water surface unit 102 is mainly composed of a temperature and humidity sensor 301, a photosensor 302, a buoyancy foam plate 303, a first capacitive accelerometer 304, a first flexible pressure sensor 305, and a first communication module. Among them, the temperature and humidity sensor 301 is responsible for collecting the temperature and humidity in the breeding environment; the photosensor 302 is responsible for collecting the light intensity in the breeding environment; the buoyancy foam plate 303 is responsible for providing buoyancy to the water surface unit; the first capacitive accelerometer 304 is responsible for collecting vibration signals from all directions of the water surface; the first flexible pressure sensor 305 is mainly used to detect the subsequent vibration signal after receiving the pressure.

[0020] like Figure 4As shown, the underwater unit 101 is mainly composed of an early warning light 401, a buoyancy module 402, a second flexible pressure sensor 403, a second communication module 404, a single-chip microcomputer 405, an underwater water quality sensor 406, a first PVC layer 407, and a second capacitive accelerometer 408. Among them, the function of the early warning light 401 is to alarm the external environment when the device detects abnormal behavior of farmed fish. The buoyancy module 402 adjusts the proportion of air therein through the single-chip microcomputer and elastically adjusts the height of the device in the water. It is mainly composed of an air pump, a solenoid valve, a shell and a piston. The air pump is used to inflate the inside of the shell, and the opening of the solenoid valve is used to deflate the inside of the shell, thereby adjusting the proportion of air in the buoyancy module 402. The second flexible pressure sensor 403 is responsible for detecting pressure signals such as fish tail flicking and wall rubbing. The second communication module 404 is responsible for collecting sensor signals to the STM32 single-chip microcomputer and communicating with the server. The water quality sensor 406 in the water includes three sensors: PH, ammonia nitrogen, and dissolved oxygen, which are responsible for collecting water quality parameters in the water body. The first PVC layer 407 serves as a structural support. The second capacitive accelerometer 408 is mainly used to detect the subsequent vibration signal after receiving the pressure.

[0021] like Figure 5 As shown, the underwater unit 103 is mainly composed of a third communication module 501, a third flexible pressure sensor 502, an underwater water quality sensor 503, a counterweight module 504, a second PVC layer 505, and a third capacitive accelerometer 506. Among them, the third communication module 501 is responsible for overcoming the problem of poor communication conditions in the water environment and sending the water quality sensor data to the underwater unit 101. The third flexible pressure sensor 502 is responsible for collecting pressure data such as fish rubbing the bottom and diving. 503 is the underwater water quality sensor 503, which includes a turbidity sensor, a pH sensor, and an ammonia nitrogen sensor, which are responsible for detecting the water quality data of the bottom of the pool. The counterweight module 504 is responsible for counterweighting the underwater unit 103 to maintain its stability. The second PVC layer 505 serves as a structural support. The third capacitive accelerometer 506 is mainly used to detect subsequent vibration signals after receiving pressure.

[0022] For flexible pressure sensors, such as Figure 6As shown, it is mainly composed of an upper encapsulation silica gel 201, a lower encapsulation silica gel 205, a nanolayer 202, a cross electrode positive electrode 203 and a cross electrode negative electrode 204, wherein the upper and lower layers of silica gel can protect internal components, provide flexible support and conduct pressure, and the nanolayer 202 is a MXene nanolayer, which is a sensitive layer and is mainly a two-dimensional material composed of carbon, nitrogen, oxygen and other atoms. When subjected to pressure, the structure of the MXene nanolayer will change, affecting the electron conduction path therein and changing its conductivity, which makes it very sensitive to detect pressure changes. The cross electrode layer composed of the cross electrode positive electrode 203 and the cross electrode negative electrode 204 can detect changes in the conductivity of the MXene layer, thereby converting the pressure signal into an electrical signal output.

[0023] When the behavior of fish generates pressure signals, the upper encapsulation silicone 201 and the lower encapsulation silicone 205 will transmit the pressure to the nanolayer 202. At this time, the nanolayer 202 breaks and the conductivity changes. The cross-electrode layer positive electrode 203 and the cross-electrode negative electrode 204 convert the conductivity signal into a voltage signal and collect it as a pressure signal for sensing the pressure generated by the fish.

[0024] For capacitive accelerometers, such as Figure 7 As shown, it is mainly composed of a micro spring 901, a movable capacitor negative plate 902, a fixed capacitor negative plate 903, and a counterweight 904. Among them, the micro spring 901 mainly receives vibration signals and provides elastic force. When the device is vibrated by the behavior of fish, a lateral acceleration force will be generated. At this time, the micro spring 901 will transfer the acceleration to the counterweight 904, and the counterweight 904 will carry the movable capacitor negative plate 902 to move laterally. Since the distance between the movable capacitor negative plate 902 and the fixed capacitor negative plate 903 changes, the capacitance between the plates will also change. At this time, the acceleration signal can be measured by measuring the change in capacitance.

[0025] In terms of data transmission, the STM32 microcontroller is responsible for collecting the data sent by all sensors, and then after calibrating the data with a position code, the data is sent to the cloud (i.e., the data processing center) for processing, and at the same time, it will also provide feedback on the control data sent from the cloud. That is, when the cloud determines that the behavior of the fish in the fish pond is abnormal, the main control system receives the corresponding abnormal signal, and its warning light will flash to alarm, and different signal lights will flash based on different abnormalities. The cloud will process the data sent by each sensor, and since the surface, underwater, and underwater units 103 are evenly and orderly arranged in the breeding pond, each round of transmitted data will be position-encoded by the STM32 microcontroller to form spatial dimension data, thereby realizing three-dimensional monitoring of the behavior of the fish school. By processing the received data, the cloud can accurately judge the health of the fish school in the breeding pond and the location of the abnormality.

[0026] In the fish health monitoring method of the present invention, the data processing center processes the data through the Stacking learning model. As an integrated learning model method, the Stacking learning model uses the prediction results of multiple base learners as input to train a new learner for final prediction. As a learning model for common classification tasks, the stacking learning model is often deployed in refined edge tasks because of its high precision that is superior to traditional machine learning and its computational complexity that is far lower than that of deep learning. As a feedback behavior caused by various inducements, the behavior of fish can collect various types of wall-rubbing spectrograms, pressure and water quality signals through accelerometers, flexible sensors, water quality sensors, etc. The Stacking learning model is very suitable for integrating the characteristics of spectrograms and water quality and pressure signals to classify their coercive behaviors. The specific processing method is as follows.

[0027] like Figure 8 and 9 The Stacking learning model shown is composed of three base learners and a meta learner, where the base learners are mainly XGBoost, RF, and SVM, and the meta learner is mainly a Softmax classifier.

[0028] Among them, XGBoost and RF are mainly used to receive vibration data from the accelerometer, and the vibration data is spectrum data transformed from Laplace to the frequency domain, which describes the location, frequency and intensity of vibration. XGBoost can sort the importance of the features of the spectrum and capture important information behind the amplitude and frequency based on parasitic infection, fungal or bacterial infection, water quality environment stress, and external damage to fish. RF has built-in strong randomness, which can effectively avoid overfitting caused by XGBoost and suppress the noise data of the vibration spectrum brought by the water flow in the breeding pond. Inputting the output of XGBoost and RF into the meta-learner can perfectly process the characteristic patterns of the vibration spectrum. After fusing their features, powerful nonlinear modeling can be completed, and the vibration spectrum data can be effectively classified. The outputs of XGBoost and RF are usually logical values ​​of the category. These output values ​​can be understood as the confidence of the model in a certain category (unnormalized). These values ​​represent the combination of each sensor position, vibration frequency and vibration intensity.

[0029] SVM is mainly used to receive water quality parameter data, including pH, dissolved oxygen, ammonia nitrogen, temperature, humidity, light intensity, turbidity, etc. It also receives pressure level data output by flexible pressure sensors (such as no pressure, slight pressure, medium pressure, etc.). SVM is good at classifying low-dimensional data. By introducing radial basis sum function (RBF Kernel), it can solve nonlinear problems by mapping low-dimensional data to high dimensions. For dissolved oxygen, pH, and ammonia nitrogen data measured by sensors, SVM easily obtains nonlinear relationships between data and classifies environmental data of real-time measurement of breeding ponds. SVM outputs the classification results of each sensor, such as the categories output by flexible pressure sensors and water quality sensors respectively. Inputting the output of SVM into the meta-learner can assist vibration data in detecting whether the wall-rubbing behavior of fish is caused by environmental stress, and make a more accurate classification of fish behavior based on the data.

[0030] The following will give a more detailed introduction to the base learners XGBoost, RF, and SVM.

[0031] XGBoost, or eXtreme Gradient Boosting, is a decision tree-based gradient boosting algorithm with high efficiency and excellent performance. Its core idea is to reduce prediction errors by training a weak model and then adjusting the subsequent training process based on the model's incorrect predictions.

[0032] In XGBoost, each iteration adds a new base learner to the model. At the tth iteration, the XGBoost model is: ,in, is the output of the XGBoost model, and x is the input of the XGBoost model; is the learning rate, which is used to control the contribution of each new tree to the final result and prevent overfitting; is the output of the decision tree.

[0033] The objective function of XGBoost is: .

[0034] in, is the loss function; is the regularization term; , is the complexity penalty coefficient of the tree; T is the number of leaf nodes in the book; is the regularization parameter; is the weight of the jth leaf node; N is the number of samples.

[0035] Substituting the XGBoost model into the XGBoost objective function, the objective function can be simplified to: .

[0036] Expand the objective function using the second-order Taylor expansion, ignoring the constant term , keeping only the first and second terms, we can get the final objective function: .

[0037] In the above formula, is the first-order derivative of the loss function, i.e., the gradient; is the second-order derivative of the loss function. The first-order derivative contained in the objective function is and the second derivative It reflects the loss information of each sample under the current model, which is used to guide the construction of each tree.

[0038] In training tree When a node is split, the loss reduction caused by the split is measured by the gain. The gain is based on the local optimization of the current split point, which determines which split can maximize the improvement of the objective function under the current tree structure. The calculation formula of the split gain G is: .

[0039] In the process of XGBoost building a tree, we hope to minimize the objective function by selecting the best split point. That is to say, XGBoost does not find the minimum value of the objective function through the common gradient descent in machine learning, but through tree structure optimization. In each round of tree structure optimization, we need to determine the optimal split node based on the current tree structure. The gain G here is to decide whether to split by measuring the gain after splitting a node. The change in the objective function after splitting at a certain split point is calculated. Since the gain measures the reduction in the loss function brought about by the split, the larger the gain, the better the effect of the split on optimizing the objective function.

[0040] In simple terms, minimizing the objective function requires first passing through the loss function in the objective function. The gradient of , then substitute the gradient into G to calculate the split gain, and then choose the split (maximizing the gain) and minimize the objective function.

[0041] In XGBoost, is the weight of each leaf node in the decision tree, indicating the contribution of the leaf node to the final prediction. Find the partial derivative, set the partial derivative = 0, and you can get the leaf node weight for: .

[0042] After multiple iterations, we get the final XGBoost model: .

[0043] RF stands for Random Forest, which is an ensemble learning algorithm that is widely used in classification and regression problems. In classification tasks, it outputs the final result by constructing multiple decision trees and voting on the prediction results. It is efficient and robust.

[0044] In RF, bagging is first introduced for autonomous sampling. It is worth mentioning that due to random sampling, the data sets randomly sampled during the training of each decision tree are different. Furthermore, RF has good randomness to suppress the overfitting problem caused by XGBoost. Then, using the idea of ​​random subspace, f features are randomly extracted, node splitting is performed, a single classification decision subtree is constructed, and sub-judgment is made. Next, the above steps are repeated, bagging performs a sampling process with replacement, and then continues to construct a single classification decision subtree. It is worth mentioning that pruning operations will not be performed in RF, which can retain more complex features and increase the generalization ability of the model. In the end, subtrees one by one eventually form a forest. The result of the model is similar to voting for all parallel subtrees in the forest and selecting the most categories as the output.

[0045] SVM stands for Support Vectior Machines. As a classic binary classification algorithm, its main purpose is to find a hyperplane to segment samples. The principle of segmentation is to maximize the interval. In the fish health evaluation index, it is used to determine whether environmental stress causes fish to rub against the wall. The features it outputs will serve as an important indicator for meta-learner classification. In SVM, the most important task is to find a hyperplane in the sample space. Find a hyperplane in Used to distinguish different samples. The support vector is the point closest to the hyperplane. In this task, there are two support vector planes for binary classification. The planes are parallel to each other, and the hyperplane is located in the center of these two support vector planes. The ultimate goal of the task is to find the maximum margin .

[0046] Based on the processing of the received data by the above-mentioned base learners XGBoost, RF, and SVM, the output results of XGBoost, RF, and SVM are the logical values ​​of the predicted categories of the sensor data in each unit, and these values ​​are not normalized to probability distribution. And the meaning represented by each output value is only to complete the classification of local information. The effect of the Softmax classifier is to combine the prediction results of multiple base learners to comprehensively output the final classification result.

[0047] The Softmax classifier is an extension of logistic regression. Its core idea is to convert the output of the linear model into a probability distribution, thereby achieving classification of multiple categories. Softmax is essentially a fully connected layer. The function value is converted into a probability distribution through the Softmax function. The probability of each output represents the confidence of a certain type of classification sample. That is, the output of Softmax is used as the final output of the model. Softmax is widely used in deep learning, especially in neural networks, as an output layer.

[0048] Softmax first needs to input data (It corresponds to the output of XGBoost, RF, and SVM) Make linear changes: .

[0049] In the above formula, ω is the weight vector; x is the input feature vector; b is the bias term. For the output of each base learner, use Indicates. Softmax will perform Apply the exponential function to get the exponential value of each category: .

[0050] At this time, all logical values ​​will be converted to positive values, especially the influence of larger values, so that categories with higher confidence will occupy a larger proportion in subsequent processing.

[0051] This output is then converted into a probability distribution through the Softmax function: .

[0052] In the above formula, Refers to the data input to the kth Softmax classifier Perform exponential transformation; It is the exponential sum of all exponentially changing values.

[0053] Therefore, for the input category of the kth Softmax classifier, we have: .

[0054] is the probability distribution value of this category.

[0055] Abnormal behaviors of fish include: rubbing against the wall, floating, gathering in groups, swimming rapidly or escaping, reducing or stopping eating, avoiding light or changing swimming state, and normal. For each abnormal behavior, a corresponding probability threshold of abnormal behavior of fish is set. After the values ​​are obtained, the maximum value is taken as the target distribution value and compared with the probability threshold of fish abnormal behavior to determine the abnormal behavior of fish.

[0056] For example, the results output by softmax are 0.658, 0.243, and 0.098, respectively. According to the classification of the probability threshold of abnormal fish behavior, 0.658 is wall-rubbing behavior, 0.243 is floating head behavior, and 0.098 is normal. In the present invention, the maximum probability is used as the output of the model. Assuming that the threshold is 0.6, then P=0.658>0.6, then the model will judge it as a wall-rubbing result, and generate a corresponding abnormal alarm signal, and simultaneously send it to the underwater unit 101, so that the warning light 401 therein lights up. The specific abnormalities and the colors of the lights are as follows. Fig.10 shown.

[0057] In the Softmax classifier, the role of the loss function is to measure the difference between the model results and the actual results. The loss function can be expressed as: .

[0058] is the corresponding true result, which is the probability value, i is the sample category belonging to k; represents the probability value output by the softmax classifier; and N represents the number of samples.

[0059] Abnormal behaviors of fish may include rubbing against the wall, floating, gathering in groups, swimming rapidly or escaping, reducing or stopping eating, avoiding light, or changing swimming status, etc. These behaviors may be caused by multiple stress factors or may occur at a certain location in the water body. The Stacking algorithm automatically calculates the severity and location of environmental stress by inputting information from the breeding pond, and alarms through warning lights at corresponding locations, thereby realizing the monitoring of abnormal fish behaviors.

[0060] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A fish health status monitoring device based on fish behavior analysis, characterized in that: The invention comprises a water surface unit (102), a submerged unit (101), a bottom unit (103) and a data processing center; the water surface unit (102), the submerged unit (101) and the bottom unit (103) are respectively installed on the water surface, in the water and under the water of the breeding pond; the water surface unit (102) and the bottom unit (103) are communicatively connected with the submerged unit (101); and the submerged unit (101) is communicatively connected with the data processing center.

2. The fish health status monitoring device based on fish behavior analysis according to claim 1 is characterized in that: The underwater unit (101) comprises an early warning light (401), a buoyancy module (402), a second communication module (404), a single chip microcomputer (405), a first PVC layer (407), a second flexible pressure sensor (403) for collecting water pressure data, an underwater water quality sensor (406) for collecting water quality parameters in water, and a second capacitive accelerometer (408) for collecting water vibration data; the early warning light (401) is installed on the top of the first PVC layer (407), the buoyancy module (402) is installed in the first PVC layer (407), and the second flexible pressure sensor (403) is installed in the first PVC layer (407). ) is installed below the first PVC layer (407), the second communication module (404), the single-chip microcomputer (405), and the second capacitive accelerometer (408) are all installed in the first PVC layer (407), the water quality sensor (406) is installed on the outside of the first PVC layer (407), the second flexible pressure sensor (403), the second communication module (404), the water quality sensor (406), and the second capacitive accelerometer (408) are all electrically connected to the single-chip microcomputer (405), and the single-chip microcomputer (405) transmits data to and from a data processing center via the second communication module (404).

3. The fish health status monitoring device based on fish behavior analysis according to claim 2 is characterized in that: The buoyancy module (402) controls the proportion of air therein through the single chip computer (405) to flexibly adjust the height of the underwater unit (101) in the water, so that the warning light (401) is located on the water surface and the second flexible pressure sensor (403) is located below the water surface.

4. The fish health status monitoring device based on fish behavior analysis according to claim 2 is characterized in that: The water surface unit (102) comprises a buoyancy foam plate (303), a first communication module, a first capacitive accelerometer (305) for collecting water surface pressure data, and a first flexible pressure sensor (304) for collecting water surface vibration data; the first communication module, the first capacitive accelerometer (305), and the first flexible pressure sensor (304) are all mounted on the buoyancy foam plate (303), and the first communication module is communicatively connected to the single-chip computer (405) via a second communication module.

5. The fish health status monitoring device based on fish behavior analysis according to claim 2 is characterized in that: The underwater unit (103) comprises a third communication module (501), a counterweight module (504), a second PVC layer (505), a third flexible pressure sensor (502) for collecting underwater pressure data, a bottom water quality sensor (503) for collecting bottom water quality parameters, and a third capacitive accelerometer (506) for collecting bottom vibration data; the third communication module (501), the counterweight module (504), and the third capacitive accelerometer (506) are all installed in the second PVC layer (505), the third flexible pressure sensor (502) is installed on the second PVC layer (505), and the bottom water quality sensor (503) is installed on the outside of the second PVC layer (505).

6. The fish health status monitoring device based on fish behavior analysis according to any one of claims 2 to 5, characterized in that: The flexible pressure sensor comprises an upper encapsulation silicone (201), a lower encapsulation silicone (205), a nanolayer (202), a cross-electrode positive electrode (203) and a cross-electrode negative electrode (204); the cross-electrode positive electrode (203) and the cross-electrode negative electrode (204) are both mounted on the lower encapsulation silicone (205), and the cross-electrode positive electrode (203) and the cross-electrode negative electrode (204) are arranged in a non-contact manner; the nanolayer (202) is mounted between the upper encapsulation silicone (201) and the lower encapsulation silicone (205); and a sealed package for the nanolayer (202), the cross-electrode positive electrode (203) and the cross-electrode negative electrode (204) is formed between the upper encapsulation silicone (201) and the lower encapsulation silicone (205).

7. The fish health status monitoring device based on fish behavior analysis according to any one of claims 2 to 5, characterized in that: The capacitive accelerometer comprises a shell and a micro spring (901), a movable capacitor negative electrode plate (902), a fixed capacitor negative electrode plate (903), and a counterweight (904) located in the shell; micro springs (901) are installed at both ends of the shell, a counterweight (904) is installed on the micro spring (901), a connecting rod is fixed between the two counterweights (904), a movable capacitor negative electrode plate (902) is fixed in the connecting rod, and the fixed capacitor negative electrode plate (903) is fixed on the inner wall of the shell.

8. A method for monitoring fish health status based on the fish health status monitoring device based on fish behavior analysis according to any one of claims 1 to 7, characterized in that: The data processing center of the fish health monitoring device receives real-time data sent from a water surface unit (102), an underwater unit (101), and an underwater unit (103); the data processing center processes the real-time data through a stacking learning model to obtain a plurality of probability distribution values ​​for describing fish behavior; the largest probability distribution value among the plurality of probability distribution values ​​is taken as a target distribution value; the target distribution value is compared with a preset fish abnormal behavior probability threshold to determine the abnormal behavior of the fish; and an abnormal alarm signal corresponding to the abnormal behavior is generated according to the abnormal behavior.

9. The method for monitoring the health status of fish according to claim 8, characterized in that: The stacking learning model is composed of XGBoost, RF, SVM as base learners, and Softmax classifier as meta learner; the real-time data includes vibration data, pressure data and water quality parameters; The vibration data are respectively input into the XGBoost and RF for processing, and the processing results are input into the Softmax classifier; the pressure data and water quality parameters are simultaneously input into the SVM for processing, and the processing results are input into the Softmax classifier, and the Softmax classifier outputs a plurality of probability distribution values ​​according to the input data.

10. The method for monitoring the health status of fish according to claim 9, characterized in that: The calculation formula of the probability distribution value is: ； In the above formula, Data input to the kth Softmax classifier The corresponding probability distribution value; Refers to the data input to the kth Softmax classifier Perform an exponential transformation.

Images