A fish feed detection and formulation system

Abstract

The present invention discloses a fish feed detection and formulation system, which includes an aquaculture environment parameter acquisition and control platform, a fish feed formulation weight ratio prediction subsystem, and a fish feed formulation firefly algorithm optimization subsystem, to achieve the detection of aquaculture environment parameters, the prediction of the weight ratio of fish feed formulation, and the optimization of fish feed formulation; the present invention effectively solves the problems that the existing fish feed formulation does not consider the non-linearity, large lag of the change of aquaculture environment parameters and the complexity of aquaculture regulation, etc., which affect the environmental benefits of aquaculture, does not predict the weight ratio of fish feed formulation and optimize the fish feed formulation, thus greatly affecting the aquaculture benefits and production management problems.

Description

Technical Field

[0001] The present invention relates to the technical field of fish feed detection and formula automation, and particularly relates to a fish feed detection and formula system. Background Art

[0002] Aquaculture is an important part of agriculture and a symbol of agricultural modernization. With the favorable market demand, fish prices, and the optimism of farmers, the fish feed market will become increasingly active. Farmers' experience shows that the quality and price of fish feed formulas have become one of the key links in aquaculture efficiency. Studying reasonable fish feed formulas is of great practical significance in improving aquaculture. A reasonable fish feed formula should meet the needs of fish growth, development, and production, while avoiding waste of any one or more nutrients. It needs to be achieved according to the fish nutrition requirements and formula optimization system to achieve the goal of obtaining the minimum cost formula under the premise of meeting fish nutrition. The fish feed detection and formula system is a visual operation system, which has the characteristics of convenient operation, simple to understand, friendly interface design, and complete various indicators. Moreover, it should minimize nutrient waste. The ingenious optimization design of this system makes the required amount and actual supply amount of various nutrients for fish clear at a glance. Users can adjust appropriately to obtain the minimum cost formula. During the design process of this system, full consideration is given to the aquatic environment and breeding conditions to meet the balanced supply of nutrients, improve the economic benefits of aquaculture, and reduce the cost of aquaculture. Summary of the Invention

[0003] The present invention provides a fish feed detection and formula system, which effectively solves the problems that the existing fish feed formula does not consider the non-linearity, large lag of changes in aquaculture environment parameters, and the complexity of aquaculture regulation, etc., which affect the environmental benefits of aquaculture, does not predict the feed weight ratio of the fish feed formula, and does not optimize the fish feed formula, thus greatly affecting aquaculture efficiency and production management.

[0004] The present invention is achieved through the following technical solutions:

[0005] A fish feed detection and formula system, the system includes an aquaculture environment parameter acquisition and control platform, a fish feed formula feed weight ratio prediction subsystem, and a fish feed formula firefly algorithm optimization subsystem, to realize the detection of fish breeding environment parameters, the prediction of the fish feed formula feed weight ratio, and the optimization of the fish feed formula.

[0006] A further technical improvement scheme of the present invention is:

[0007] The aquaculture environment parameter acquisition and control platform consists of detection nodes, control nodes, gateway nodes, on-site monitoring terminals, cloud platforms, and remote monitoring computers. The detection nodes collect aquaculture environment parameters of fish and upload them to the cloud platform through the gateway nodes. The cloud platform provides the aquaculture environment parameters of fish to the remote monitoring computer for Web visualization management of the aquaculture environment parameter interface. The remote monitoring computer issues commands to the control nodes to implement remote environment control. Data is stored and information is published at the cloud platform end. The detection nodes and control nodes are responsible for collecting aquaculture environment parameters of fish and controlling aquaculture environment equipment. Bidirectional communication among the detection nodes, control nodes, on-site monitoring terminals, cloud platform, and remote monitoring computer is achieved through the gateway nodes, realizing the acquisition of aquaculture environment parameters and the control of aquaculture equipment. The aquaculture environment parameter acquisition and control platform is shown in Figure 1 。

[0008] A further technical improvement scheme of the present invention is:

[0009] The feed weight ratio prediction subsystem for fish feed formula includes a fuzzy C-means clustering algorithm, multiple CNN convolutional neural network models, multiple LSTM neural network models, a GRNN neural network model, a fuzzy recurrent neural network model, a NARX neural network model, a time-delay neural network model, a feed weight ratio trend prediction module, and an environment evaluation module. The fish feed formula is used as the input of the fuzzy C-means clustering algorithm. The fish feed formulas of multiple categories output by the fuzzy C-means clustering algorithm are respectively used as the inputs of the corresponding multiple CNN convolutional neural network models. The outputs of the multiple CNN convolutional neural network models are respectively used as the inputs of the corresponding multiple LSTM neural network models. The outputs of the multiple LSTM neural network models are used as the input of the GRNN neural network model. The outputs of the GRNN neural network model, the time-delay neural network model, the feed weight ratio trend prediction module, and the environment evaluation module are used as the input of the fuzzy recurrent neural network model. The output of the fuzzy recurrent neural network model is used as the input of the NARX neural network model. The output of the NARX neural network model is used as the input of the time-delay neural network model. The output value of the NARX neural network model is used as the feed weight ratio of the fish feed formula. The feed weight ratio prediction subsystem for fish feed formula is shown in Figure 2 。

[0010] A further technical improvement scheme of the present invention is:

[0011] The feed weight ratio trend prediction module includes a CNN convolutional neural network model, an LSTM neural network model, an ARIMA model, and a NARX neural network model. The historical data of the fish feed weight ratio are respectively used as the inputs of the CNN convolutional neural network model and the ARIMA model. The outputs of the CNN convolutional neural network model and the ARIMA model are used as the input of the NARX neural network model. The output value of the NARX neural network model is used as the output of the feed weight ratio trend prediction module. The feed weight ratio trend prediction module is shown in Figure 2 。

[0012] A further technical improvement solution of the present invention is as follows:

[0013] The environmental assessment module includes multiple LSTM neural network models, multiple autoassociative neural network models, and NARX neural network models. The outputs of multiple groups of temperature, dissolved oxygen, and pH value sensors are respectively used as the inputs of the corresponding multiple LSTM neural network models. The outputs of the multiple LSTM neural network models are respectively used as the inputs of each autoassociative neural network model. The outputs of the multiple autoassociative neural network models are used as the inputs of the NARX neural network model. The output value of the NARX neural network model is used as the output of the environmental assessment module; see the environmental assessment module Figure 2 .

[0014] A further technical improvement solution of the present invention is as follows:

[0015] The firefly algorithm optimization subsystem for fish feed formula includes five links: initialization of firefly population, determination of fluorescence brightness, calculation of objective function, update of firefly position, and determination of the optimal fish feed formula. The firefly population is randomly distributed in the search space as the initial solution. Each firefly is regarded as a fish feed formula. It will be attracted by fireflies that are brighter than it. The attractiveness of fireflies is positively correlated with brightness. For any two fireflies, one firefly will move towards the other firefly that is brighter than it. The brightness decreases with the increase of distance, and the fireflies gather around the fireflies with high brightness; the firefly individuals of each fish feed formula are used as the inputs of the fish feed formula weight ratio prediction subsystem. The output of the fish feed formula weight ratio prediction subsystem is used as the predicted value of the weight ratio of the firefly individuals of the fish feed formula. The reciprocal of the predicted value of the weight ratio of the firefly individuals of each fish feed formula is used as the fluorescence brightness of the firefly individuals of the fish feed formula. The larger the reciprocal of the predicted value of the weight ratio of the firefly individuals of each fish feed formula, the higher the fluorescence brightness of the firefly individuals of the fish feed formula. The sum of the reciprocals of the predicted values of the weight ratios of the firefly individuals of each fish feed formula in the firefly population is the total fluorescence brightness of the firefly population. See the firefly algorithm optimization subsystem for fish feed formula Figure 3 .

[0016] Compared with the prior art, the present invention has the following obvious advantages:

[0017] 1. The present invention utilizes the advantages that the CNN convolutional neural network model can extract the features of fish feed formula and historical data of fish feed weight ratio, and can shorten the feature extraction time, and the LSTM neural network model can memorize the relationship between the fish feed formula and historical data of fish feed weight ratio with strong memory dependence, to solve the problems of spatial feature extraction of the activity sequence data of fish feed formula and historical data of fish feed weight ratio and data dependence of time features; First, the preprocessed sequence data of fish feed formula and historical data of fish feed weight ratio are input into the CNN convolutional neural network model to extract the corresponding spatial feature vectors; Secondly, the different activity spatial feature vectors of the fish feed formula and historical data of fish feed weight ratio extracted in the previous step are used as the input of the LSTM neural network model, and the data interaction of the input gate, forget gate and output gate in the LSTM neural network model is used to process and predict the time feature interaction problem between the activity sequence data of fish feed formula weight ratio, so as to improve the accuracy and time efficiency of predicting the weight ratio of fish feed formula.

[0018] 2. The present invention uses the CNN convolutional neural network model to extract the high-dimensional spatial features of fish feed formula and historical data of fish feed weight ratio, so as to realize the feature extraction of fish feed formula and historical data of fish feed weight ratio; At the same time, the LSTM neural network model is selected to process the spatial feature sequence output by the CNN convolutional neural network model, to mine the time series information in the fish feed formula and historical data of fish feed weight ratio, to extract the time features of fish feed formula and historical data of fish feed weight ratio in the time dimension, and to realize the accurate prediction of fish feed formula weight ratio.

[0019] 3. The main advantage of the convolutional layer of the CNN convolutional neural network model of the present invention lies in extracting weight sharing and sparse connection in the fish feed formula and historical data of fish feed weight ratio. Weight sharing means that the weight of the convolutional kernel of the CNN convolutional neural network model remains unchanged during the convolution operation, and the weight of each convolutional kernel is the same for the entire area of fish feed formula and historical data of fish feed weight ratio; Sparse connection means that each convolutional kernel of the CNN convolutional neural network model only uses specific local area data in the previous layer of data for operation, and does not use the global fish feed formula and historical data of fish feed weight ratio; The characteristics of weight sharing and sparse connection of the convolutional kernel of the CNN convolutional neural network model greatly reduce the number of spatial feature parameters of the fish feed formula and historical data of fish feed weight ratio, thus preventing overfitting of the CNN convolutional neural network model, accelerating the training speed of the CNN convolutional neural network model and improving the prediction accuracy of fish feed formula.

[0020] IV. The LSTM neural network model of the present invention is similar to a standard network with a recurrent hidden layer. The only change is that a memory module is used to replace the original hidden layer units. By means of the self-feedback of the internal state of the memory cell and the truncation of the error of the input and output, the problems of gradient vanishing and explosion are solved. Compared with the BP neural network and ordinary RNN, the LSTM adds one state unit c and three control gates, which greatly increases the feature inclusion ability and memory ability of the model, and avoids underfitting and gradient vanishing. The function of the LSTM neural network model aims at the correlation existing in the fish feed formula, the historical data of the fish feed weight ratio, and the aquaculture environment data, remembers this relationship and the change of this relationship over time, so as to obtain more accurate results. The LSTM neural network model realizes the prediction of the fish feed formula weight ratio and the water quality parameter level of the aquaculture pond environment, and improves the prediction accuracy.

[0021] V. The LSTM neural network model of the present invention has a chain-like repeating network structure similar to that of the standard RNN. The repeating network in the standard RNN is very simple, while the repeating network in the LSTM neural network model has four interaction layers, including three gate layers and one tanh layer. The processor state is a key variable in the LSTM neural network model. It carries the information of the previous steps of the feed weight ratio prediction and gradually passes through the entire LSTM neural network model. The gates in the interaction layer can partially delete the processor state of the previous step and add new information of the feed weight ratio prediction to the processor state of the current step according to the hidden state of the previous step and the input of the current step. The input of each repeating network includes the hidden state and the processor state of the feed weight ratio prediction of the previous step and the input of the current step. The processor state is updated according to the calculation results of the four interaction layers. The updated processor state and the hidden state form the output and are passed to the next step.

[0022] VI. The LSTM neural network model of the present invention is a recurrent neural network with four interacting layers in the repeating network. It can not only extract information from the feed weight ratio prediction sequence data like a standard recurrent neural network, but also retain the information with long-term correlation from the previous relatively distant steps. The feed weight ratio prediction data is sequence data, and its change trend is meaningful. In addition, due to the relatively small sampling interval of the feed weight ratio prediction, there is long-term spatial correlation in the feed weight ratio prediction, and the LSTM neural network model has sufficient long-term memory to handle this problem.

[0023] VII. In the cascaded LSTM neural network model of the present invention, first, the relatively easy-to-predict feed conversion ratio data is reconstructed at the shallow level, and then the generated feed conversion ratio data is used as the input for the next level. The prediction result at the deep level is not only based on the input values in the feed conversion ratio data training data, but also affected by the feed conversion ratio data result at the shallow level. This method can more effectively extract the information contained in the input data of the feed conversion ratio data and improve the accuracy of the model in predicting the feed conversion ratio data.

[0024] VIII. The present invention combines the fuzzy C-means clustering (FCM), CNN convolutional neural network model, LSTM neural network model and GRNN neural network model technologies and applies them to the measurement of the feed conversion ratio of fish feed formulations. First, the FCM method is used to classify the fish feed formulation samples, and then the CNN convolutional neural network model and the LSTM neural network model are used in series to establish local measurement models for the feed conversion ratio of fish feed formulations. The outputs of multiple local models are fused through the GRNN neural network model. The results show that the fish feed conversion ratio measurement model constructed by the above method has a good training speed and high measurement accuracy.

[0025] IX. The present invention adopts a dynamic recursive network of the NARX neural network model to establish the NARX neural network model through the delay module and feedback of the feed conversion ratio of the feed formulation. It is a data correlation modeling idea that realizes and simulates the function along the sequence of multiple time feed conversion ratio parameters extended in the time axis direction of the feed conversion ratio parameter. This method establishes a feed conversion ratio combination model through the feed conversion ratio parameters within a period of time. The feed conversion ratio parameters output by the model are used as inputs in the feedback effect for closed-loop training to improve the calculation accuracy of the neural network. The NARX network model realizes continuous dynamic prediction of the feed conversion ratio. The input includes the feed conversion ratio input and output history feedback of the feed formulation for a period of time. This part of the feedback input can be considered to contain the historical information of the fish feed conversion ratio state for a period of time and participate in the prediction of the fish feed conversion ratio. For a suitable feedback time length, good prediction results are obtained. The NARX neural network prediction mode of this patent provides an effective method for predicting the fish feed conversion ratio.

[0026] X. The present invention uses the NARX neural network to establish a feed conversion ratio prediction model. Since a dynamic recursive network of the model is established by introducing a delay module and output feedback, it introduces the input and output vector delay feedback into the network training to form a new input vector, which has good non-linear mapping ability. The input of the network model not only includes the original input data, but also the output data after training. The generalization ability of the network is improved, making it have better prediction accuracy and adaptability than traditional static neural networks in non-linear feed conversion ratio time series prediction.

[0027] XI. The GRNN neural network model adopted by the present invention is simple and complete in structure. Its internal structure is determined with the determination of sample points. It has less requirements for data samples. As long as there are input and output samples, even if the data is scarce, it can converge to the regression surface. It has clear probability meaning, good generalization ability, local approximation ability and fast learning characteristics, can approximate any type of function, and in the process of establishing and learning the network model, only the smoothing factor needs to be adjusted and selected to finally determine the model. The process of establishing the network is also the training process of the network, without special training. In terms of the fusion effect on the dynamic system, the GRNN neural network model has the characteristics of simple network establishment process, few influencing factors, strong local approximation ability, fast learning speed and good simulation performance. Therefore, the GRNN neural network model is very suitable for the fusion of feed-to-weight ratio. This patent takes advantage of the self-adaptability, self-learning and non-linear approximation with arbitrary precision of the GRNN neural network model. Therefore, this patent uses the GRNN neural network model to perform the fusion of feed-to-weight ratio, which better meets the robustness and fault tolerance of the feed-to-weight ratio fusion model.

[0028] XII. The GRNN neural network model adopted by the present invention has strong non-linear mapping ability, flexible network structure, high fault tolerance and robustness, and is suitable for the fusion of feed-to-weight ratio. The GRNN neural network model has stronger advantages than the RBF network in terms of approximation ability and learning speed. The network finally converges to the optimized regression surface with more accumulated sample quantities, and when the sample data is less, the network can also process unstable data, and the effect of fusing the feed-to-weight ratio is also better. The GRNN neural network model has strong generalization ability, high fusion accuracy and stable algorithm. The GRNN neural network model also has the advantages of fast convergence speed, few adjusted parameters and not easy to fall into local minimum. The network fusion operation speed is fast, and it has good application prospects for the fusion of feed-to-weight ratio.

[0029] XIII. The ARIMA model adopted by the present invention is based on the fact that the original data of the feed-to-weight ratio fusion follows a time series distribution. Using the principle that the changes in the feed-to-weight ratio fusion all have a certain inertial trend, the original time series variables of the feed-to-weight ratio fusion that integrate trend factors, periodic factors and random errors and other factors are transformed into a stationary random sequence with zero mean through methods such as differential data conversion. By repeatedly identifying and comparing model diagnostics and selecting an ideal model for numerical fitting and prediction of the feed-to-weight ratio fusion. This method combines the advantages of autoregressive and moving average methods, has the characteristics of not being restricted by data types and strong applicability, and is a model with better short-term prediction effect for the feed-to-weight ratio fusion.

[0030] 14. The present invention adopts a fuzzy recursive neural network structure. By introducing internal variables in the fuzzy rule layer, the static network is endowed with dynamic characteristics. The activation degree of each rule at time K of the network includes not only the activation degree value calculated from the current input, but also the contribution of the activation degree values of all rules at the previous moment. Therefore, the accuracy of network identification is improved, and the prediction of the feed-to-weight ratio can be better completed. A fuzzy recursive neural network model is used to establish a prediction model for the feed-to-weight ratio. It is a typical dynamic recursive neural network. Its feedback connection consists of a group of "structural" units, which are used to memorize the past state of the hidden layer and, together with the network input at the next moment, serve as the input of the hidden layer units. This property enables some recursive networks to have a dynamic memory function, thus making them suitable for establishing a prediction model for the time series feed-to-weight ratio. Simulation experiments show that the model has good dynamic performance, high prediction accuracy, and stable prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is the aquaculture environment parameter acquisition and control platform of this patent;

[0032] Figure 2 is the fish feed formula feed-to-weight ratio prediction subsystem of this patent;

[0033] Figure 3 is the flowchart of the firefly algorithm optimization subsystem for the fish feed formula of this patent;

[0034] Figure 4 is the detection node of this patent;

[0035] Figure 5 is the control node of this patent;

[0036] Figure 6 is the gateway node of this patent;

[0037] Figure 7 is the on-site monitoring terminal software of this patent. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] Combined with the attached Figures 1-7 , the technical solution of this application is further described:

[0039] 1. Design of the overall system functions

[0040] A fish feed detection and formulation system of the present invention realizes the detection of aquaculture environment parameters, the prediction of the feed weight ratio of fish feed formulations, and the optimization of feed formulations. The system consists of three parts: an aquaculture environment parameter acquisition and control platform, a fish feed formulation weight ratio prediction subsystem, and a fish feed formulation firefly algorithm optimization subsystem. The aquaculture environment parameter acquisition and control platform includes a detection node, a control node, a gateway node, a on-site monitoring terminal, a cloud platform, and a remote monitoring terminal for aquaculture environment parameters. The detection node and the control node are constructed into a LoRa network communication to realize the LoRa network communication between the detection node, the control node, and the gateway node; the detection node sends the detected aquaculture environment parameters to the on-site monitoring terminal and the cloud platform through the gateway node, and realizes the two-way transmission of aquaculture environment parameters and related control information between the gateway node, the cloud platform, the on-site monitoring terminal, and the remote monitoring terminal. The aquaculture environment parameter acquisition and control platform is shown in Figure 1 as follows.

[0041] II. Design of the Detection Node

[0042] A large number of detection nodes 1 based on the LoRa communication network are used as the aquaculture environment parameter perception terminals, and the detection nodes realize the information interaction between the on-site monitoring terminals through the LoRa communication network. The detection node includes sensors for collecting aquaculture environment temperature, dissolved oxygen, pH value, and salinity parameters and corresponding signal conditioning circuits, an STM32 microprocessor, and an SX1278 radio frequency module for LoRa network communication; the software of the detection node mainly realizes the LoRa network communication and the acquisition and preprocessing of aquaculture environment parameters. The software is designed using the C language program, with high compatibility, greatly improving the work efficiency of software design and development, and enhancing the reliability, readability, and portability of the program code. The structure of the detection node is shown in Figure 4 .

[0043] III. Design of the Control Node

[0044] The control node realizes the information interaction with the gateway node through the LoRa network. The control node includes 4 digital-to-analog conversion circuits corresponding to controlling external devices, an STM32 microprocessor, 4 external device controllers, and an SX1278 radio frequency module for the LoRa communication network; the 4 external device controllers are a temperature controller, a dissolved oxygen controller, a pH value controller, and a salinity controller. The control node is shown in Figure 5 .

[0045] IV. Design of the Gateway Node

[0046] The gateway node includes SX1278, NB-IoT module, STM32 single-chip microcomputer and RS232 interface. The gateway node includes an SX1278 radio frequency module to implement a LoRa communication network for communication between the detection node and the control node. The NB-IoT module realizes two-way data interaction between the gateway and the cloud platform. The RS232 interface is connected to the on-site monitoring terminal to realize information interaction between the gateway and the on-site monitoring terminal. The gateway node is shown in Figure 6 .

[0047] V. Cloud Platform Design

[0048] The cloud platform supports multiple transmission protocols to provide high-quality services such as simple massive connection, cloud storage, message distribution, and big data analysis for various cross-platform Internet of Things applications and industry solutions, and has good visualization applications. First, create a product for aquaculture environment monitoring on the cloud platform, and connect the detection node, control node, gateway node, on-site monitoring terminal, and remote monitoring computer to the created product according to the platform's transmission protocol, and complete operations such as establishing a TCP connection with the cloud platform server and transmitting data on the Internet, so as to realize two-way transmission of data and information between them.

[0049] VI. Remote Monitoring Computer Design

[0050] The remote monitoring computer manages the Web visualization interface of aquaculture environment parameters, issues instructions to the control node to implement remote environment control, stores data and publishes information at the cloud platform end. Aquaculture personnel use the browser of the remote monitoring computer based on the B / S architecture to access and view real-time aquaculture environment information, query and export historical data, and implement remote control of aquaculture equipment. The Web page of the remote monitoring computer has an automatic alarm function for managers to take measures in time.

[0051] VII. On-Site Monitoring Terminal Software Design

[0052] The on-site monitoring terminal is an industrial control computer. The on-site monitoring terminal mainly realizes the collection and processing of aquaculture environment parameters, and realizes information interaction with the gateway node. The main functions of the on-site monitoring terminal are communication parameter setting, data analysis and data management, fish feed formula weight ratio prediction subsystem and fish feed formula firefly algorithm optimization subsystem. The structures of the fish feed formula weight ratio prediction subsystem and the fish feed formula firefly algorithm optimization subsystem are shown in Figure 2 and Figure 3 . The management software selects Microsoft Visual++6.0 as the development tool and calls the system's Mscomm communication control to design the communication program. The functions of the on-site monitoring terminal software are shown in Figure 7 . The design processes of the fish feed formula weight ratio prediction subsystem and the fish feed formula firefly algorithm optimization subsystem are as follows:

[0053] (1). Fish Feed Formula Weight Ratio Prediction Subsystem

[0054] It includes the fuzzy C-means clustering algorithm, multiple CNN convolutional neural network models, multiple LSTM neural network models, GRNN neural network model, fuzzy recurrent neural network model, NARX neural network model, time-delay neural network model, weight ratio trend prediction module and environmental evaluation module. The design of each model is as follows:

[0055] 1. Design of the Fuzzy C-Means Clustering Algorithm

[0056] The fish feed formula is used as the input of the fuzzy C-means clustering algorithm. The fish feed formulas of multiple categories output by the fuzzy C-means clustering algorithm are respectively used as the inputs of the corresponding multiple CNN convolutional neural network models. Let the finite set X = {x1, x2,... x n} be a set composed of n fish feed formula samples, which are fish feed formulas respectively. C is the predetermined number of categories, and m i (i = 1, 2,... c) is the center of each cluster, and μ j (x i ) is the membership degree of the i-th sample with respect to the j-th category. The clustering criterion function is defined by the membership function as:

[0057]

[0058] Where: ||x i -m j || is the Euclidean distance between x i and m j ; b is the fuzzy weighted power exponent, which is a parameter that can control the fuzziness of the clustering result; M is the fuzzy C-partition matrix of X, and V is the set of clustering centers of X. The result of the fuzzy C-means clustering algorithm is to obtain M and V that minimize the criterion function. In the fuzzy C-means clustering method, it is required that the sum of the membership degrees of the samples to each cluster is 1, that is:

[0059]

[0060] The FCM algorithm can be completed according to the following iterative steps:

[0061] A. Set the number of clusters c, parameter b, algorithm termination threshold ε, iteration number t = 1, and the maximum allowable number of iterations is t max ; B. Initialize each clustering center m i ; C. Calculate the membership function using the current clustering centers; D. Update the clustering centers of each category using the current membership function; E. Select an appropriate matrix norm. If ||V(t + 1) - V(t)|| ≤ ε or t ≥ t max, Stop the operation; otherwise, t = t + 1, and return to step C. When the algorithm converges, the clustering centers of each class and the membership degrees of each sample to each class are obtained, and the fuzzy clustering division is completed. Finally, the fuzzy clustering result is defuzzified to transform the fuzzy clustering into a deterministic classification, realizing the final clustering segmentation.

[0062] 2. CNN Convolutional Neural Network Model Design

[0063] The fish feed formulas of multiple categories output by the fuzzy C-means clustering algorithm are respectively used as the inputs of the corresponding multiple CNN convolutional neural network models, and the outputs of the multiple CNN convolutional neural network models are respectively used as the inputs of the corresponding multiple LSTM neural network models. The CNN convolutional neural network model can directly automatically mine and extract sensitive spatial features representing the system state from a large amount of fish feed formula and fish feed weight ratio historical data. The structure of the CNN convolutional neural network model mainly includes 4 parts: ① Input layer (Input). The input layer is the input of the CNN convolutional neural network model. Generally, the original data or preprocessed signals of the fish feed formula and fish feed weight ratio historical data are directly input after normalization. ② Convolution layer (Conv). Since the data dimension of the input layer is large, it is difficult for the CNN convolutional neural network model to directly and comprehensively perceive all the input information of the fish feed formula and fish feed weight ratio historical data. It is necessary to divide the input data into several parts for local perception, and then obtain global information through weight sharing, while reducing the complexity of the CNN convolutional neural network model structure. This process is the main function of the convolution layer. The specific process is to use a convolution kernel of a specific size to traverse and perform convolution operations on the input signal at a fixed step size, so as to realize the mining and extraction of sensitive features of the input signal of the fish feed formula and fish feed weight ratio historical data. ③ Pooling layer (Pool, also known as downsampling layer). Since the data samples obtained after the convolution operation still have a large dimension, it is necessary to compress the data volume and extract key information to avoid excessive model training time and overfitting. Therefore, a pooling layer is connected after the convolution layer to reduce the dimension. Considering the peak characteristics of the defect features, the maximum pooling method is used for downsampling. ④ Fully connected layer. After all the convolution operations and pooling operations, the feature extraction data enters the fully connected layer. Each neuron in this layer is fully connected to all the neurons in the previous layer to integrate the local feature information extracted by the convolution layer and the pooling layer. At the same time, to avoid overfitting, the dropout technology is added to this layer. The output value of the last fully connected layer will be passed to the output layer, and the pooling results of the last layer are connected together in a head-to-tail manner to form the output layer.

[0064] 3. LSTM Neural Network Model Design

[0065] The outputs of multiple CNN convolutional neural network models are respectively used as the inputs of the corresponding multiple LSTM neural network models, and the outputs of the multiple LSTM neural network models are used as the inputs of the GRNN neural network model. The LSTM neural network model is a time-recurrent neural network (RNN) composed of long short-term memory (LSTM) cells, which is called the LSTM neural network model time-recurrent neural network and is usually also called the LSTM neural network model network. The LSTM neural network model introduces the mechanisms of memory cells and hidden layer states to control the information transfer between hidden layers. There are 3 gate calculation structures in the memory cell of an LSTM neural network model, namely the input gate, the forget gate, and the output gate. Among them, the input gate can control the addition or filtering of new information; the forget gate can forget the information to be discarded and retain the useful information in the past; the output gate can make the memory cell only output the information related to the current time step. These 3 gate structures perform operations such as matrix multiplication and non-linear summation in the memory cell, so that the memory will not decay during continuous iteration. The long short-term memory (LSTM) structural unit consists of a cell, an input gate, an output gate, and a forget gate. The LSTM neural network model is a model that can maintain short-term memory for a long time and is suitable for predicting the dynamic change of the feed weight ratio in the fish feed formula. The LSTM neural network model effectively prevents the gradient disappearance during the training of RNN. The long short-term memory (LSTM) network is a special RNN. The LSTM neural network model can learn long-term dependence information and avoid the problem of gradient disappearance at the same time. In the internal structure of the neuron, the LSTM neural network model adds a structure called a memory cell in the neural nodes of the hidden layer of the RNN to remember the dynamic change information of the feed weight ratio of the past fish feed formula, and adds three gate structures (Input, Forget, Output) to control the use of the historical information of the feed weight ratio of the fish feed formula. Let the time series values of the input feed weight ratio of the fish feed formula be (x1, x2, …, x T ) and the hidden layer states be (h1, h2, …, h T ), then at time t:

[0066] i t = sigmoid(W hi h t-1 + W xi X t ) (3)

[0067] f t= sigmoid(W hf h t-1 + W hf X t ) (4)

[0068] c t = f t ⊙ c t-1 + i t ⊙ tanh(W hc h t-1 + W xc X t ) (5)

[0069] o t = sigmoid(W ho h t-1 + W hx X t + W co c t ) (6)

[0070] h t = o t ⊙ tanh(c t ) (7)

[0071] where i t , f t , o t represent the input gate, forget gate, and output gate, c t represents the cell unit, W h represents the weight of the recurrent connection, W x represents the weight from the input layer to the hidden layer, sigmoid and tanh are two activation functions. Four LSTM neural network models using long short-term memory are used to predict the weight ratio of fish feed formulations. This method first establishes an LSTM neural network model. A training set is established using the preprocessed weight ratio data of fish feed formulations and the model is trained. The LSTM neural network model takes into account the temporal and non-linear nature of the weight ratio changes in fish feed formulations and has a high prediction accuracy for the dynamic weight ratio of fish feed formulations.

[0072] 4. GRNN Neural Network Model Design

[0073] The outputs of multiple LSTM neural network models serve as the inputs of the GRNN neural network model, and the outputs of the GRNN neural network model serve as the inputs of the fuzzy recurrent neural network model. The GRNN (Generalized Regression Neural Network) neural network model is a local approximation network. It is based on mathematical statistics and has a clear theoretical basis. Once the learning samples are determined, the network structure and connection values are also determined. During the training process, only one variable, the smoothing parameter, needs to be determined. The learning of the GRNN neural network depends entirely on data samples, and it has stronger advantages than the BRF network in terms of approximation ability and learning speed. It has a strong non-linear mapping, a flexible network structure, and high fault tolerance and robustness, and is particularly suitable for the rapid approximation of functions and the processing of unstable data. The GRNN neural network model has very few manually adjustable parameters, and the learning of the network depends entirely on data samples. This characteristic enables the network to minimize the impact of human subjective assumptions on the prediction results. The GRNN neural network has a powerful prediction ability under small samples, and also has characteristics such as fast training and strong robustness, and is basically not troubled by the multicollinearity of input data. The GRNN neural network model structure constructed in this patent consists of an input layer, a pattern layer, a summation layer, and an output layer. The input vector X of the GRNN neural network model is an n-dimensional vector, and the network output vector Y is a k-dimensional vector X = {x1, x2, …, x n} T and Y = {y1, y2, …, y k} T . The number of neurons in the pattern layer is equal to the number of training samples m, and each neuron corresponds to a training sample one by one. The transfer function p i of the neurons in the pattern layer is:

[0074] p i = exp{-[(x - x i ) T (x - x i )] / 2σ}, (i = 1, 2, …, m) (8)

[0075] The outputs of the neurons in the above formula enter the summation layer for summation. The summation layer functions are divided into two categories, which are respectively:

[0076]

[0077]

[0078] Among them, y ij is the jth element value in the output vector of the ith training sample. According to the aforementioned GRNN algorithm, the estimated value of the jth element of the network output vector Y is:

[0079] y j= s Nj / s D , (j = 1, 2, … k) (11)

[0080] The GRNN neural network model is based on mathematical statistics and can approximate the implicit mapping relationship according to sample data. The output result of the network can converge to the optimal regression surface. Especially in the case of scarce sample data, satisfactory prediction results can also be obtained. The GRNN neural network model has strong classification ability and fast learning speed. It is mainly used to solve function approximation problems and also has high parallelism in terms of structure.

[0081] 5. Design of Fuzzy Recurrent Neural Network Model

[0082] The outputs of the GRNN neural network model, the time-delay neural network model, the material weight ratio trend prediction module, and the environmental evaluation module are used as the inputs of the fuzzy recurrent neural network model. The fuzzy recurrent neural network (HRFNN) is a multi-input single-output network topology. The network consists of four layers: the input layer, the membership function layer, the rule layer, and the output layer. The network contains n input nodes, where each input node corresponds to m conditional nodes, m represents the number of rules, nm rule nodes, and 1 output node. In the figure, the first layer introduces the input into the network; the second layer fuzzifies the input, and the membership function used is the Gaussian function; the third layer corresponds to fuzzy inference; the fourth layer corresponds to defuzzification operation. Let represent the input and output of the i-th node in the k-th layer respectively. Then the signal transmission process inside the network and the input-output relationship between layers can be described as follows. The first layer: input layer, each input node in this layer is directly connected to the input variable. The input and output of the network are expressed as:

[0083]

[0084] In the formula and are the input and output of the i-th node in the input layer of the network, and N represents the number of iterations.

[0085] The second layer: membership function layer, the nodes in this layer fuzzify the input variable. Each node represents a membership function, and the Gaussian function is used as the membership function. The input and output of the network are expressed as:

[0086]

[0087] In the formula m i j and σ ij represent the mean center and width value of the j-th Gaussian function of the i-th linguistic variable in the second layer respectively. m is the total number of linguistic variables corresponding to the input node.

[0088] Layer III: Fuzzy reasoning layer, i.e. rule layer, adds dynamic feedback to make the network have better learning efficiency. The feedback link introduces internal variables h k , the sigmoid function is selected as the activation function of the internal variables of the feedback link. The input and output of the network are expressed as:

[0089]

[0090] Where ω jk is the connection weight of the recursive part. The neurons in this layer represent the antecedent part of the fuzzy logic rule. The nodes in this layer perform π operation on the output of the second layer and the feedback of the third layer. is the output of the third layer, and m represents the number of rules when fully connected. The feedback link mainly calculates the value of the internal variable and the activation strength of the corresponding membership function of the internal variable. The activation strength is related to the matching degree of the rule node of the third layer. The internal variables introduced in the feedback link include two types of nodes: receiving nodes and feedback nodes. The receiving node uses weighted summation to calculate the internal variable to achieve the defuzzification function; the internal variable represents the result of fuzzy reasoning of the hidden rule. The feedback node uses the sigmoid function as the fuzzy membership function to achieve the fuzzification of the internal variable. The membership function layer of the HRFNN network uses a local membership function. The difference is that the feedback part uses a global membership function on the domain of the internal variable to simplify the network structure and realize the feedback of global historical information. The number of receiving nodes is equal to the number of feedback nodes; the number of receiving nodes is equal to the number of nodes in the rule layer. The feedback is connected to the third layer as the input of the fuzzy rule layer. The output of the feedback node contains the historical information of the activation strength of the fuzzy rule.

[0091] Layer IV: Defuzzification layer, also known as output layer. The nodes in this layer perform summation operations on the input quantities. The input and output of the network are expressed as:

[0092]

[0093] In the formula, jThey are the connection weights of the output layer. The fuzzy recurrent neural network has the performance of approximating highly nonlinear dynamic systems. The training error and testing error of the fuzzy recurrent neural network with internal variables added are significantly reduced respectively. The prediction effect of this network is better than that of the fuzzy recurrent neural network with self-feedback and the fuzzy neural network for dynamic modeling. This shows that the learning ability of the network is enhanced after adding internal variables and it can more fully reflect the dynamic characteristics of the sewage treatment system. The simulation results prove the effectiveness of the network. The fuzzy recurrent neural network HRFNN of this patent uses the gradient descent algorithm with cross-validation to train the weights of the neural network. The HRFNN is used to predict the material weight ratio parameter. The HRFNN introduces internal variables in the feedback link, weights and sums the output of the rule layer and then defuzzifies the output as the feedback quantity, and takes the feedback quantity and the output of the membership function layer together as the input of the next moment of the rule layer. The network output contains the activation intensity of the rule layer and the historical information of the output, enhancing the ability of the HRFNN to adapt to nonlinear dynamic systems. Experiments show that the HRFNN can accurately predict the material weight ratio parameter. Comparing the simulation results with those obtained by other networks, the model established by the method of this patent has the smallest network scale and the smallest prediction error when applied to the prediction of the material weight ratio, indicating the effectiveness of this method.

[0094] 6. Design of the NARX Neural Network Model

[0095] The output of the fuzzy recurrent neural network model is used as the input of the NARX neural network model, the output of the NARX neural network model is used as the input of the time-delay neural network model, and the output value of the NARX neural network model is used as the material weight ratio of the fish feed formula; the NARX neural network model is a dynamic recurrent neural network with output feedback connection. In terms of topological connection relationship, it can be equivalently regarded as a BP neural network with input time delay plus a time-delay feedback connection from output to input. Its structure consists of an input layer, a time-delay layer, a hidden layer and an output layer. Among them, the input layer nodes are used for signal input, the time-delay layer nodes are used for the time delay of the input signal and the output feedback signal, the hidden layer nodes perform nonlinear operations on the time-delayed signal using the activation function, and the output layer nodes are used to linearly weight the output of the hidden layer to obtain the final network output. The NARX neural network has characteristics such as nonlinear mapping ability, good robustness and self-adaptability, and is suitable for predicting the material weight ratio of fish feed. x(t) represents the external input of the neural network, that is, the output value of the fuzzy recurrent neural network model; m represents the delay order of the external input; y(t) is the output of the neural network, that is, the predicted value of the material weight ratio in the next period; n is the output delay order; s is the number of neurons in the hidden layer; thus, the output of the jth hidden unit can be obtained as follows:

[0096]

[0097] In the above formula, w jiis the connection weight between the i-th input and the j-th hidden neuron, and b j is the bias value of the j-th hidden neuron. The value of the network output y(t + 1) is:

[0098] y(t + 1) = f[y(t), y(t - 1), …, y(t - n), x(t), x(t - 1), …, x(t - m + 1); W] (17)

[0099] The NARX neural network model of this invention patent is a dynamic feedforward neural network. The NARX neural network is a non-linear autoregressive network with external inputs, which is the output of a fuzzy recurrent neural network model. It has a dynamic characteristic of multi-step time delay and is a closed network with several layers that connect the output value to the network input through feedback material weight ratio. The NARX neural network model is one of the most widely used dynamic neural networks in non-linear dynamic systems, and its performance is generally better than that of the full regression neural network. A typical NARX regression neural network mainly consists of an input layer, a hidden layer, an output layer, and input and output time delays. Generally, the time delay order of the input and output and the number of hidden neurons need to be determined in advance before application. The current output material weight ratio of the NARX neural network model not only depends on the output material weight ratio at the past y(t - n) moment, but also depends on the output of the current fuzzy recurrent neural network model as the input vector X(t) and the time delay order of the input vector, etc. Among them, the output of the fuzzy recurrent neural network model as the input signal is transmitted to the hidden layer through the time delay layer. The hidden layer processes the input signal and then transmits it to the output layer. The output layer linearly weights the output signal of the hidden layer to obtain the final output material weight ratio of the NARX neural network model. The time delay layer delays the signal feedback from the output material weight ratio of the NARX neural network model and the output of the fuzzy recurrent neural network model as the signal output by the input layer, and then transports it to the hidden layer.

[0100] 7. Time Delay Neural Network Model

[0101] The output of the NARX neural network model is used as the input of the time-delay neural network model, and the output of the time-delay neural network model is used as the corresponding input of the fuzzy recurrent neural network model. A time-delay neural network (TDNN neural network) is an adaptive linear network. Its input enters from the left side of the network. Through the action of a single-step delay line D, after d steps of delay, it becomes the input of a d+1-dimensional vector, which is composed of the signals output by the NARX neural network model at the current K moments and the signals output by the NARX neural network model d-1 moments before K. The neurons adopt linear activation functions, and the time-delay neural network belongs to a variant of traditional artificial neural networks. The structure of the time-delay neural network consists of an input layer, an output layer, and one or several hidden layers, and a mapping relationship between "input-output" is established by the neural network. Different from traditional neural networks, the time-delay neural network realizes the memory of previous inputs by delaying the input at the input layer. By delaying the input at the input layer, the network can use the previous d steps of input and the current input to jointly predict the output at the current time point. For a time-delay neural network with a delay step of d at the input layer, R is the forward propagation operator of the time-delay neural network. The relationship between the input sequence X and the output sequence Y can be simply expressed in the following form:

[0102] Y(t) = R(X(t), X(t - 1), …, X(t - d)) (18)

[0103] 8. Design of Feed-Weight Ratio Trend Prediction Module

[0104] The feed-weight ratio trend prediction module includes a CNN convolutional neural network model, an LSTM neural network model, an ARIMA model, and a NARX neural network model. The historical data of the fish feed-weight ratio are used as the inputs of the CNN convolutional neural network model and the ARIMA model respectively. The outputs of the CNN convolutional neural network model and the ARIMA model are used as the input of the NARX neural network model. The output value of the NARX neural network model is used as the output of the feed-weight ratio trend prediction module; the feed-weight ratio trend prediction module is shown in Figure 2 . The designs of the CNN convolutional neural network model, the LSTM neural network model, and the NARX neural network model respectively refer to the above design process. The design process of the ARIMA model is as follows:

[0105] The ARIMA model is a modeling method for predicting the feed-to-weight ratio based on time series proposed by Box et al., which can be extended to analyze the time series of the predicted feed-to-weight ratio. In this patent, for the study of the time series characteristics of the feed-to-weight ratio of the ARIMA model, three parameters are used to analyze the time series of the feed-to-weight ratio changes, namely the autoregressive order (p), the number of differences (d), and the moving average order (q). The ARIMA model is written as: ARIMA(p, d, q). The ARIMA model equation with p, d, and q as parameters can be expressed as follows:

[0106]

[0107] Δ d y t represents y t the sequence after d - times of difference transformation, ε t is the random error at time t, which is an independent white noise sequence and follows a normal distribution with a mean of 0 and a constant variance of σ 2 φ i (i = 1, 2, …, p) and θ j (j = 1, 2, …, q) are the parameters to be estimated in the ARIMA model, and p and q are the orders of the ARIMA dynamic prediction feed-to-weight ratio model. The ARIMA dynamic prediction feed-to-weight ratio model essentially belongs to a linear model, and the modeling and prediction include four steps: (1) Sequence stationary processing. If the feed-to-weight ratio data sequence is non-stationary, such as having a certain growth or decline trend, etc., then the data needs to be differenced. Commonly used tools are the autocorrelation function graph and the partial autocorrelation function graph. If the autocorrelation function rapidly approaches zero, then the feed-to-weight ratio time series is a stationary time series. If the time series has a certain trend, then the feed-to-weight ratio data needs to be differenced. If there is a seasonal pattern, seasonal differencing is also required. If the time series has heteroscedasticity, then the feed-to-weight ratio data needs to be logarithmically transformed first. (2) Model identification. The orders p, d, and q of the ARIMA dynamic prediction feed-to-weight ratio model are mainly determined through the autocorrelation coefficient and the partial autocorrelation coefficient. (3) Estimate the parameters of the model and model diagnosis. The estimated values of all parameters in the ARIMA dynamic prediction feed-to-weight ratio model are obtained by maximum likelihood estimation, and tests including the significance test of the parameters and the randomness test of the residuals are carried out, and then it is judged whether the established feed-to-weight ratio model is acceptable. Use the ARIMA dynamic prediction feed-to-weight ratio model with appropriate parameters to predict the feed-to-weight ratio; and conduct tests in the model to determine whether the model is appropriate. If it is not appropriate, re-estimate the parameters. (4) Use the feed-to-weight ratio model with appropriate parameters to predict the change trend of the feed-to-weight ratio.

[0108] 9. Design of the environmental assessment module

[0109] The environmental assessment module includes multiple LSTM neural network models, multiple auto-associative neural network models, and a NARX neural network model. The outputs of multiple groups of temperature, dissolved oxygen, and pH value sensors are respectively used as the inputs of the corresponding multiple LSTM neural network models. The outputs of the multiple LSTM neural network models are respectively used as the inputs of each auto-associative neural network model. The outputs of the multiple auto-associative neural network models are used as the input of the NARX neural network model. The output value of the NARX neural network model is used as the output of the environmental assessment module; The environmental assessment module is shown in Figure 2 . The design process of the auto-associative neural network model is as follows: The auto-associative neural network model (Auto-associative neural networ, AANN), a feedforward neural network with a special structure, and the structure of the auto-associative neural network model includes an input layer, a certain number of hidden layers, and an output layer. First, the data information output by multiple LSTM neural network models is compressed through the input layer, mapping layer, and bottleneck layer of the auto-associative neural network model. The most representative low-dimensional subspace reflecting the structure of the aquaculture environment assessment level system is extracted from the high-dimensional parameter space output by multiple LSTM neural network models. At the same time, the noise and measurement errors in the input data of the aquaculture environment assessment level are effectively filtered out. Then, the decompression of the aquaculture environment assessment level data is realized through the bottleneck layer, demapping layer, and output layer, and the previously compressed information is restored to each parameter value, so as to realize the reconstruction of the input data of each aquaculture environment assessment level. In order to achieve the purpose of compressing the aquaculture environment assessment level information, the number of nodes in the bottleneck layer of the auto-associative neural network model is significantly smaller than that of the input layer. In order to prevent the formation of a simple single mapping between the input and output layers of the aquaculture environment assessment level, except that the activation function of the output layer uses a linear function, other layers all use non-linear activation functions. Essentially, the first layer of the hidden layer of the auto-associative neural network model is called the mapping layer, and the node transfer function of the mapping layer may be an S-type function or other similar non-linear functions; The second layer of the hidden layer is called the bottleneck layer, and the dimension of the bottleneck layer is the smallest in the network. Its transfer function may be linear or non-linear. The bottleneck layer avoids the one-to-one mapping relationship where the output and input are easily equal, and it enables the auto-associative neural network model to encode and compress the aquaculture environment assessment level signal to obtain a related model of the input aquaculture environment assessment level, and perform aquaculture environment assessment level decoding and decompression after the bottleneck layer to generate an estimated value of the aquaculture environment assessment level input signal; The third or last layer of the hidden layer is called the demapping layer, and the node transfer function of the demapping layer is usually a non-linear S-type function. The auto-associative neural network uses the error backpropagation algorithm for training.

[0110] (2). Design of the Firefly Algorithm Optimization Subsystem for Fish Feed Formulation

[0111] The firefly algorithm optimization subsystem for fish feed formula mainly includes five steps: initialization of the firefly population, determination of fluorescence brightness, calculation of the objective function, update of the firefly position, and determination of the optimal fish feed formula. The firefly population is randomly distributed in the search space as the initial solution. Each firefly is regarded as a fish feed formula, and it will be attracted by fireflies that are brighter than it. The attractiveness of the firefly is positively correlated with the brightness. For any two fireflies, one firefly will move towards the other firefly that is brighter than it. The brightness decreases with the increase of the distance, and the fireflies gather around the fireflies with high brightness. The optimization process of the firefly algorithm optimization subsystem for fish feed formula is shown in Figure 3 , and the optimization process of the firefly algorithm optimization subsystem for fish feed formula is as follows:

[0112] 1. Initialization of the firefly population

[0113] Suppose the fish feed formula consists of rice bran, wheat bran, soybean cake, fish meal, and yeast powder, and the content of each material in each fish feed formula has a certain range, which is used as the constraint condition of the fish feed formula. Only when the content of each material is within this range can it meet the nutritional needs of fish growth. The limitation range of the content of each material (unit: kg) in each fish feed formula is as follows: rice bran is [35, 45], wheat bran is [35, 45], soybean cake is [9, 13], fish meal is [9, 13], and yeast powder is [1.5, 3]. According to the constraint conditions of the fish feed formula, randomly generate 5 real numbers representing the 5 materials in the fish feed formula. They are arranged together to form a firefly individual of the fish feed formula, which is defined as a 5-dimensional search space of a firefly. They are B, C, D, E, and F respectively. Continuously generate M such firefly individuals of the fish feed formula. M is the size of the firefly population, that is, the number of firefly individuals of the fish feed formula in each generation. Suppose the number of fireflies is M, and the spatial position of the i-th firefly is A i = [B i , C i , D i , E i , F i ; The light intensity absorption coefficient is γ, the maximum attractiveness is β0, and the step size factor is α; The maximum and minimum weights are w max , w min , and the maximum number of update generations is S max . Randomly distribute the fireflies in the 5-dimensional space, and set the distance d ij between the i-th firefly and the j-th firefly as the Euclidean distance, that is, the following formula:

[0114]

[0115] 2. Determine the fluorescence brightness

[0116] The fluorescence brightness function is used to evaluate the quality of firefly individuals in fish feed formulations, serving as the basis for survival of the fittest in the optimization process of fish feed formulations. The optimization of fish feed formulations involves optimizing the combination of material contents in the fish feed formulation, striving for the best benefit of the fish feed formulation on the premise of meeting the fish feed nutrition standards; the firefly individuals of each fish feed formulation are used as the input of the fish feed formulation feed weight ratio prediction subsystem, and the output of the fish feed formulation feed weight ratio prediction subsystem is used as the predicted value of the feed weight ratio of the firefly individuals of the feed formulation. The reciprocal of the predicted value of the feed weight ratio of the firefly individuals of each fish feed formulation is used as the fluorescence brightness of the firefly individuals of the fish feed formulation. The larger the reciprocal of the predicted value of the feed weight ratio of the firefly individuals of each fish feed formulation, the higher the fluorescence brightness of the firefly individuals of the fish feed formulation. The sum of the reciprocals of the predicted values of the feed weight ratios of the firefly individuals of each fish feed formulation in the population is the total fluorescence brightness of the population.

[0117] 3. Objective function calculation

[0118] If the calculation reaches the specified maximum number of generations, the firefly algorithm stops after reaching it; or the difference between the fluorescence brightness of the firefly individuals of the fish feed formulation and the fluorescence brightness of the firefly individuals of the specified fish feed formulation is less than the set threshold, which is used as the objective function of the firefly algorithm, and the firefly individuals of the fish feed formulation are the optimal solutions.

[0119] 4. Firefly position update

[0120] Fireflies move in the direction of fireflies with higher fluorescence brightness than themselves, and the degree of attraction between them determines the moving step size of the fireflies. The degree of attraction between fireflies i and j is:

[0121]

[0122] From formula (21), the position update formula of the fish feed formulation fireflies can be further obtained as:

[0123]

[0124] Where: R is a random number uniformly distributed on [0, 1]; S is the current update generation of the fish feed formulation fireflies. And check one by one whether the corresponding new fish feed formulation firefly individuals meet the constraint conditions of the fish feed formulation. If they meet, the new fish feed formulation firefly individuals are used as new generation members; otherwise, the new fish feed formulation firefly individuals are discarded.

[0125] 5. Determination of the optimal fish feed formulation

[0126] After meeting the stop conditions of the firefly algorithm, calculate the fluorescence brightness of each firefly individual of the fish feed formulation. The firefly individual of the feed formulation with the highest fluorescence brightness is the optimal firefly individual of the fish feed formulation, and the optimal fish feed formulation is obtained.

[0127] VIII. Design Example of a Fish Feed Detection and Formulation System

[0128] According to the actual situation of the big data detection system for aquaculture environment, the system arranges a plane layout installation drawing of the aquaculture parameter acquisition platform and the detection nodes, control nodes, gateway nodes, and on-site monitoring terminals for control. Among them, the sensors of the detection nodes are evenly arranged in all directions of the aquaculture pond according to the detection needs, and the aquaculture environment parameters are collected through this system.

[0129] The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A fish feed detection and formulation system, characterized in that: The system includes an aquaculture environment parameter acquisition and control platform, a fish feed formula weight ratio prediction subsystem, and a fish feed formula firefly algorithm optimization subsystem, which can detect the aquaculture environment parameters, predict the weight ratio of the fish feed formula, and optimize the fish feed formula; The fish feed formula weight ratio prediction subsystem includes a fuzzy C-means clustering algorithm, a CNN convolutional neural network model, an LSTM neural network model, a GRNN neural network model, a fuzzy recurrent neural network model, a NARX neural network model, a time-delay neural network model, a weight ratio trend prediction module, and an environmental evaluation module. The fish feed formula is used as the input of the fuzzy C-means clustering algorithm. The fish feed formulas of multiple categories output by the fuzzy C-means clustering algorithm are respectively used as the inputs of the corresponding multiple CNN convolutional neural network models. The outputs of the multiple CNN convolutional neural network models are respectively used as the inputs of the corresponding multiple LSTM neural network models. The outputs of the multiple LSTM neural network models are used as the input of the GRNN neural network model. The outputs of the GRNN neural network model, the time-delay neural network model, the weight ratio trend prediction module, and the environmental evaluation module are used as the input of the fuzzy recurrent neural network model. The output of the fuzzy recurrent neural network model is used as the input of the NARX neural network model. The output of the NARX neural network model is used as the input of the time-delay neural network model. The output value of the NARX neural network model is used as the weight ratio of the fish feed formula; The fish feed formula firefly algorithm optimization subsystem includes five links: firefly population initialization, determination of fluorescence brightness, objective function calculation, firefly position update, and determination of the optimal fish feed formula. The firefly population is randomly distributed in the search space as the initial solution. Each firefly is regarded as a fish feed formula. It will be attracted by fireflies that are brighter than it. The attractiveness of fireflies is positively correlated with their brightness. For any two fireflies, one firefly will move towards another firefly that is brighter than it. The brightness decreases with the increase of distance, and the fireflies gather around the fireflies with high brightness; The firefly individual of each fish feed formula is used as the input of the fish feed formula weight ratio prediction subsystem. The output of the fish feed formula weight ratio prediction subsystem is used as the predicted value of the weight ratio of the firefly individual of this fish feed formula. The reciprocal of the weight ratio predicted value of the firefly individual of each fish feed formula is used as the fluorescence brightness of the firefly individual of this fish feed formula. The larger the reciprocal of the weight ratio predicted value of the firefly individual of each fish feed formula, the higher the fluorescence brightness of the firefly individual of this fish feed formula. The sum of the reciprocals of the weight ratio predicted values of the firefly individuals of each fish feed formula in the firefly population is the total fluorescence brightness of this firefly population; The feed conversion ratio trend prediction module includes a CNN convolutional neural network model, an LSTM neural network model, an ARIMA model, and a NARX neural network model. The historical data of the feed conversion ratio of fish feed is used as the input of the CNN convolutional neural network model and the ARIMA model respectively. The outputs of the CNN convolutional neural network model and the ARIMA model are used as the input of the NARX neural network model. The output value of the NARX neural network model is used as the output of the feed conversion ratio trend prediction module. The environmental evaluation module includes multiple LSTM neural network models, multiple autoassociative neural network models, and a NARX neural network model. The outputs of multiple groups of temperature, dissolved oxygen, and pH value sensors are used as the inputs of the corresponding multiple LSTM neural network models respectively. The outputs of the multiple LSTM neural network models are used as the inputs of each autoassociative neural network model respectively. The outputs of the multiple autoassociative neural network models are used as the input of the NARX neural network model. The output value of the NARX neural network model is used as the output of the environmental evaluation module.

2. The fish feed detection and formulation system according to claim 1, characterized in that: The aquaculture environment parameter acquisition and control platform consists of detection nodes, control nodes, gateway nodes, on-site monitoring terminals, cloud platforms, and remote monitoring computers.

3. The fish feed detection and formulation system according to claim 2, characterized in that: The detection nodes collect the fish farming environment parameters and upload them to the cloud platform through the gateway nodes. The cloud platform provides the fish farming environment parameters to the remote monitoring computer for Web visualization of the fish farming environment parameter interface management. The remote monitoring computer issues commands to the control nodes to implement remote environmental control. Data is stored and information is published at the cloud platform end. The detection nodes and control nodes are responsible for collecting the fish farming environment parameters and controlling the fish farming environment equipment. Bidirectional communication among the detection nodes, control nodes, on-site monitoring terminals, cloud platforms, and remote monitoring computers is achieved through the gateway nodes, realizing the acquisition of fish farming environment parameters and the control of fish farming equipment.

Images