CN113255739A - Fish feed detection and formula system - Google Patents

Fish feed detection and formula system Download PDF

Info

Publication number
CN113255739A
CN113255739A CN202110496297.5A CN202110496297A CN113255739A CN 113255739 A CN113255739 A CN 113255739A CN 202110496297 A CN202110496297 A CN 202110496297A CN 113255739 A CN113255739 A CN 113255739A
Authority
CN
China
Prior art keywords
neural network
fish feed
network model
weight ratio
firefly
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202110496297.5A
Other languages
Chinese (zh)
Inventor
陈佳豪
马从国
丁晓红
王苏琪
张庆宇
肖炳宇
马海波
张利兵
金德飞
刘伟
周恒瑞
王建国
陈亚娟
李亚洲
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huaiyin Institute of Technology
Original Assignee
Huaiyin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huaiyin Institute of Technology filed Critical Huaiyin Institute of Technology
Priority to CN202110496297.5A priority Critical patent/CN113255739A/en
Publication of CN113255739A publication Critical patent/CN113255739A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/02—Agriculture; Fishing; Mining

Abstract

The invention discloses a fish feed detection and formula system, which comprises an aquaculture environment parameter acquisition and control platform, a fish feed formula material weight ratio prediction subsystem and a fish feed formula firefly algorithm optimization subsystem, and realizes detection of fish culture environment parameters, prediction of fish feed formula material weight ratio and optimization of fish feed formula; the invention effectively solves the problems that the conventional fish feed formula has no influence on the benefits of the aquaculture environment according to the nonlinearity and large hysteresis of the change of the parameters of the aquaculture environment, the complex aquaculture adjustment and the like, and the feed-to-weight ratio of the fish feed formula is not predicted and the fish feed formula is not optimized, so that the benefits of the aquaculture and the production management are greatly influenced.

Description

Fish feed detection and formula system
Technical Field
The invention relates to the technical field of fish feed detection and formula automation, in particular to a fish feed detection and formula system.
Background
Aquaculture is an important part of agriculture and is a mark of agricultural modernization. With the market demand, the fish price and the farmer, the fish feed market is becoming active, the experience of the farmer shows that the quality and the price of the fish feed formula become one of the key links of the aquaculture benefit, the research on the reasonable formula of the fish feed has very important practical significance in improving the aquaculture. A reasonable fish feed formulation should meet the needs of fish growth, development and production while avoiding the waste of any one or more nutrients and needs to be achieved according to the fish nutritional needs and a formulation optimization system to achieve the goal of obtaining a minimum cost formulation while meeting the fish nutritional needs. The fish feed detection and formula system is a visual operation system, has the characteristics of convenient operation, simplicity, comprehensiveness, friendly interface design, complete indexes and the like, and reduces the waste of nutrients as much as possible. The system fully considers the aquaculture environment and the aquaculture conditions in the design process so as to meet the balanced supply of nutrients, improve the economic benefit of aquaculture and reduce the aquaculture cost.
Disclosure of Invention
The invention provides a fish feed detection and formula system, which effectively solves the problems that the conventional fish feed formula does not influence the benefits of aquaculture environment according to the nonlinearity and large hysteresis of the change of parameters of the aquaculture environment, the complex aquaculture regulation and the like, and the feed-weight ratio of the fish feed formula is not predicted and the fish feed formula is not optimized, so that the benefits of aquaculture and production management are greatly influenced.
The invention is realized by the following technical scheme:
a fish feed detection and formula system comprises an aquaculture environment parameter acquisition and control platform, a fish feed formula material weight ratio prediction subsystem and a fish feed formula firefly algorithm optimization subsystem, and realizes detection of fish culture environment parameters, prediction of fish feed formula material weight ratio and optimization of fish feed formula.
The invention further adopts the technical improvement scheme that:
the aquaculture environment parameter acquisition and control platform consists of a detection node, a control node, a gateway node, an on-site monitoring end, a cloud platform and a remote monitoring computer, wherein the detection node acquires fish culture environment parameters and uploads the fish culture environment parameters to the cloud platform through the gateway node, the cloud platform provides the fish culture environment parameters to the remote monitoring computer for Web visual fish culture environment parameter interface management, the remote monitoring computer implements environment remote control on instructions under the control node, and data and release information are stored at the cloud platform end; the detection nodes and the control nodes are responsible for collecting fish culture environment parameters and controlling fish culture environment equipment, and the bidirectional communication among the detection nodes, the control nodes, the field monitoring end, the cloud platform and the remote monitoring computer is realized through the gateway nodes, so that the fish culture environment parameter collection and the fish culture equipment control are realized; the aquiculture environment parameter acquisition and control platform is shown in figure 1.
The invention further adopts the technical improvement scheme that:
the fish feed formula material weight ratio prediction subsystem comprises a fuzzy C-means clustering algorithm, a plurality of CNN convolutional neural network models, a plurality of LSTM neural network models, a GRNN neural network model, a fuzzy recurrent neural network model, an NARX neural network model, a time delay neural network model, a material weight ratio trend prediction module and an environment evaluation module, wherein a fish feed formula is used as the input of the fuzzy C-means clustering algorithm, a plurality of categories of fish feed formulas output by the fuzzy C-means clustering algorithm are respectively used as the input of the corresponding CNN convolutional neural network models, the outputs of the CNN convolutional neural network models are respectively used as the input of the corresponding LSTM neural network models, the outputs of the LSTM neural network models are used as the input of the GRNN neural network model, and the outputs of the GRNN neural network model, the time delay neural network model, the material 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, and the output value of the NARX neural network model is used as the material weight ratio of the fish feed formula; the fish feed formula feed weight ratio prediction subsystem is shown in figure 2.
The invention further adopts the technical improvement scheme that:
the material weight ratio trend prediction module comprises a CNN convolutional neural network model, an LSTM neural network model, an ARIMA model and a NARX neural network model, fish feed material weight ratio historical data are respectively used as the input of the CNN convolutional neural network model and the ARIMA model, the output of the CNN convolutional neural network model and the ARIMA model is used as the input of the NARX neural network model, and the output value of the NARX neural network model is used as the output of the material weight ratio trend prediction module; the material weight ratio trend prediction module is shown in figure 2.
The invention further adopts the technical improvement scheme that:
the environment evaluation module comprises a plurality of LSTM neural network models, a plurality of self-association neural network models and an NARX neural network model, the outputs of a plurality of groups of temperature, dissolved oxygen and PH value sensors are respectively used as the inputs of the corresponding plurality of LSTM neural network models, the outputs of the plurality of LSTM neural network models are respectively used as the input of each self-association neural network model, the outputs of the plurality of self-association neural network models are used as the inputs of the NARX neural network models, and the output value of the NARX neural network model is used as the output of the environment evaluation module; the environmental assessment module is shown in figure 2.
The invention further adopts the technical improvement scheme that:
the fish feed formula firefly algorithm optimization subsystem comprises 5 links including firefly population initialization, fluorescence brightness determination, objective function calculation, firefly position updating and optimal fish feed formula determination, the firefly populations are randomly distributed in a search space as an initial solution, each firefly is regarded as a fish feed formula and attracted by the firefly brighter than the firefly population, the attraction force of the firefly is positively correlated with the brightness, for any two fireflies, one firefly moves towards the other firefly brighter than the firefly population, the brightness decreases along with the increase of the distance, and the fireflies gather to the periphery of the firefly with high brightness; the firefly individual of each fish feed formula is used as the input of the fish feed formula material weight ratio prediction subsystem, the output of the fish feed formula material weight ratio prediction subsystem is used as the predicted value of the firefly individual material weight ratio of the fish feed formula, the reciprocal of the predicted value of the firefly individual material weight ratio of each fish feed formula is used as the fluorescence brightness of the firefly individual of the fish feed formula, the larger the reciprocal of the predicted value of the firefly individual material weight ratio of each fish feed formula is, the higher the fluorescence brightness of the firefly individual of the fish feed formula is, and the sum of the reciprocals of the predicted values of the material weight ratios of the firefly individuals of each fish feed formula in the firefly population is the total fluorescence brightness of the firefly population. The fish feed formulation firefly algorithm optimization subsystem is shown in figure 3.
Compared with the prior art, the invention has the following obvious advantages:
firstly, the invention can realize the feature extraction of fish feed formula and fish feed weight ratio historical data by utilizing a CNN convolutional neural network model, shorten the feature extraction time, and memorize the advantage of the relationship between the fish feed formula and the fish feed weight ratio historical data in the fish culture process with strong dependency by utilizing an LSTM neural network model, thereby solving the problems of the spatial feature extraction and the time feature data dependency of the fish feed formula and fish feed weight ratio historical data activity sequence data; firstly, inputting sequence data preprocessed by a fish feed formula and fish feed weight ratio historical data into a CNN convolutional neural network model to extract corresponding spatial feature vectors; and secondly, taking different activity space characteristic vectors of historical data of the fish feed formula and the fish feed weight ratio extracted in the previous step as input of an LSTM neural network model, and processing the problem of time characteristic interaction among activity sequence data of the predicted fish feed formula weight ratio by utilizing the data interaction of an input gate, a forgetting gate and an output gate in the LSTM neural network model, so that the accuracy and the time efficiency of the predicted fish feed formula weight ratio are improved.
Secondly, extracting high-dimensional spatial characteristics of the fish feed formula and fish feed material weight ratio historical data by using a CNN convolutional neural network model to realize characteristic extraction of the fish feed formula and fish feed material weight ratio historical data; meanwhile, an LSTM neural network model is selected to process a spatial feature sequence output by the CNN convolutional neural network model, time sequence information in fish feed formula and fish feed material weight ratio historical data is mined, time features of the fish feed formula and the fish feed material weight ratio historical data are extracted in a time dimension, and accurate prediction of the fish feed formula material weight ratio is achieved.
Thirdly, the convolutional layer of the CNN convolutional neural network model has the main advantages that weight sharing and sparse connection in historical data of fish feed formula and fish feed weight ratio are extracted, the weight sharing means that the weight of a convolutional kernel of the CNN convolutional neural network model is kept unchanged when convolution operation is carried out, and the weight of each convolutional kernel is the same as that of the historical data of the fish feed formula and the fish feed weight ratio in the whole area; sparse connection means that each convolution kernel of the CNN convolution neural network model only uses specific local area data in the data of the upper layer to carry out operation, and does not use the overall fish feed formula and fish feed weight ratio historical data; the weight sharing and sparse connection characteristics of the convolution kernel of the CNN convolutional neural network model greatly reduce the number of spatial characteristic parameters of the fish feed formula and the fish feed weight ratio historical data, so that overfitting of the CNN convolutional neural network model is prevented, the training speed of the CNN convolutional neural network model is increased, and the prediction accuracy of the fish feed formula is improved.
The LSTM neural network model is similar to a standard network containing a recursion hidden layer, the only change is that a memory module is used for replacing an original hidden layer unit, the problems of gradient disappearance and sharp increase are solved through self-feedback of the internal state of a memory cell and truncation of errors of input and output, compared with a BP neural network and a common RNN, the LSTM adds 1 state unit c and 3 control gates, the feature inclusion capacity and the memory capacity of the model are greatly increased, and under-fitting and gradient disappearance are avoided. The function of the LSTM neural network model is to learn the relationships and changes in the relationships over time that exist in fish feed formulations, fish feed weight ratio history data and aquaculture environment data, and to obtain more accurate results. The LSTM neural network model realizes the prediction of the feed weight ratio of the fish feed formula and the water quality parameter grade of the aquaculture pond environment, and improves the prediction accuracy.
And fifthly, the LSTM neural network model has a chain-like repeating network structure similar to that of the standard RNN, the repeating network in the standard RNN is very simple, and the repeating network in the LSTM neural network model has 4 interaction layers comprising 3 gate layers and 1 tanh layer. Processor state is a key variable in the LSTM neural network model that carries information from the previous step of the material-to-weight ratio prediction and steps through the entire LSTM neural network model. The gates in the interaction layer may partially delete the processor state of the previous step and add the fuel-to-weight ratio prediction new information into the processor state of the current step based on the hidden state of the previous step and the input of the current step. The inputs to each repeating network include the implicit and processor states of the previous step fuel-to-weight ratio prediction and the input of the current step. The processor state is updated according to the calculation results of the 4 interaction layers. The updated processor state and hidden state constitute the output and are passed on to the next step.
Sixthly, the LSTM neural network model is a recurrent neural network with 4 interaction layers in a repetitive network. It can not only extract information from the material weight ratio prediction sequence data like a standard recurrent neural network, but also retain information with long-term correlation from a previous distant step. The material weight ratio prediction data are sequence data, and the change trend of the sequence data is rich in meaning. In addition, since the sampling interval of the material weight ratio prediction is relatively small, the material weight ratio prediction has long-term spatial correlation, and the LSTM neural network model has enough long-term memory to deal with the problem.
In the cascade LSTM neural network model, firstly, material weight ratio data which are easy to predict are reconstructed at a shallow level, and then the generated material weight ratio data are used as input of the next level. The deep-level prediction result is not only based on the input value in the material-to-weight ratio data training data, but also influenced by the shallow-level material-to-weight ratio data result, the method can more effectively extract the information contained in the material-to-weight ratio data input data, and the accuracy of the model for predicting the material-to-weight ratio data is improved.
Eighthly, the method combines the technologies of fuzzy C-means clustering (FCM), CNN convolutional neural network model, LSTM neural network model and GRNN neural network model, and is applied to the material-to-weight ratio measurement of the fish feed formula. Firstly, classifying fish feed formula samples by adopting an FCM method, then establishing a local measurement model of the material weight ratio of the fish feed formula by using a CNN convolutional neural network model and an LSTM neural network model in series, and fusing the outputs of a plurality of local models through a GRNN neural network model, wherein the results show that the fish feed material weight ratio measurement model established by the method has better training speed and higher measurement precision.
The invention adopts a dynamic recursive network of an NARX neural network model established through a time delay module and feedback realization of the feed ratio of a feed formula, and the dynamic recursive network is realized by a sequence of a plurality of time material ratio parameters expanded along the time axis direction of the material ratio parameters and a data relevance modeling idea of a function simulation function. The input comprises a feed ratio input and an output historical feedback of a feed formula for a period of time, the feedback input can be considered to comprise historical information of the feed ratio state of the fish feed for the period of time to participate in the prediction of the feed ratio of the fish, and the prediction has good effect on a proper feedback time length.
The invention utilizes NARX neural network to establish material weight ratio prediction model, because of introducing the dynamic recursive network of delay module and output feedback establishment model, it introduces the input and output vector delay feedback into the network training, forms new input vector, has good nonlinear mapping ability, the input of the network model not only includes the original input data, but also includes the output data after training, the generalization ability of the network is improved, make it have better prediction accuracy and adaptive ability than the traditional static neural network in the nonlinear material weight ratio time series prediction.
The GRNN neural network model adopted by the invention has simple and complete structure, the internal structure of the model is determined along with the determination of the sample points, the requirement on data samples is less, and the data can be converged on a regression surface even if the data is rare as long as people are input and the samples are output. The method has the characteristics of definite probability significance, better generalization capability, local approximation capability and quick learning, can approximate functions of any healing type, and finally determines the model only by adjusting and selecting the smooth factor in the process of establishing and learning the network model. The network establishing process is the network training process, and special training is not needed. On the fusion effect of a dynamic system, the GRNN neural network model has the characteristics of simple network establishment process, few influence factors, strong local approximation capability, high learning speed and good simulation performance. Therefore, the GRNN neural network model is well suited for material-to-weight ratio fusion. The GRNN neural network model is used for carrying out material-weight ratio fusion by utilizing the characteristics that the GRNN neural network model has self-adaptability, self-learning, nonlinear approximation with any precision and the like, so that the robustness and the fault tolerance of the material-weight ratio fusion model are better met.
The GRNN neural network model adopted by the method has strong nonlinear mapping capability, a flexible network structure, high fault tolerance and robustness, and is suitable for material-to-weight ratio fusion. The GRNN neural network model has stronger advantages than an RBF network in approximation ability and learning speed, the network finally converges on an optimized regression surface with more sample size accumulation, and when the sample data is less, the network can also process unstable data, and the fusion material weight ratio effect is better. The GRNN neural network model has the advantages of strong generalization capability, high fusion precision and stable algorithm, the GRNN neural network model also has the advantages of high convergence speed, few adjustment parameters, difficulty in falling into local minimum values and the like, the network fusion operation speed is high, and the GRNN neural network model has a good application prospect on material-to-weight ratio fusion.
Thirteen, the invention adopts an ARIMA model to obey time sequence distribution based on the raw data of material-to-weight ratio fusion, utilizes the principle that the fusion change of the material-to-weight ratio has certain inertia trend, integrates the raw time sequence variable of the material-to-weight ratio fusion of the factors such as trend factor, periodic factor, random error and the like, converts a non-stationary sequence into a stationary random sequence with zero mean value by methods such as differential data conversion and the like, and performs material-to-weight ratio fusion numerical value fitting and prediction by repeated identification, model diagnosis and comparison and selection of an ideal model. The method combines the advantages of autoregressive and moving average methods, has the characteristics of no data type constraint and strong applicability, and is a model with better short-term prediction material-to-weight ratio fusion effect.
Fourteen, the invention adopts the fuzzy recurrent neural network structure, and introduces the internal variable in the fuzzy rule layer, so that the static network has dynamic characteristics; the activation degree of each rule of the network at the K moment not only comprises the activation degree value calculated by the current input, but also comprises the contribution of all the rule activation degree values at the previous moment, so that the accuracy of network identification is improved, and the material-to-weight ratio prediction can be well completed. The fuzzy recursion neural network model is used for establishing a prediction model of the material weight ratio, is a typical dynamic recursion neural network, and the feedback connection of the fuzzy recursion neural network model is composed of a group of 'structure' units and is used for memorizing the past state of a hidden layer, and the feedback connection and the network input are used as the input of the hidden layer unit at the next moment.
Drawings
FIG. 1 is an aquaculture environment parameter acquisition and control platform of the present patent;
FIG. 2 is a feed weight ratio prediction subsystem of the fish feed formulation of this patent;
FIG. 3 is a flow chart of a fish feed formulation firefly algorithm optimization subsystem of the present patent;
FIG. 4 is a detection node of the present patent;
FIG. 5 is a control node of the present patent;
FIG. 6 is a gateway node of the present patent;
fig. 7 shows the site monitoring software of the present patent.
Detailed Description
The technical scheme of the application is further described by combining the attached drawings 1-7:
design of overall system function
The invention relates to a fish feed detection and formula system, which realizes detection of aquaculture environment parameters, prediction of fish feed formula material weight ratio and optimization of feed formula. The aquaculture environment parameter acquisition and control platform comprises a detection node, a control node, a gateway node, an on-site monitoring end, a cloud platform and a remote monitoring end of aquaculture environment parameters, wherein the detection node and the control node construct LoRa network communication to realize the LoRa network communication among 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, bidirectional transmission of the aquaculture environment parameters and related control information is realized among the gateway node, the cloud platform, the on-site monitoring terminal and the remote monitoring terminal, and the aquaculture environment parameter acquisition and control platform is shown in figure 1.
Design of detection node
A large number of detection nodes 1 based on an LoRa communication network are adopted as aquaculture environment parameter sensing terminals, and the detection nodes realize mutual information interaction between on-site monitoring terminals through the LoRa communication network. The detection node comprises sensors for acquiring parameters of temperature, dissolved oxygen, pH value and salinity of the aquaculture environment, a corresponding signal conditioning circuit, an STM32 microprocessor and an SX1278 radio frequency module in LoRa network communication; the software of the detection node mainly realizes LoRa network communication and acquisition and pretreatment of aquaculture environment parameters. The software is designed by adopting a C language program, so that the compatibility degree is high, the working efficiency of software design and development is greatly improved, and the reliability, readability and transportability of program codes are enhanced. The structure of the detection node is shown in fig. 4.
Design of control node
The control node realizes mutual information interaction with the gateway node through an LoRa network, and comprises 4 digital-to-analog conversion circuits corresponding to control external equipment, an STM32 microprocessor, 4 external equipment controllers and an LoRa communication network module SX1278 radio frequency module; the 4 external equipment controllers are respectively a temperature controller, a dissolved oxygen controller, a pH value controller and a salinity controller. The control node is shown in figure 5.
Fourth, gateway node design
The gateway node comprises an SX1278, an NB-IoT module, an STM32 single chip microcomputer and an RS232 interface, the gateway node comprises an SX1278 radio frequency module to realize an LoRa communication network communicated with the detection node and the control node, the NB-IoT module realizes data bidirectional interaction between the gateway and the cloud platform, and the RS232 interface is connected with the field monitoring terminal to realize information interaction between the gateway and the field monitoring terminal. The gateway node is shown in figure 6.
Design of five cloud platforms
The cloud platform supports multiple transmission protocols to provide high-quality services such as simple and convenient mass connection, cloud storage, message distribution and big data analysis for various cross-platform Internet of things applications and industry solutions, and has good visual application. Firstly, an aquaculture environment monitoring product is created on a cloud platform, a detection node, a control node, a gateway node, a field monitoring end and a remote monitoring computer are connected into the created product according to a transmission protocol of the platform, the TCP connection with an internet and a cloud platform server is established, data transmission and other operations are completed, and the bidirectional transmission of data and information between the detection node, the control node, the gateway node, the field monitoring end and the remote monitoring computer is realized.
Design of six, remote monitoring computer
The remote monitoring computer performs Web visual aquaculture environment parameter interface management, remote environment control is performed on instructions under control nodes, data are stored and information is issued at a cloud platform end, aquaculture personnel access and view aquatic real-time environment information, inquire and derive historical data and perform remote control on aquaculture equipment by using a browser of the remote monitoring computer based on a B/S framework, and a Web page of the remote monitoring computer has an automatic alarm function so that management personnel can take measures in time.
Seven, field monitoring terminal software design
The on-site monitoring terminal is an industrial control computer, mainly collects and processes aquaculture environment parameters and realizes information interaction with gateway nodes, and the on-site monitoring terminal mainly has the functions of communication parameter setting, data analysis and data management, a fish feed formula material weight ratio prediction subsystem and a fish feed formula firefly algorithm optimization subsystem. The structures of the fish feed formula material weight ratio prediction subsystem and the fish feed formula firefly algorithm optimization subsystem are shown in figures 2 and 3. The management software selects Microsoft Visual + +6.0 as a development tool, calls the Mscomm communication control of the system to design a communication program, and the functions of the field monitoring end software are shown in figure 7. The design process of the fish feed formula material weight ratio prediction subsystem and the fish feed formula firefly algorithm optimization subsystem is as follows:
(I) fish feed formula feed weight ratio prediction subsystem
The method comprises a fuzzy C-means clustering algorithm, a plurality of CNN convolutional neural network models, a plurality of 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 material-to-weight ratio trend prediction module and an environment evaluation module. The design of each model is as follows:
1. fuzzy C-means clustering algorithm design
The fish feed formula is used as the input of a fuzzy C-means clustering algorithm, the fish feed formulas of a plurality of categories output by the fuzzy C-means clustering algorithm are respectively used as the input of a plurality of corresponding CNN convolutional neural network models, and a finite set X ═ X is set1,x2,…xnIs a set of n samples of fish feed formulas, each of which is a fish feed formula, C is a predetermined category, mi(i ═ 1,2, … c) is the center of each cluster, μj(xi) Is the membership of the ith sample with respect to the jth class, and the clustering criterion function is defined by the membership function as:
Figure BDA0003054449260000101
in the formula, | | xi-mjIs xiTo mjThe euclidean distance between; b is fuzzy weighted power exponent, which is a parameter capable of controlling the fuzzy degree of the clustering result; m is a fuzzy C partition matrix of X, V is a clustering center set of X, and the result of the fuzzy C-means clustering algorithm is to obtain M and V which enable the criterion function to be minimum. In the fuzzy C-means clustering method, the sum of the membership degrees of the samples to each cluster is required to be 1, namely:
Figure BDA0003054449260000102
the FCM algorithm can be done in the following iterative steps:
A. setting the clustering number c and the parameter b, stopping the threshold value epsilon of the algorithm, setting the iteration number t to be 1, and allowing the maximum iteration number to be tmax(ii) a B. Initializing each cluster center mi(ii) a C. Computing affiliation with current cluster centersA function of membership; D. updating various clustering centers by using the current membership function; E. selecting a proper matrix norm, if | | | V (t +1) -V (t) | | | is less than or equal to epsilon or t is more than or equal to tmaxStopping the operation; otherwise, returning to the step C when t is t + 1. And when the algorithm is converged, obtaining various clustering centers and the membership degree of each sample to various classes, and finishing fuzzy clustering division. And finally, defuzzifying the fuzzy clustering result, converting the fuzzy clustering into deterministic classification, and realizing final clustering segmentation.
2. CNN convolutional neural network model design
And the fish feed formulas of a plurality of categories output by the fuzzy C-means clustering algorithm are respectively used as the input of a plurality of corresponding CNN convolutional neural network models, and the outputs of the CNN convolutional neural network models are respectively used as the input of a plurality of corresponding LSTM neural network models. The CNN convolutional neural network model can automatically mine and extract sensitive spatial features representing the system state from a large amount of fish feed formulas and fish feed weight ratio historical data, and mainly comprises 4 parts: input layer (Input). The input layer is the input of the CNN convolutional neural network model, and the fish feed formula and the fish feed weight ratio historical data original data or the preprocessed signals are generally normalized and then directly input. ② a convolutional layer (Conv). Because the data dimension of the input layer is large, the CNN convolutional neural network model is difficult to directly and comprehensively sense all fish feed formula and fish feed material weight ratio historical data input information, the input data needs to be divided into a plurality of parts for local sensing, then the global information is obtained through weight sharing, and meanwhile the complexity of the CNN convolutional neural network model structure is reduced. And a pooling layer (Pool, also known as a down-sampling layer). Because the dimensionality of the data samples obtained after the convolution operation is still large, the data size needs to be compressed and key information needs to be extracted to avoid overlong model training time and overfitting, and therefore a pooling layer is connected behind the convolution layer to reduce the dimensionality. And taking the peak characteristic of the defect characteristic into consideration, performing down-sampling by adopting a maximum pooling method. And fourthly, a full connection layer. After all convolution operations and pooling operations, feature extraction data enter a full-connection layer, each nerve layer in the layer is in full connection with all neurons in the previous layer, and local feature information extracted by the convolution layer and the pooling layer is integrated. Meanwhile, in order to avoid the over-fitting phenomenon, a lost data (dropout) technology is added in the layer, the output value passing through the last layer of full connection layer is transmitted to the output layer, and the pooling results of the last layer are connected together in an end-to-end mode to form the output layer.
3. LSTM neural network model design
The outputs of the CNN convolutional neural network models are respectively used as the inputs of the corresponding LSTM neural network models, and the outputs of the LSTM neural network models are used as the inputs of the GRNN neural network model. The temporal Recurrent Neural Network (RNN) of the LSTM neural network model, which is composed of long-short term memory (LSTM) units, is referred to as the LSTM neural network model temporal recurrent neural network, and is also commonly referred to as the LSTM neural network model network. The LSTM neural network model introduces mechanisms of Memory cells (Memory cells) and hidden layer states (Cell states) to control the transfer of information between hidden layers. The memory unit of an LSTM neural network model neural network is internally provided with 3 Gate (Gates) computing structures which are an Input Gate (Input Gate), a forgetting Gate (Forget Gate) and an Output Gate (Output Gate). Wherein, the input gate can control the adding or filtering of new information; the forgetting door can forget the information to be lost and keep the useful information in the past; the output gate enables the memory unit to output only information related to the current time step. The 3 gate structures carry out operations such as matrix multiplication, nonlinear summation and the like in the memory unit, so that the memory still cannot be attenuated in continuous iteration. The long-short term memory unit (LSTM) structure unit is composed of a unit (Cell), an Input Gate (Input Gate), an Output Gate (Output Gate) and a forgetting Gate (Forget Gate). The LSTM neural network model is suitable for predicting the dynamic change of the material-to-weight ratio of the fish feed formula in time series, and can effectively prevent the gradient disappearance and the long-term storage during RNN trainingShort Term Memory (LSTM) networks are a special type of RNN. The LSTM neural network model can learn long-term dependency information while avoiding the gradient vanishing problem. The LSTM neural network model adds a structure called a Memory Cell (Memory Cell) in a neural node of a hidden layer of a neuron internal structure RNN for memorizing the material weight ratio dynamic change information of a past fish feed formula, and adds three gate structures (Input, Forget and Output) for controlling the use of the material weight ratio historical information of the fish feed formula. Setting the time sequence value of the feed weight ratio of the input fish feed formula as (x)1,x2,…,xT) The hidden layer state is (h)1,h2,…,hT) Then, time t has:
it＝sigmoid(Whiht-1+WxiXt) (3)
ft＝sigmoid(Whfht-1+WhfXt) (4)
ct＝ft⊙ct-1+it⊙tanh(Whcht-1+WxcXt) (5)
ot＝sigmoid(Whoht-1+WhxXt+Wcoct) (6)
ht＝ot⊙tanh(ct) (7)
wherein it、ft、otRepresenting input, forget and output doors, ctRepresenting a cell, WhRepresenting the weight of the recursive connection, WxSigmoid and tanh represent the weights from the input layer to the hidden layer, and are two activation functions. The method comprises the steps of firstly establishing an LSTM neural network model, establishing a training set by utilizing the data of the material weight ratio of the pretreated fish feed formula and training the model, wherein the LSTM neural network model considers the time sequence and nonlinearity of the change of the material weight ratio of the fish feed formula to the material weight ratio of the fish feed formula and has higher dynamic material weight ratio of the fish feed formulaAnd (4) predicting the precision.
4. GRNN neural network model design
The output of a plurality of LSTM Neural network models is used as the input of a GRNN Neural network model, the output of the GRNN Neural network model is used as the input of a fuzzy recurrent Neural network model, the GRNN (generalized Regression Neural network) Neural network model is a local approximation network, the GRNN Neural network model is established on the basis of mathematical statistics and has definite theoretical basis, the network structure and the connection value are determined accordingly after the learning sample is determined, and only one variable of a smoothing parameter needs to be determined in the training process. The learning of the GRNN neural network completely depends on data samples, has stronger advantages than the BRF network in approximation capability and learning speed, has strong nonlinear mapping and flexible network structure and high fault tolerance and robustness, and is particularly suitable for fast approximation of functions and processing unstable data. The artificial adjustment parameters of the GRNN neural network model are few, and the learning of the network completely depends on data samples, so that the influence of artificial subjective assumption on the prediction result can be reduced to the maximum extent by the network. The GRNN neural network has strong prediction capability under a small sample, has the characteristics of high training speed, strong robustness and the like, and is basically not disturbed by multiple collinearity of input data. The GRNN neural network model structure constructed by the method comprises an input layer, a mode layer, a summation layer and an output layer, wherein an input vector X of the GRNN neural network model is an n-dimensional vector, and a network output vector Y is a k-dimensional vector X ═ X1,x2,…,xn}TAnd Y ═ Y1,y2,…,yk}T. The number of neurons in the mode layer is equal to the number m of training samples, each neuron corresponds to a training sample one by one, and the transfer function p of the neurons in the mode layeriComprises the following steps:
pi＝exp{-[(x-xi)T(x-xi)]/2σ},(i＝1,2,…,m) (8)
the neuron outputs in the above formula enter a summation layer for summation, and the summation layer functions are divided into two types, which are respectively:
Figure BDA0003054449260000141
Figure BDA0003054449260000142
wherein, yijThe jth element value in the vector is output for the ith training sample. According to the GRNN algorithm, the estimated value of the jth element of the network output vector Y is:
yj＝sNj/sD,(j＝1,2,…k) (11)
the GRNN neural network model is established on the basis of mathematical statistics, the implicit mapping relation of the GRNN neural network model can be approximated according to sample data, the output result of the network can be converged on an optimal regression surface, and a satisfactory prediction effect can be obtained particularly under the condition that the sample data is rare. The GRNN neural network model has strong classification capability and high learning speed, is mainly used for solving the problem of function approximation and has high parallelism in the aspect of structure.
5. Fuzzy recurrent neural network model design
And the outputs of the GRNN neural network model, the time delay neural network model, the material weight ratio trend prediction module and the environment 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, and the network consists of 4 layers: an input layer, a membership function layer, a rule layer, and an output layer. The network comprises n input nodes, wherein each input node corresponds to m condition nodes, m represents a rule number, nm rule nodes and 1 output node. Layer I in the figure introduces the input into the network; the second layer fuzzifies the input, and the adopted membership function is a Gaussian function; the third layer corresponds to fuzzy reasoning; layer iv corresponds to the defuzzification operation. By using
Figure BDA0003054449260000143
Representing the input and output of the ith node of the kth layer, respectively, the signal transfer process inside the network and the input-output relationship between the layers can be described as follows. Layer I: an input layer, each of the layersThe input nodes are directly connected to the input variables, and the input and output of the network are represented as:
Figure BDA0003054449260000151
in the formula
Figure BDA0003054449260000152
And
Figure BDA0003054449260000153
for the input and output of the ith node of the network input layer, N represents the number of iterations.
Layer II: and in the membership function layer, nodes of the membership function layer fuzzify input variables, each node represents a membership function, and a Gaussian function is adopted as the membership function. The inputs and outputs of the network are represented as:
Figure BDA0003054449260000154
in the formula mij and σijRespectively representing the mean center and width value of the j term Gaussian function of the ith linguistic variable of the II layer, wherein m is the number of all linguistic variables corresponding to the input node.
Layer III: the fuzzy inference layer, namely the rule layer, adds dynamic feedback to ensure that the network has better learning efficiency, and the feedback link introduces an internal variable hkAnd selecting a sigmoid function as an activation function of the internal variable of the feedback link. The inputs and outputs of the network are represented as:
Figure BDA0003054449260000155
in the formula of omegajkIs the connecting weight value of the recursion part, the neuron of the layer represents the front-piece part of the fuzzy logic rule, the node of the layer performs pi operation on the output quantity of the second layer and the feedback quantity of the third layer,
Figure BDA0003054449260000156
is the output of the third layer, and m represents the number of rules in a full connection. 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 rule node matching degree of the layer 3. The internal variables introduced by the feedback link comprise two types of nodes: and the receiving node and the feedback node. The carrying node calculates an internal variable by using weighted summation to realize the defuzzification function; the result of fuzzy inference of hidden rules represented by internal variables. And the feedback node adopts a sigmoid function as a fuzzy membership function to realize the fuzzification of the internal variable. The membership function layer of the HRFNN network uses a local membership function, which is different from the local membership function: the feedback part adopts a global membership function on the domain of an internal variable to simplify the network structure and realize the feedback of global historical information. The number of the receiving nodes is equal to the number of the feedback nodes; the number of the bearing nodes is equal to the number of the nodes of the rule layer. The feedback quantity is connected to the 3 rd layer and serves as the input quantity of the fuzzy rule layer, and the output of the feedback node contains historical information of the activation strength of the fuzzy rule.
A fourth layer: the deblurring layer, i.e., the output layer. The layer node performs a summation operation on the input quantities. The inputs and outputs of the network are represented as:
Figure BDA0003054449260000161
in the formula lambdajIs the connection weight of the output layer. The fuzzy recurrent neural network has the performance approaching to a highly nonlinear dynamic system, the training error and the testing error of the fuzzy recurrent neural network added with the internal variable are respectively obviously reduced, the network prediction effect is superior to that of the fuzzy neural network with the self-feedback fuzzy recurrent neural network and the dynamic modeling, which shows that the learning capacity of the network is enhanced after the internal variable is added, and the dynamic characteristic of the sewage treatment system is more fully reflected. The simulation result proves the effectiveness of the network. The fuzzy recurrent neural network HRFNN of the patent is adoptedAnd training the weight of the neural network by using a gradient descent algorithm added with cross validation. And (5) using HRFNN to predict the material weight ratio parameter. The HRFNN introduces an internal variable in a feedback link, performs weighted summation on the output quantity of the rule layer, then performs defuzzification output as a feedback quantity, and uses the feedback quantity and the output quantity of the membership function layer as the input of the rule layer at the next moment. The network output comprises the activation intensity of the rule layer and the output historical information, and the capability of the HRFNN to adapt to a nonlinear dynamic system is enhanced. Experiments show that the HRFNN can accurately predict the material-weight ratio parameters. The simulation result is compared with results obtained by other networks, when the model established by the method is applied to material-to-weight ratio prediction, the network scale is minimum, the prediction error is small, and the effectiveness of the method is indicated.
6. NARX neural network model design
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, which can be equivalent to a BP neural network with input time delay and added with time delay feedback connection from output to input on a topological connection relation, and the structure of the NARX neural network model is composed of an input layer, a time delay layer, a hidden layer and an output layer, wherein an input layer node is used for signal input, a time delay layer node is used for time delay of an input signal and an output feedback signal, the hidden layer node performs nonlinear operation on the delayed signal by using an activation function, and an output layer node is used for performing linear weighting on hidden layer output to obtain final network output. The NARX neural network has the characteristics of nonlinear mapping capability, good robustness, adaptability and the like, and is suitable for predicting the feed-to-feed ratio of the fish. x (t) represents the external input of the neural network, namely 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, namely the predicted value of the material-to-weight ratio in the next time interval; n is the output delay order; s is the number of hidden layer neurons; the output of the jth implicit element can thus be found as:
Figure BDA0003054449260000171
in the above formula, wjiAs a connection weight between the ith input and the jth implicit neuron, bjIs the bias value of the jth implicit neuron, and the output y (t +1) of the network has the value:
y(t+1)＝f[y(t),y(t-1),…,y(t-n),x(t),x(t-1),…,x(t-m+1)；W] (17)
the NARX neural network model of the invention is a dynamic feedforward neural network, the NARX neural network is a nonlinear autoregressive network with the output of a fuzzy recurrent neural network model with external input, has the dynamic characteristic of multi-step time delay and is connected to a plurality of layers of closed networks of network input through feedback material weight ratio output values, the NARX neural network model is a dynamic neural network which is most widely applied in a nonlinear dynamic system, and the performance of the NARX neural network model is generally superior to that of a full recurrent neural network. A typical NARX recurrent neural network is mainly composed of an input layer, a hidden layer, an output layer, and input and output delays, before application, the delay order and the number of hidden layer neurons of the input and output are generally determined in advance, and the current output material-to-weight ratio of the NARX neural network model depends not only on the output material-to-weight ratio at the past y (t-n) time, but also on the current output of the fuzzy recurrent neural network model as an input vector x (t), the delay order of the input vector, and the like. The output of the fuzzy recurrent neural network model is used as an input signal and is transmitted to the hidden layer through the time delay layer, the hidden layer processes the input signal and then transmits the processed input signal to the output layer, the output layer linearly weights the output signal of the hidden layer to obtain a final output material-to-weight ratio of the NARX neural network model, and the time delay layer delays a signal fed back by the output material-to-weight ratio of the NARX neural network model and a signal output by the fuzzy recurrent neural network model and serves as the output signal of the input layer and then transmits the delayed signal to the hidden layer.
7. Time delay 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 of the time delay neural network model is used as the corresponding input of the fuzzy recurrent neural network model. The Time Delay Neural Network (TDNN) is an adaptive linear network, the input of which enters from the left side of the network, and becomes the input of a D + 1-dimensional vector after D-step Delay under the action of a single-step Delay line D, the vector is formed by combining signals output by a current K-Time NARX Neural network model and signals output by D-1 NARX Neural network models before K, a neuron adopts a linear activation function, and the Delay Neural network belongs to the variation of the traditional artificial Neural network. The time delay neural network structure consists of an input layer, an output layer and one or a plurality of hidden layers, and the neural network establishes a mapping relation between input and output. Different from the traditional neural network, the time delay neural network realizes the memory of preamble input by delaying input at an input layer, and the input is delayed at the input layer, so that the network can jointly predict the output of the current time point by using the input of previous d steps and the current input, and for the time delay neural network with the delay step number of d at an input layer, R is a forward propagation operator of the time delay neural network, the relation between an input sequence X and an output sequence Y can be simply expressed as follows:
Y(t)＝R(X(t),X(t-1),…,X(t-d)) (18)
8. material-to-weight ratio trend prediction module design
The material weight ratio trend prediction module comprises a CNN convolutional neural network model, an LSTM neural network model, an ARIMA model and a NARX neural network model, fish feed material weight ratio historical data are respectively used as the input of the CNN convolutional neural network model and the ARIMA model, the output of the CNN convolutional neural network model and the ARIMA model is used as the input of the NARX neural network model, and the output value of the NARX neural network model is used as the output of the material weight ratio trend prediction module; the material weight ratio trend prediction module is shown in figure 2. The CNN convolutional neural network model, the LSTM neural network model and the NARX neural network model respectively refer to the design process, and the ARIMA model is designed as follows:
the ARIMA model is a time series predictive fuel-to-weight ratio modeling method proposed by Box et al, and extends to the analysis of time series of predicted fuel-to-weight ratios. According to the study on the material-weight ratio time series characteristics of the ARIMA model, 3 parameters are adopted to analyze the time series of the material-weight ratio change, namely the autoregressive order (p), the difference times (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:
Figure BDA0003054449260000191
Δdytdenotes ytSequence after d differential conversions,. epsilontIs a random error of time, is a white noise sequence which is independent of each other, and has a mean value of 0 and a variance of a constant sigma2Normal distribution of phii(i ═ 1,2, …, p) and θj(j ═ 1,2, …, q) are parameters to be estimated for the ARIMA model, and p and q are orders of the ARIMA dynamic prediction fuel-to-weight ratio model. The ARIMA dynamic prediction fuel-to-weight ratio model belongs to a linear model essentially, and the modeling and prediction comprise 4 steps of (1) sequence stabilization treatment. If the material weight ratio data sequence is not stable, if a certain increase or decrease trend exists, the data needs to be differentially processed. Common tools are autocorrelation function maps and partial autocorrelation function maps. If the autocorrelation function rapidly approaches zero, the material-to-weight time series is a stationary time series. If the time sequence has a certain trend, the material weight ratio data needs to be subjected to difference processing, if seasonal rules exist, seasonal difference is also needed, and if the time sequence has heteroscedasticity, logarithmic conversion needs to be carried out on the material weight ratio data. (2) And (5) identifying the model. The orders p, d and q of the ARIMA dynamic prediction fuel-to-weight ratio model are mainly determined through autocorrelation coefficients and partial autocorrelation coefficients. (3) Estimating parameters of the model and diagnosing the model. Obtaining estimated values of all parameters in an ARIMA dynamic prediction material-to-weight ratio model by using maximum likelihood estimation, checking the estimated values including parameter significance check and residual randomness check, judging whether the built material-to-weight ratio model is available or not, and performing material-to-weight ratio prediction by using the ARIMA dynamic prediction material-to-weight ratio model with selected proper parameters; and checks are made in the model to determine if the model is adequate and if not, the parameters are re-estimated. (4) By using a suitable parameterAnd predicting the change trend of the material-to-weight ratio by the material-to-weight ratio model.
9. Environment assessment module design
The environment evaluation module comprises a plurality of LSTM neural network models, a plurality of self-association neural network models and an NARX neural network model, the outputs of a plurality of groups of temperature, dissolved oxygen and PH value sensors are respectively used as the inputs of the corresponding plurality of LSTM neural network models, the outputs of the plurality of LSTM neural network models are respectively used as the input of each self-association neural network model, the outputs of the plurality of self-association neural network models are used as the inputs of the NARX neural network models, and the output value of the NARX neural network model is used as the output of the environment evaluation module; the environmental assessment module is shown in figure 2. The self-association neural network model design process is as follows: an Auto-associative neural network (AANN) model is a feedforward neural network with a special structure, and the model structure of the Auto-associative neural network includes an input layer, a number of hidden layers and an output layer. The method comprises the steps of firstly compressing data information output by a plurality of LSTM neural network models through an input layer, a mapping layer and a bottleneck layer, extracting the most representative low-dimensional subspace reflecting the structure of an aquaculture environment evaluation grade system from a high-dimensional parameter space output by the plurality of LSTM neural network models, effectively filtering noise and measurement errors in aquaculture environment evaluation grade input data, decompressing the aquaculture environment evaluation grade data through the bottleneck layer, the demapping layer and the output layer, and restoring the previously compressed information to each parameter value, so that reconstruction of the aquaculture environment evaluation grade input data is realized. In order to achieve the purpose of compressing the information of the evaluation grade of the aquaculture environment, the number of nodes of a network bottleneck layer of a self-associative neural network model is obviously smaller than that of an input layer, and in order to prevent the formation of simple single mapping between the input and output layers of the evaluation grade of the aquaculture environment, except that the excitation function of the output layer adopts a linear function, the excitation functions of other layers all adopt non-linear excitation functions. In essence, the first layer of the hidden layer of the self-associative neural network model is called as a mapping layer, and the node transfer function of the mapping layer can be an S-shaped function or other similar nonlinear functions; the second layer of the hidden layer is called a bottleneck layer, the dimension of the bottleneck layer is the minimum in the network, the transfer function of the bottleneck layer can be linear or nonlinear, the bottleneck layer avoids the mapping relation that the output and the input are equal and can be easily realized in a one-to-one way, the bottleneck layer enables a self-association neural network model to encode and compress the aquaculture environment evaluation grade signal to obtain a relevant model of the input aquaculture environment evaluation grade, and the aquaculture environment evaluation grade is decoded and decompressed behind the bottleneck layer to generate an estimated value of the aquaculture environment evaluation grade input signal; the third layer or the last layer of the hidden layer is called a demapping layer, the node transfer function of the demapping layer is a generally nonlinear S-shaped function, and the self-associative neural network is trained by an error back propagation algorithm.
(II) fish feed formula firefly algorithm optimization subsystem design
The fish feed formula firefly algorithm optimization subsystem mainly comprises 5 links including firefly population initialization, fluorescence brightness determination, objective function calculation, firefly position updating and optimal fish feed formula determination, the firefly population is randomly distributed in a search space as an initial solution, each firefly is regarded as a fish feed formula and can be attracted by the firefly brighter than the firefly, the attraction of the firefly is positively correlated with the brightness, for any two fireflies, one firefly can move towards the other firefly brighter than the firefly, the brightness is reduced along with the increase of the distance, and the fireflies are gathered around the firefly with high brightness; the fish feed formulation firefly algorithm optimization process is shown in figure 3, and the optimization process of the fish feed formulation firefly algorithm optimization subsystem is as follows:
1. firefly population initialization
The fish feed formula consists of rice bran, bean cakes, fish meal and yeast powder, the content of each material in each part of the fish feed formula is within a certain range, as a constraint condition of the fish feed formula, the content of each material can meet the nutritional requirement of fish growth within the range, and the limit range of the content (unit of kilogram) of each material in each part of the fish feed formula is as follows: the rice bran is [35,45 ]]Bran is [35,45 ]]The bean cake is [9,13 ]]The fish meal is [9,13 ]]The yeast powder is [1.5,3 ]]. According to the constraints of the fish feed formulation, 5 real numbers representing 5 materials in the fish feed formulation are randomly generated, and are arranged together to form a firefly individual of the fish feed formulation, which is defined as a firefly 5-dimensional search space, and are B, C, D, E and F respectively. Continuously generating M firefly individuals of the fish feed formula, wherein M is the scale of the firefly population, namely the number of the firefly individuals of each generation of fish feed formula, and the number of the firefly is M, and the spatial position of the ith firefly is Ai＝[Bi,Ci,Di,Ei,Fi](ii) a The light intensity absorption coefficient is gamma, and the maximum attraction degree is beta0Step size factor is alpha; the maximum and minimum weight is wmax，wminThe maximum update algebra is Smax. Randomly distributing the fireflies in 5-dimensional space, and setting the distance d between the ith and the jth firefliesijj is the Euclidean distance, i.e. the following formula:
Figure BDA0003054449260000211
2. determining fluorescence brightness
The fluorescence brightness function is used for evaluating the advantages and disadvantages of fireflies of the fish feed formula and is used as a basis for the advantages and disadvantages of the fish feed formula optimization process, the fish feed formula optimization is the optimized combination of the material content of the fish feed formula, and the best benefit of striving for the fish feed formula is achieved on the premise of reaching the nutritional standard of the fish feed; the firefly individual of each fish feed formula is used as the input of the fish feed formula material weight ratio prediction subsystem, the output of the fish feed formula material weight ratio prediction subsystem is used as the predicted value of the firefly individual material weight ratio of the feed formula, the reciprocal of the predicted value of the firefly individual material weight ratio of each fish feed formula is used as the fluorescence brightness of the firefly individual of the fish feed formula, the larger the reciprocal of the predicted value of the firefly individual material weight ratio of each fish feed formula is, the higher the fluorescence brightness of the firefly individual of the fish feed formula is, and the sum of the reciprocals of the predicted values of the material weight ratios of the firefly individuals of each fish feed formula in the population is the total fluorescence brightness of the population.
3. Calculation of an objective function
If the specified maximum algebra is calculated, stopping the firefly algorithm after the firefly algorithm is reached; or the difference value between the individual fluorescence brightness of the firefly of the fish feed formula and the individual fluorescence brightness of the firefly of the designated fish feed formula is smaller than a set threshold value and is used as a target function of a firefly algorithm, and the firefly of the fish feed formula is an optimal solution.
4. Firefly location update
The firefly moves toward the direction of the firefly with higher fluorescence brightness than itself, and the degree of attraction between them determines the movement step length of the firefly. The attraction between fireflies i, j is:
Figure BDA0003054449260000221
the position updating formula of the fish feed formula firefly can be further obtained by the formula (21) as follows:
Figure BDA0003054449260000222
in the formula, R is random number which is uniformly distributed on [0, 1 ]; s is the current update generation of the firefly in the fish feed formula. Checking whether the corresponding firefly individuals in the new fish feed formula meet the constraint conditions of the fish feed formula one by one, and if so, taking the firefly individuals in the new fish feed formula as members of a new generation; otherwise, the new fish feed formulation firefly individuals are discarded.
5. Optimal fish feed formulation determination
After the firefly algorithm stopping condition is met, the fluorescence brightness of the firefly individual of each fish feed formula is calculated, the firefly individual of the feed formula with the highest fluorescence brightness is the firefly individual of the optimal fish feed formula, and the optimal fish feed formula is obtained.
Design example of eight-fish feed detection and formula system
According to the actual condition of the aquaculture environment big data detection system, the system is provided with an aquaculture parameter acquisition platform and a plane arrangement installation diagram of a detection node, a control node, a gateway node and a field monitoring end, wherein sensors of the detection node are arranged in all directions of an aquaculture pond in a balanced manner according to the detection requirement, and the aquaculture environment parameters are acquired through the system.
The technical means disclosed in the invention scheme are not limited to the technical means disclosed in the above embodiments, but also include the technical scheme formed by any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and such improvements and modifications are also considered to be within the scope of the present invention.

Claims (6)

1. A fish feed detection and formula system is characterized in that: the system comprises an aquaculture environment parameter acquisition and control platform, a fish feed formula material weight ratio prediction subsystem and a fish feed formula firefly algorithm optimization subsystem, and is used for detecting fish aquaculture environment parameters, predicting the fish feed formula material weight ratio and optimizing a fish feed formula;
the fish feed formula material weight ratio prediction subsystem comprises 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, an NARX neural network model, a time-delay neural network model, a material weight ratio trend prediction module and an environment evaluation module, wherein the fish feed formula is used as the input of the fuzzy C-means clustering algorithm, a plurality of classes of fish feed formulas output by the fuzzy C-means clustering algorithm are respectively used as the input of a plurality of corresponding CNN convolutional neural network models, the output of the CNN convolutional neural network models is respectively used as the input of a plurality of corresponding LSTM neural network models, the output of the LSTM neural network models is used as the input of the GRNN neural network model, and the output of the GRNN neural network model, the time-delay neural network model, the material weight ratio trend prediction module and the environment evaluation module is used as the input of the recurrent fuzzy 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, and the output value of the NARX neural network model is used as the material weight ratio of the fish feed formula.
2. The fish feed detection and formulation system of claim 1, wherein: the fish feed formula firefly algorithm optimization subsystem comprises 5 links of firefly population initialization, fluorescence brightness determination, objective function calculation, firefly position updating and optimal fish feed formula determination, the firefly populations are randomly distributed in a search space as an initial solution, each firefly is regarded as a fish feed formula and attracted by the firefly brighter than the firefly population, the attraction force of the firefly is positively correlated with the brightness, for any two fireflies, one firefly moves towards the other firefly brighter than the firefly population, the brightness decreases along with the increase of the distance, and the fireflies gather to the periphery of the firefly with high brightness;
the firefly individual of each fish feed formula is used as the input of the fish feed formula material weight ratio prediction subsystem, the output of the fish feed formula material weight ratio prediction subsystem is used as the predicted value of the firefly individual material weight ratio of the fish feed formula, the reciprocal of the predicted value of the firefly individual material weight ratio of each fish feed formula is used as the fluorescence brightness of the firefly individual of the fish feed formula, the larger the reciprocal of the predicted value of the firefly individual material weight ratio of each fish feed formula is, the higher the fluorescence brightness of the firefly individual of the fish feed formula is, and the sum of the reciprocals of the predicted values of the material weight ratios of the firefly individuals of each fish feed formula in the firefly population is the total fluorescence brightness of the firefly population.
3. The fish feed detection and formulation system of claim 1, wherein: the material weight ratio trend prediction module comprises a CNN convolutional neural network model, an LSTM neural network model, an ARIMA model and a NARX neural network model, fish feed material weight ratio historical data are respectively used as the input of the CNN convolutional neural network model and the ARIMA model, the output of the CNN convolutional neural network model and the ARIMA model is used as the input of the NARX neural network model, and the output value of the NARX neural network model is used as the output of the material weight ratio trend prediction module.
4. The fish feed detection and formulation system of claim 1, wherein: the environment evaluation module comprises a plurality of LSTM neural network models, a plurality of self-association neural network models and an NARX neural network model, the outputs of a plurality of groups of temperature, dissolved oxygen and PH value sensors are respectively used as the inputs of the corresponding plurality of LSTM neural network models, the outputs of the plurality of LSTM neural network models are respectively used as the input of each self-association neural network model, the outputs of the plurality of self-association neural network models are used as the inputs of the NARX neural network models, and the output value of the NARX neural network model is used as the output of the environment evaluation module.
5. The fish feed detection and formulation system of claim 1, wherein: the aquaculture environment parameter acquisition and control platform consists of a detection node, a control node, a gateway node, a field monitoring terminal, a cloud platform and a remote monitoring computer.
6. The fish feed detection and formulation system of claim 5, wherein: the detection nodes acquire fish culture environment parameters and upload the fish culture environment parameters to the cloud platform through the gateway nodes, the cloud platform provides the fish culture environment parameters to the remote monitoring computer for Web visual fish culture environment parameter interface management, the remote monitoring computer implements environment remote control on instructions under the control nodes, and data and release information are stored at the cloud platform end; the detection nodes and the control nodes are responsible for collecting fish culture environment parameters and controlling fish culture environment equipment, and bidirectional communication among the detection nodes, the control nodes, the field monitoring terminal, the cloud platform and the remote monitoring computer is realized through the gateway nodes, so that the fish culture environment parameter collection and the fish culture equipment control are realized.
CN202110496297.5A 2021-05-07 2021-05-07 Fish feed detection and formula system Pending CN113255739A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110496297.5A CN113255739A (en) 2021-05-07 2021-05-07 Fish feed detection and formula system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110496297.5A CN113255739A (en) 2021-05-07 2021-05-07 Fish feed detection and formula system

Publications (1)

Publication Number Publication Date
CN113255739A true CN113255739A (en) 2021-08-13

Family

ID=77223918

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110496297.5A Pending CN113255739A (en) 2021-05-07 2021-05-07 Fish feed detection and formula system

Country Status (1)

Country Link
CN (1) CN113255739A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114323136A (en) * 2021-12-28 2022-04-12 旭昇智能科技(常熟)有限公司 Production environment monitoring method and system based on Internet of things
CN115016275A (en) * 2022-06-17 2022-09-06 淮阴工学院 Intelligent feeding and livestock and poultry house big data internet of things system

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114323136A (en) * 2021-12-28 2022-04-12 旭昇智能科技(常熟)有限公司 Production environment monitoring method and system based on Internet of things
CN115016275A (en) * 2022-06-17 2022-09-06 淮阴工学院 Intelligent feeding and livestock and poultry house big data internet of things system
CN115016275B (en) * 2022-06-17 2023-06-06 淮阴工学院 Intelligent feeding and livestock house big data Internet of things system

Similar Documents

Publication Publication Date Title
CN111639748B (en) Watershed pollutant flux prediction method based on LSTM-BP space-time combination model
CN110647980A (en) Time sequence prediction method based on GRU neural network
CN110647979B (en) Greenhouse environment multi-parameter intelligent monitoring system based on Internet of things
CN113281465A (en) Livestock and poultry house breeding environment harmful gas detection system
CN113126676B (en) Livestock and poultry house breeding environment parameter intelligent control system
CN112904756B (en) Pipe network big data detection system
CN115016276B (en) Intelligent water content adjustment and environment parameter Internet of things big data system
CN113301127B (en) Livestock feed detection system
CN113255739A (en) Fish feed detection and formula system
CN113219871B (en) Curing room environmental parameter detecting system
CN110633846A (en) Gas load prediction method and device
CN115687995A (en) Big data environmental pollution monitoring method and system
CN115128978A (en) Internet of things environment big data detection and intelligent monitoring system
CN114911185A (en) Security big data Internet of things intelligent system based on cloud platform and mobile terminal App
CN111292121A (en) Garden load prediction method and system based on garden image
CN113283642A (en) Poultry feed detection and formula system
CN112911533B (en) Temperature detection system based on remove end App
CN112903929A (en) Food quality detection system
CN115062764B (en) Intelligent illuminance adjustment and environmental parameter Internet of things big data system
CN114970745B (en) Intelligent security and environment big data system of Internet of things
CN115759343A (en) E-LSTM-based user electric quantity prediction method and device
CN114995248A (en) Intelligent maintenance and environmental parameter big data internet of things system
CN114372418A (en) Wind power space-time situation description model establishing method
CN112906856A (en) Livestock and poultry house ambient temperature detection system based on cloud platform
CN115016275B (en) Intelligent feeding and livestock house big data Internet of things system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination