Aquaculture water quality prediction method based on IMS-NARX

Abstract

The invention discloses an aquaculture water quality prediction method based on an IMSSA-NARX (International Mass Spectroscopy Amplification-Networking ARX Collecting aquaculture water quality parameters, establishing a data set, and completing preprocessing; multi-bit control parameters are introduced into a chaotic equation of an original Tent, so that chaotic mapping of the improved Tent is realized; through combination of a dogvessel alveolar group algorithm and improved Tent chaotic mapping, an improved dogvessel alveolar algorithm IMSSA is realized so as to improve global exploration and local development capability of the dogvessel alveolar algorithm. Training a nonlinear autoregressive neural network NARX model by using an IMSSA algorithm to optimize the weight and bias of the NARX neural network, so as to establish an aquaculture water quality prediction model based on the IMSSA-NARX; and completing an aquaculture water quality prediction experiment based on IMS-NARX, and comparing the method with a conventional method. Compared with the prior art, the model provided by the invention has the highest prediction accuracy, and the IMSSA-NARX network model has good performance in the aspect of aquaculture water quality condition prediction.

Description

Technical Field

[0001] The present invention relates to the field of aquaculture water quality prediction, and in particular to an aquaculture water quality prediction method based on IMSSA-NARX. Background Art

[0002] In order to meet people's dietary needs, my country has been vigorously developing aquaculture. However, in the process of aquaculture, the parameters of aquaculture water have a great impact on the survival of fish and other animals. Among them, dissolved oxygen DO, pH value, and temperature are important indicators of sustainable aquaculture. These parameters will greatly affect the output of aquatic products. If the water quality parameter values ​​are within a reasonable range, aquatic plants and animals can grow and develop normally. For example, the amount of dissolved oxygen required by aquaculture farms usually depends on the type of farmed fish, pH value and water temperature. When the concentration is lower than 3 mg / L, the mortality and morbidity rates of most species will increase. If the status of dissolved water quality parameters can be accurately predicted, aquaculture farmers can take emergency corrective measures to avoid abnormal conditions in aquatic products by increasing the dissolved oxygen concentration and changing the pH value and temperature. In addition, it can also help maintain the water quality management of the ecosystem.

[0003] However, current domestic aquaculture environmental monitoring technologies are relatively backward, with limited use of real-time, continuous water quality monitoring systems. Traditional water quality parameter monitoring typically requires transporting collected water samples to a laboratory for analysis, requiring manual maintenance. Consequently, water environment prediction plays a crucial role in modern aquaculture management. Proper water quality control can maintain the concentrations of water environment parameters within optimal ranges, improve the growth rate and feed utilization of aquatic organisms, and reduce the incidence of disease in aquacultured organisms. However, the changing trends of aquaculture water quality parameters are complex, characterized by multivariate, nonlinear, and time-lag characteristics, making aquatic organisms sensitive to changes in the aquaculture water environment. Numerous models have been developed to predict water environment parameters, but most are complex and suffer from poor accuracy and generalization, resulting in low efficiency and accuracy in aquaculture parameter prediction. Consequently, these models cannot fully meet the requirements for refined aquaculture management. Summary of the Invention

[0004] Purpose of the invention: In response to the problems in the background technology, the present invention discloses an aquaculture water quality prediction method based on IMSSA-NARX, which is used to analyze the key parameters of aquaculture (temperature, dissolved oxygen, pH value) and predict the aquaculture water quality status, and has good performance in predicting aquaculture water quality status.

[0005] Technical solution: The present invention discloses a method for predicting aquaculture water quality based on IMSSA-NARX, comprising the following steps:

[0006] Step 1: Collect aquaculture water quality parameters, establish a data set, and complete preprocessing;

[0007] Step 2: By introducing multi-bit control parameters β and λ into Tent's chaotic equation, the Tent chaotic map is improved, and the salp swarm algorithm is combined with the improved Tent chaotic map to obtain the improved salp algorithm IMSSA;

[0008] Step 3: The improved Salp Intense Squid algorithm IMSSA is used to train the NARX neural network model to optimize the weights and bias of the NARX neural network, thereby establishing an aquaculture water quality prediction model based on IMSSA-NARX.

[0009] Step 4: Use the optimized IMSSA-NARX-based aquaculture water quality prediction model to predict aquaculture water quality.

[0010] Furthermore, in step 1, multiple groups of aquaculture water quality parameters including temperature, pH value, and dissolved oxygen (DO) are collected, and after normalization, they are divided into training data and test data in a ratio of 6:4.

[0011] Furthermore, in step 2, multiple control parameters β and λ are introduced into Tent's chaos equation, specifically:

[0012]

[0013] Among them, h n is the nth iteration value of the mapping variable h, λ is a multi-bit control parameter, and β is the initial control parameter. When β is 0.5, the value of the mapping variable h is prevented from falling into the range (0.2, 0.4, 0.6, 0.8), which may cause the system to exhibit a short-period state, or from falling into the range (0, 0.25, 0.5, 0.75), which may cause a fixed point to appear later. The control parameters are selected based on the set {h|h=t / (D+1), t=1, 2, …, D, h≠β}, where D represents the dimensionality of the search space.

[0014] Furthermore, the improved salp algorithm IMSSA is based on the search of improved Tent chaos map global perturbation, specifically as follows:

[0015] Determine the initial value vector of the Salp Algorithm, obtain the chaotic sequence distributed on [0,1] by the following formula to form the population position, and map the elements in the sequence to the corresponding solution domain:

[0016]

[0017] in, are the values ​​of the nth and n+1th salp individuals represented by chaotic variables in dimension d, ub d and lbd are the upper and lower bounds of the dth dimension of the corresponding solution domain, represents the value of the nth salp in dimension d, The value of the salp individual represented by the nth chaotic variable in dimension d generate;

[0018] In the process of using the improved Tent chaotic map to perform global perturbation search for the salp algorithm SSA, first, the improved Tent chaotic map is used to generate N salps, which are mixed with the original N salps and sorted according to their fitness values. Then, the top N excellent salps after mixing are selected to replace the original N salps, thus completing the global perturbation.

[0019] Furthermore, in step 3, the improved salp algorithm IMSSA is used to train the NARX neural network model, as follows:

[0020] Step 3.1: Construct the NARX neural network model. The NARX neural network model is expressed as:

[0021] y(n)=f NARX (x(n),x(n-1),…,x(n-α),…x(nk),y(n-1),…,y(n-γ),…,y(nm))

[0022] As can be seen from the above formula, the output y(n) is not only related to the current input x(n), but also to the past input x(n-α); α = 1, 2, ..., k, and the past output y(n-γ); γ = 1, 2, ..., m, where k is the maximum delay length of the input and m is the maximum delay length of the output feedback.

[0023] In the specific implementation of the NARX neural network model, the output of the first layer of neurons is defined as:

[0024]

[0025] Where, is the output of the βth neuron in the first layer, K represents the number of inputs x received by the βth neuron in the first layer, is the weight of the αth input x received by the βth neuron, is the input, i.e., x(n) and K delayed signals of x(n), i.e., x(n-α); α=1,2,…,K, is the weight of the γth delayed output y received by the βth neuron, represents the γth delayed output received by the βth neuron from the past, i.e. y(n-γ); γ=1,2,…,M, where M represents the number of delayed outputs y, is the bias, f 1 is the hidden layer activation function;

[0026] Similarly, the output of the second layer of neurons is Defined as:

[0027]

[0028] Where, is the output of the βth neuron in the second layer, h represents the number of βth neurons in the second layer that receive outputs from the first layer, is the weight of the αth input received by the βth neuron, is the input, which represents the output signal from the first layer is the bias, f 2 is the output layer activation function;

[0029] Step 3.2: Use the improved Salp Sea Squid algorithm IMSSA to train the NARX neural network model to optimize the weights of the NARX neural network and bias

[0030] Furthermore, the improved salp algorithm IMSSA trains the NARX neural network model, specifically:

[0031] Step 1: Set initial parameters, including the maximum number of iterations l max , population number N, dimension D, initialization population X N×D , randomly initialize the individual position coordinates of the population;

[0032] Step 2: Calculate the individual fitness value of each salp according to the objective function;

[0033] Step 3: Select the initial location of the food source, sort the fitness values ​​of the salps, and set the coordinates of the salp with the best fitness as the target location, which is the location of the food source;

[0034] Step 4: Determine the leader and followers. Set the individuals in the first half of the salp chain as leaders, and the rest as followers.

[0035] Step 5: Update the position of the salp leader:

[0036]

[0037] Among them, F d is the target position value of the dth dimension, ub d , lb dare the upper and lower bounds of the dth dimension, respectively. c1, c2, and c3 are control parameters. c2 and c3 are random numbers between [0, 1] that determine the direction and step size of the leader position update. c1 is the convergence factor, which is used to balance the convergence speed of the algorithm during the iteration process.

[0038] Step 6: Update the position of the salp follower:

[0039]

[0040] Where n≥2, Respectively represent the position coordinates of the nth and n-1th salps in dimension d;

[0041] Step 7: Perform boundary processing on each dimension of the updated individual, calculate the fitness value of each salp, and compare it with the fitness value of the current target position to update the global optimal salp target position, that is, update the location of the food source;

[0042] Step 8: Determine whether the number of iterations is met. If so, output the result; otherwise, go to Step 4 to continue iterative evolution, and finally output the current optimal fitness and target position.

[0043] Beneficial effects:

[0044] 1. The present invention proposes an improved salp algorithm (IMSSA). First, by introducing multi-bit control parameters on the original Tent chaotic equation, an improved Tent chaotic mapping is realized, so that it has better chaotic characteristics, can generate chaotic space within a larger parameter range, and has good security and pseudo-randomness. Further, considering that the improved Tent chaotic mapping has the advantage of better ergodicity, the salp swarm algorithm is combined with the improved Tent chaotic mapping, that is, the improved Tent chaotic mapping is used to perform a global perturbation search of the salp population, so as to obtain population individuals uniformly distributed in the problem solution space, generate a salp population with good diversity, improve the optimization performance of the salp swarm algorithm, and thus realize the improved salp algorithm IMSSA, which uses the mutual linkage between leaders and followers as an adaptive mechanism, effectively balances the exploration and development of the group, and obtains an accurate estimate of the optimal solution during the optimization period, thereby ensuring that its global exploration and local development capabilities are greatly improved.

[0045] 2. The present invention proposes an aquaculture water quality prediction model based on IMSSA-NARX. First, the NARX network is used to complete the preliminary construction of the prediction model. Since NARX takes into account the effects of input, output and their delay on the model, the accuracy and robustness of the constructed model are greatly improved by introducing delay and feedback. Secondly, the NARX neural network model is trained using the improved Salp Insipid Algorithm (IMSSA) to optimize the weights and thresholds of NARX, thereby constructing an aquaculture water quality prediction model based on IMSSA-NARX. Based on the strong global optimization capability of the IMSSA algorithm, IMSSA-NARX can reduce the search range to near the better weights. Experimental results show that for the training set and test set data, the prediction accuracy of the IMSSA-NARX model is 90.02% and 84.66% respectively. Compared with other conventional models, this method also shows better prediction effects. It can be seen that the proposed model can better improve the shortcomings of existing methods and provide more accurate predictions for aquaculture, thereby laying the foundation for improving the management process of aquaculture farms and early detection of water quality parameters, and further promoting the development of aquaculture industry towards green and sustainable direction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is the implementation block diagram of the aquaculture water quality prediction model based on IMSSA-NARX;

[0047] Figure 2 It is a multi-sensor acquisition system structure;

[0048] Figure 3 It is the overall optimization process of the Improved Salp Squid Algorithm (IMSSA);

[0049] Figure 4 It is a search process based on the global perturbation of the improved Tent chaos map;

[0050] Figure 5 It is the basic structure of NARX neural network;

[0051] Figure 6 It is a prediction model structure based on the NARX neural network of IMSSA. DETAILED DESCRIPTION

[0052] In order to better explain the present invention and facilitate understanding, the technical solution of the present invention is described in detail below. The following examples are provided to explain the present invention, but the present invention is not limited to the following examples.

[0053] This paper proposes an aquaculture water quality prediction method based on IMSSA-NARX. First, an improved Tent chaotic map is introduced and combined with a salp swarm algorithm to implement the improved salp algorithm (IMSSA). Second, the improved salp algorithm (IMSSA) is used to train a NARX neural network model to optimize the weights and thresholds of the NARX neural network, thereby establishing an aquaculture water quality prediction model based on IMSSA-NARX.

[0054] Specifically, the present invention applies multi-sensor fusion technology to the detection of aquaculture water environmental parameters. The feature layer adopts a prediction model based on the improved Salp Intense Squid Algorithm (IMSSA) and nonlinear autoregressive neural network (NARX), namely the IMSSA-NARX model, to predict the information layer data. The input parameters are pre-processed temperature, pH value, and dissolved oxygen data, and the output is the condition of the aquaculture water, including poor, bad, average, good, and good. The implementation block diagram of the proposed fusion method is shown in the figure below. Figure 1 The specific steps are as follows:

[0055] Step (1) collects aquaculture water quality parameters, establishes a data set, and completes preprocessing.

[0056] Step 1.1: Build a multi-sensor acquisition system

[0057] Construct a multi-sensor acquisition system to collect aquaculture water quality parameters. The acquisition part includes a sensor group (temperature, pH and dissolved oxygen electrodes) and its conditioning circuit. During the working process, the water quality signals collected by various sensors on site are converted into digital signals after being conditioned by amplifiers, A / D converters and other conditioning circuits. The real water quality data is then sent to the computer for storage and processing. The structure of the multi-sensor acquisition system is as follows: Figure 2 shown.

[0058] The specific process is as follows:

[0059] (1) Complete temperature data collection

[0060] DS18B20 is selected as the temperature acquisition sensor. DS18B20 is a single bus digital temperature sensor. Its output signal does not require A / D conversion. The lowest measurement temperature is -55℃, the highest temperature is +125℃, and the maximum resolution is 0.0625℃. It can achieve multi-point measurement and is low cost. Figure 3 shown.

[0061] (2) Complete dissolved oxygen data collection

[0062] The DO-957 electrode is used to collect dissolved oxygen data. The sensor is durable, has a short electrode response time, a working measurement range of 0 to 20 mg / L, and an operating temperature of 5 to 40°C.

[0063] (3) Complete pH value data collection

[0064] The pH value indicates the acidity or alkalinity of water. When the pH value in water is below 5.5 or above 9, it is not suitable for aquaculture. This patent uses a pH sensor with a composite electrode to detect the pH value of aquaculture water. Because the output impedance of this electrode is very high, the first stage of the amplifier circuit must use a high-input impedance op amp for impedance matching. To eliminate the influence of power frequency signal interference on the electrode probe, a low-pass filter is also required. After a series of conditioning, it is converted into a voltage signal, and finally into a digital signal for processing.

[0065] In the specific implementation, LC (inductor, capacitor) filtering is used in the low-pass filtering link, and the high-input impedance operational amplifier CA3140EZ is used for impedance matching and amplification. Its specific parameters are shown in Table 1. ADC conversion uses ADC14X250, and its performance parameters are shown in Table 2.

[0066] Table 1 CA3140EZ performance parameters

[0067]

[0068] Table 2 ADC14X250 performance parameters

[0069]

[0070] Step 1.2: Create a dataset of aquatic environmental parameters

[0071] A certain aquatic environment monitoring field was selected, and the above-selected sensors were used to build an acquisition system. The system was deployed at different locations within the water quality monitoring area. The aquaculture water environment data collected every day included temperature, pH value, and dissolved oxygen (DO). Finally, a sample data set of 3,000 temperature, pH value, and dissolved oxygen (DO) was established, and it was divided into training data and test data in a ratio of 6:4. The number of training generations was 500.

[0072] Table 3 shows the descriptive statistics of temperature, dissolved oxygen (DO), and pH in the data set, including the mean, standard deviation, minimum, and maximum values ​​of each parameter. This information summarizes the trend and variability of the data set, as well as the range of variation of each observation.

[0073] Table 3 Detailed description of the observation dataset

[0074]

[0075]

[0076] Step 1.3: Normalization

[0077] A good data source is the foundation for ensuring accurate predictions. Because aquaculture environmental parameters are susceptible to weather, physical and chemical influences, and human activity, and because the presence of aquatic plants can limit the monitoring capabilities of sensor equipment, data collected by monitoring systems often contain missing or outliers. To improve data quality and ensure accurate predictions, the collected data must be preprocessed before being used for model training and evaluation to eliminate any outliers or missing values.

[0078] In the specific implementation, after obtaining the data set through sensor monitoring, due to the different dimensions of each variable, all variable environmental parameters need to be normalized and converted into dimensionless expressions using the following formula to ensure that the normalized value range is between [0, 1].

[0079]

[0080] Where: X nor is the normalized data, X is the original data, and X max 、X min are the maximum and minimum values ​​of each environmental parameter data respectively.

[0081] Step 1.4: Define prediction accuracy

[0082] The prediction accuracy is used to evaluate the performance of the proposed model. In addition, it is also used to compare the performance with other conventional models. The definition formula is as follows:

[0083]

[0084] Among them, Accuracy is the accuracy rate; N is the total number of samples; T N is the correct number of samples.

[0085] Step (2), implement the improved salp sea squirt algorithm (IMSSA)

[0086] The Salp Swarm Algorithm (SSA) has few control parameters, high exploratory power, and a simple, easy-to-understand implementation principle. This algorithm achieves search evolution through inter-group interaction. This involves controlling the continuous movement of individual salps and updating their group positions, ultimately converging at the location with the most abundant food sources, thereby achieving the global optimal solution. During the optimization process, leading salps adjust their positions solely based on the direction of the food source, while followers follow and slowly move, ensuring the group continuously explores and exploits the optimal food source. Furthermore, with each iteration, the group retains the best solution to date, as well as search trends and directions.

[0087] However, the salp swarm algorithm also has the problems of being prone to falling into local extreme values ​​and slow convergence in the late stage of evolution. Therefore, based on the advantage of the improved Tent chaotic map with better ergodicity, the present invention combines the salp swarm algorithm with the improved Tent chaotic map and proposes an improved salp algorithm (IMSSA), thereby giving full play to the advantages of the two algorithms, improving the optimization performance of the salp swarm algorithm, and enhancing the global exploration and local development capabilities of the algorithm during the optimization process.

[0088] In the specific implementation, first, the optimization of the salp alveolar algorithm (SSA) is completed. When the SSA algorithm runs to a certain number of generations, the current optimal position of each particle is sorted according to the fitness value to form a new population; then, the improved Tent chaotic map is used to perform a global perturbation search of the SSA, that is, the improved Tent chaotic map is used to generate N salps, which are mixed with the original first N salps and sorted according to the fitness value. Finally, the first N excellent salps after mixing are selected to replace the original N salps, thereby completing the global perturbation.

[0089] The overall optimization process of the improved salp sea squirt algorithm (IMSSA) is as follows Figure 3 As shown, the specific steps include:

[0090] Step 2.1: Complete the optimization based on the traditional SSA algorithm

[0091] The main implementation process of the traditional SSA algorithm is as follows:

[0092] (1) Initialize the salp population

[0093] The target environment is defined as an N×D dimensional space, where N represents the number of populations and D represents the dimension of the space. i Defined as in, is the d-th dimension position of the i-th salp, and the target position of the i-th food source is represented by F i It is specifically defined as: F i =[F1,F2,...,F d ,...F D ], where F d is the d-th dimension data of the ith food source, i.e., the target location. The search range of each dimension includes the upper bound Ub=[ub1,ub2,...,ub d ,...,ub D ] and the lower bound Lb=[lb1,lb2,...,lb d ,...,lb D ], where ub d , lb d are the upper and lower bounds of the d-th dimension respectively.

[0094] Finally, the initial population position is randomly obtained according to formula (3):

[0095] X N×D =rand(N,D)×(Ub-Lb)+Lb (3)

[0096] (2) Leader position update

[0097] The position of each dimension in the population that defines the leader is expressed as The position of each dimension of the follower is expressed as Where n = 2, 3, ... N, d represents the dimension of the individual salps, and N represents the number of populations.

[0098] The leader's position update is related to the target position, but must have a certain degree of randomness, so that it can play a leading role in the entire environment search and lead the followers to move towards the target. The update is performed according to formula (4):

[0099]

[0100] In the above formula, F d is the target position value of the dth dimension, c1, c2, and c3 are control parameters, where c2 and c3 are random numbers between [0, 1], which determine the direction and step size of the leader position update. c1 is the convergence factor, which is used to balance the convergence speed of the algorithm in the iterative process. It is obtained from formula (5):

[0101]

[0102] In the formula, l represents the current number of iterations, l max Indicates the maximum number of iterations.

[0103] (3) Follower location update

[0104] Followers are updated based on the leader, not randomly. In the specific implementation, according to Newton's laws of motion, the follower's velocity (vf, v0) and acceleration (a) are used to update the follower's position. The specific update formula is as follows:

[0105]

[0106] Where t is the iteration time, a is the individual motion acceleration, expressed as a = (v f -v0) / t, v0 is the initial velocity, and v0 is 0 at the beginning of each iteration.

[0107] Since the position update of the follower is only related to the position of the previous salp, the speed v f It is expressed as follows:

[0108]

[0109] Therefore, the position update formula It can be expressed as:

[0110]

[0111] Where n≥2, They represent the position coordinates of the nth and n-1th salps in dimension d, respectively.

[0112] In specific implementation, the specific execution process of the traditional SSA algorithm is as follows:

[0113] Step 1: Set initial parameters, including the maximum number of iterations l max , population number N, dimension D, initialization population X N×D , randomly initialize the individual position coordinates of the population.

[0114] Step 2: Calculate the individual fitness value of each salp based on the objective function.

[0115] Step 3: Select the initial location of the food source. Sort the fitness values ​​of the salps and set the coordinates of the salp with the best fitness as the target location, which is the location of the food source.

[0116] Step 4: Determine the leader and followers. Set the individuals in the first half of the salp chain as leaders, and the rest as followers.

[0117] Step 5: Update the position of the salp leader according to equation (4).

[0118] Step 6: Update the position of the salp follower according to equation (8).

[0119] Step 7: Perform boundary processing on each dimension of the updated individual, calculate the fitness value of each salp, and compare it with the fitness value of the current target position to update the global optimal salp target position, that is, update the location of the food source.

[0120] Step 8: Determine whether the number of iterations is met. If so, output the result; otherwise, go to Step 4 to continue iterative evolution. Finally, output the current optimal fitness and target position.

[0121] Step 2.2: Implementing the Improved Tent Chaos Map

[0122] Chaos is the study of randomness and determinism within nonlinear systems. Chaotic sequences are a type of mapping array generated by the iteration of simple functions, characterized by aperiodicity, ergodicity, and long-term unpredictability. While chaotic motion may superficially resemble a random process, it is fundamentally different. While random processes are caused by chaotic external noise, chaotic motion is governed by the inherent characteristics of objective laws and is an external manifestation of these laws, hence the name "deterministic chaos." Chaotic motion shares the characteristics of deterministic motion: it lacks a period, yet it proceeds in a consistent order and exhibits strong sensitivity to initial values. These properties are lacking in random motion, making chaotic motion predictable in the short term, while random motion is unpredictable.

[0123] Common chaotic mapping methods include Logistic, Tent, and PWLCM. However, most of these methods are sensitive to initial values; different initial values ​​will generate different sequences. Relatively speaking, Tent mapping produces more uniform sequences than other mappings. Therefore, this paper uses Tent mapping to complete the optimization search based on chaotic mapping. Furthermore, to further enhance the effectiveness of chaotic mapping, an improved Tent chaotic mapping is introduced. Furthermore, it is used to perform a global perturbation search on a salp population, thereby obtaining individuals that are uniformly distributed across the solution space, resulting in a salp population with good diversity.

[0124] The improved Tent's chaotic mapping equation is introduced, and multi-bit control parameters β and λ are introduced into the original equation. The specific definitions are as follows:

[0125]

[0126] Among them, h n is the value of the mapping variable h at the nth iteration. λ is a multi-bit control parameter, and β is the initial control parameter. It should be noted that when β is set to 0.5, it is necessary to avoid h falling into the range (0.2, 0.4, 0.6, 0.8), which would cause the system to exhibit a short-period state, or falling into the range (0, 0.25, 0.5, 0.75), which would cause a subsequent fixed point. Experimental verification shows that setting the parameter β to 0.4 and λ to ∈ (0, 0.3) can achieve better chaotic characteristics. The above control parameters are selected based on the set {h|h=t / (D+1), t=1,2,…,D, h≠β}, where D represents the dimensionality of the search space.

[0127] The improved Tent obtained in this way has better chaotic characteristics, can generate chaotic space in a larger parameter range, and has good security and pseudo-randomness.

[0128] Step 2.3: Implement the Improved Salp Squid Algorithm (IMSSA)

[0129] SSA has the advantages of a simple structure, few control parameters, strong search capabilities, and ease of implementation. In basic SSA, the leader continuously perturbs the vicinity of the food source, and the search range is unrestricted, resulting in slow convergence in the early stages of the algorithm and low convergence accuracy in the later stages of the iteration. Furthermore, the location of the food source in the traditional SSA algorithm plays a crucial role in the optimization process of the salp group. During the search, the location of the individual with the best fitness value is defined as the current food source location. Led by the leader, the salp chain moves towards the food source location. Once the food source location falls into a local optimum, it can easily cause the salp chain to fall into the local optimal area, resulting in a search stall.

[0130] Considering the improved Tent chaotic map has better ergodicity, the salp swarm algorithm is combined with the improved Tent chaotic map to give full play to the respective advantages of the two algorithms and improve the optimization performance of the salp swarm algorithm. Thus, the improved salp algorithm (IMSSA) is realized. During the optimization process, the global exploration and local development capabilities of the algorithm are enhanced. The specific steps are as follows:

[0131] Determine the initial value vector of the Salp Algorithm, obtain the chaotic sequence distributed on [0,1] by formula (10), and use it to form the population position. Then, map the elements in the sequence to the corresponding solution domain by formula (11).

[0132]

[0133] in, are the values ​​of the nth and n+1th salp individuals represented by chaotic variables in dimension d, ub d and lb d are the upper and lower bounds of the dth dimension of the corresponding solution domain, represents the value of the nth salp in dimension d, The value of the salp individual represented by the nth chaotic variable in dimension d generate.

[0134] The search process based on the improved Tent chaos map global perturbation is as follows: Figure 4 In the process of using the improved Tent chaotic map to perform SSA global perturbation search, first, the improved Tent chaotic map is used to generate N salps, which are mixed with the original N salps and sorted according to fitness values. Then, the top N excellent salps after mixing are selected to replace the original N salps, thus completing the global perturbation. The specific steps are as follows:

[0135] Step 1: Set n = 1;

[0136] Step 2: Use the improved Tent chaos map to perform global perturbation design and generate N Salp chaotic variables with different trajectories in, D is the dimension, N is the number of salps, Indicates the nth salp in dimension d The value on

[0137] Step 3: Set d = 1;

[0138] Step 4: All components of are mapped to the optimization variable value interval [x min ,x max ], and then get

[0139]

[0140] Among them, x j-max and x j-min are the maximum and minimum values ​​of the d-th dimension vector respectively.

[0141] Step 5: If d ≥ D, go to Step 7, otherwise go to Step 6;

[0142] Step 6: Set d = d + 1 and go to Step 4;

[0143] Step 7: If n ≥ N, go to Step 9, otherwise go to Step 8;

[0144] Step 8: Set n = n + 1 and go to Step 2;

[0145] Step 9: Calculate the fitness value of each salp, mix it with the original salps, and sort it from large to small according to the fitness value. Replace the original salps with the N salps with the highest fitness value to form a new population, completing the global search based on chaotic mapping.

[0146] Step (3): Establish an aquaculture water quality prediction model based on IMSSA-NARX

[0147] Step 3.1: Build the NARX neural network model

[0148] Dynamic neural networks have more advantages in nonlinear system modeling and are widely used in the modeling of nonlinear dynamic systems with time-delay characteristics. Common dynamic neural networks include fixed-delay neural networks (Focused Time-delay Neural Network, FTDNN) and nonlinear autoregressive neural networks (Nonlinear Auto Regressive Exogenous, NARX). Among them, FTDNN is used to solve time series regression problems, but FTDNN ignores the important correlation between the input time series and the predicted output value. Moreover, the FTDNN model only delays the input and is not related to the delayed signal of the predicted output, so there are certain limitations. Compared with FTDNN, NARX) neural network is a nonlinear dynamic neural network with external output feedback. NARX takes into account the effects of input and output and their delays on the model, and greatly improves the accuracy and robustness of the constructed model by introducing delay and feedback links. Another advantage of NARX is that it requires less historical data. Therefore, this patent is completed using NARX neural network. The basic structure of NARX neural network is as follows Figure 5 :

[0149] The NARX network model can be expressed as:

[0150] y(n)=f NARX (x(n),x(n-1),…,x(n-α),…x(nk),y(n-1),…,y(n-γ),…,y(nm)) (13)

[0151] As can be seen from the above formula, the output y(n) is not only related to the current input x(n), but also to the past input x(n-α); α = 1, 2, ..., k, and the past output y(n-γ); γ = 1, 2, ..., m, where k is the maximum delay length of the input and m is the maximum delay length of the output feedback.

[0152] In the specific implementation of the NARX network model, the output of the first layer of neurons is defined as:

[0153]

[0154] Where, is the output of the βth neuron in the first layer, K represents the number of inputs x received by the βth neuron in the first layer, is the weight of the αth input x received by the βth neuron, is the input, i.e., represents the input x(n) and K delayed signals of x(n), i.e., x(n-α); α=1,2,…,K. is the weight of the γth delayed output y received by the βth neuron, represents the γth delayed output from the past received by the βth neuron, i.e., y(n-γ); γ = 1, 2, …, M, where M represents the number of delayed outputs y. is the bias, f 1 is the hidden layer activation function. In implementation, the tansig function is used.

[0155] Similarly, the output of the second layer of neurons is Defined as:

[0156]

[0157] Where, is the output of the βth neuron in the second layer, h represents the number of βth neurons in the second layer that receive outputs from the first layer, is the weight of the αth input received by the βth neuron, is the input, which represents the output signal from the first layer is the bias, f 2 It is the activation function of the output layer. In implementation, purelin function is used.

[0158] It can be seen that on the basis of having an input delay structure, the NARX neural network also adds an output delay structure, which not only takes into account the current and past inputs, but also the influence of past outputs.

[0159] Step 3.2: Build the NARX neural network prediction model based on IMSSA

[0160] Although the NARX neural network has great advantages in modeling nonlinear dynamic systems, the model also has certain disadvantages. The initial weights of the NARX neural network are somewhat random, and the random and relatively blind initial weights will directly affect the convergence speed of the network, resulting in a lack of network stability. Therefore, this paper proposes an improved Salp Squid algorithm (IMSSA) and uses IMSSA to train the NARX neural network model to optimize the weights and thresholds of the NARX neural network. The IMSSA algorithm has a strong global optimization ability and can reduce the search range to the vicinity of the better weights. The proposed prediction model structure of the NARX neural network based on IMSSA is as follows: Figure 6 shown.

[0161] Step (4) is to complete the aquaculture water quality prediction experiment based on IMSSA-NARX and compare this method with the conventional method. The details are as follows:

[0162] A data set was obtained from an experimental fish pond, and the performance of the proposed model was evaluated through these experimental data. The aquaculture water quality prediction experiment based on IMSSA-NARX was completed. Based on the above experimental environment, acquisition system and experimental data set, the IMSSA-NARX model proposed in this patent was used to process and predict three aquaculture water quality parameters. Table 4 shows the experimental results of a set of environmental parameters detected by the sensor after prediction:

[0163] Table 4 Prediction results of aquaculture water conditions

[0164]

[0165] Secondly, in order to examine the effectiveness and comparability of the proposed model, we used support vector machine (SVM), RBF, BP network and the IMSSA-NARX network proposed in this paper to establish water quality prediction models, predict the aquaculture water quality environment conditions, and compare the prediction effects of each model. The experimental results are shown in Table 5:

[0166] Table 5 Comparison of prediction performance of different algorithms

[0167]

[0168] As shown in Table 5, the accuracy rates of the training and test sets of the SVM are 85% and 81.76%, respectively; the accuracy rates of the training and test sets of the IMSSA-NARX are 90.02% and 84.66%, respectively; the accuracy rates of the training and test sets of the RBF are 79.48% and 76.7%, respectively; and the accuracy rates of the training and test sets of the BP are 76.08% and 71.44%, respectively. This shows that compared with other conventional models, this method also shows better monitoring results. The IMSSA-NARX network model performs best, followed by the SVM. The performance of the RBF and BP models is similar, but their results are relatively poor. This shows that the IMSSA-NARX network model has good performance in predicting aquaculture water quality conditions. It can reasonably predict the water environment conditions and provide more accurate predictions for aquaculture, thereby laying the foundation for improving the management process of aquaculture farms and early detection of water quality parameters, and further promoting the green and sustainable development of the aquaculture industry.

[0169] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A method for predicting aquaculture water quality based on IMSSA-NARX, characterized in that: The following steps are involved: Step 1: Collect aquaculture water quality parameters, establish a data set, and complete preprocessing; Step 2: By introducing multi-bit control parameters β and λ into Tent's chaotic equation, the Tent chaotic map is improved, and the salp swarm algorithm is combined with the improved Tent chaotic map to obtain the improved salp algorithm IMSSA; Step 3: The improved Salp Intense Squid algorithm IMSSA is used to train the NARX neural network model to optimize the weights and bias of the NARX neural network, thereby establishing an aquaculture water quality prediction model based on IMSSA-NARX. Step 4: Use the optimized IMSSA-NARX-based aquaculture water quality prediction model to predict aquaculture water quality.

2. A method for predicting aquaculture water quality based on IMSSA-NARX according to claim 1, characterized in that: In the step 1, multiple groups of aquaculture water quality parameters including temperature, pH value, and dissolved oxygen (DO) are collected, and after normalization, they are divided into training data and test data in a ratio of 6:

4.

3. A method for predicting aquaculture water quality based on IMSSA-NARX according to claim 1, characterized in that: In step 2, multiple control parameters β and λ are introduced into Tent's chaos equation, specifically: Among them, h n is the nth iteration value of the mapping variable h, λ is a multi-bit control parameter, and β is the initial control parameter. When β is 0.5, the value of the mapping variable h is prevented from falling into the range (0.2, 0.4, 0.6, 0.8), which may cause the system to exhibit a short-period state, or from falling into the range (0, 0.25, 0.5, 0.75), which may cause a fixed point to appear later. The control parameters are selected based on the set {h|h=t / (D+1), t=1, 2, …, D, h≠β}, where D represents the dimensionality of the search space.

4. A method for predicting aquaculture water quality based on IMSSA-NARX according to claim 3, characterized in that: The improved Salp Intense Squid algorithm IMSSA is based on the search of improved Tent chaos map global perturbation, as follows: Determine the initial value vector of the Salp Algorithm, obtain the chaotic sequence distributed on [0,1] by the following formula to form the population position, and map the elements in the sequence to the corresponding solution domain: in, are the values ​​of the nth and n+1th salp individuals represented by chaotic variables in dimension d, ub d and lb d are the upper and lower bounds of the dth dimension of the corresponding solution domain, represents the value of the nth salp in dimension d, The value of the salp individual represented by the nth chaotic variable in dimension d generate; In the process of using the improved Tent chaotic map to perform global perturbation search for the salp algorithm SSA, first, the improved Tent chaotic map is used to generate N salps, which are mixed with the original N salps and sorted according to their fitness values. Then, the top N excellent salps after mixing are selected to replace the original N salps, thus completing the global perturbation.

5. A method for predicting aquaculture water quality based on IMSSA-NARX according to claim 1, characterized in that: In step 3, the improved salp integument algorithm IMSSA is used to train the NARX neural network model, as follows: Step 3.1: Construct the NARX neural network model. The NARX neural network model is expressed as: y(n)=f NARX (x(n),x(n-1),…,x(n-α),…x(nk),y(n-1),…,y(n-γ),…,y(nm)) As can be seen from the above formula, the output y(n) is not only related to the current input x(n), but also to the past input x(n-α); α = 1, 2, ..., k, and the past output y(n-γ); γ = 1, 2, ..., m, where k is the maximum delay length of the input and m is the maximum delay length of the output feedback. In the specific implementation of the NARX neural network model, the output of the first layer of neurons is defined as: Where, is the output of the βth neuron in the first layer, K represents the number of inputs x received by the βth neuron in the first layer, is the weight of the αth input x received by the βth neuron, is the input, i.e., x(n) and K delayed signals of x(n), i.e., x(n-α); α=1,2,…,K, is the weight of the γth delayed output y received by the βth neuron, represents the γth delayed output received by the βth neuron from the past, i.e. y(n-γ); γ=1,2,…,M, where M represents the number of delayed outputs y, is the bias, f 1 is the hidden layer activation function; Similarly, the output of the second layer of neurons is Defined as: Where, is the output of the βth neuron in the second layer, h represents the number of βth neurons in the second layer that receive outputs from the first layer, is the weight of the αth input received by the βth neuron, is the input, which represents the output signal from the first layer is the bias, f 2 is the output layer activation function; Step 3.2: Use the improved Salp Sea Squid algorithm IMSSA to train the NARX neural network model to optimize the weights of the NARX neural network and bias 6. A method for predicting aquaculture water quality based on IMSSA-NARX according to claim 5, characterized in that: The improved Salp Intense Squid Algorithm (IMSSA) is used to train the NARX neural network model. Specifically: Step 1: Set initial parameters, including the maximum number of iterations l max , population number N, dimension D, initialization population X N×D , randomly initialize the individual position coordinates of the population; Step 2: Calculate the individual fitness value of each salp according to the objective function; Step 3: Select the initial location of the food source, sort the fitness values ​​of the salps, and set the coordinates of the salp with the best fitness as the target location, which is the location of the food source; Step 4: Determine the leader and followers. Set the individuals in the first half of the salp chain as leaders, and the rest as followers. Step 5: Update the position of the salp leader: Among them, F d is the target position value of the dth dimension, ub d , lb d are the upper and lower bounds of the dth dimension, respectively. c1, c2, and c3 are control parameters. c2 and c3 are random numbers between [0, 1] that determine the direction and step size of the leader position update. c1 is the convergence factor, which is used to balance the convergence speed of the algorithm during the iteration process. Step 6: Update the position of the salp follower: Where n≥2, Respectively represent the position coordinates of the nth and n-1th salps in dimension d; Step 7: Perform boundary processing on each dimension of the updated individual, calculate the fitness value of each salp, and compare it with the fitness value of the current target position to update the global optimal salp target position, that is, update the location of the food source; Step 8: Determine whether the number of iterations is met. If so, output the result; otherwise, go to Step 4 to continue iterative evolution, and finally output the current optimal fitness and target position.

Images