Rice noodle drying system and method based on artificial neural network prediction

Abstract

The present application relates to rice line drying technical field, disclose based on artificial neural network prediction rice line drying system and method, including S1, data acquisition;S2, artificial neural network design;S3, optimization algorithm: the introduction of three kinds of algorithm through group iterative search effectively reduces the uncertainty brought by random initialization to improve the prediction effect of BP neural network;S4, network model specific parameter setting: control variable method is used, all model's basic network parameter is kept completely identical, the population size and evolution generation of optimization algorithm are also kept uniform;S5, data statistical analysis;S6, result and analysis.The present application is based on the rice line hot air drying kinetics test, introduces artificial neural network, carries out intelligent prediction rice line moisture ratio research, through the construction of basic BP neural network and adds three kinds of intelligent optimization algorithm to build the improved model and carries out comparative analysis, so that the prediction result of BP neural network to rice line hot air drying moisture ratio is better.

Description

Technical Field

[0001] This invention relates to the field of rice noodle drying technology, specifically a rice noodle drying system and method based on artificial neural network prediction. Background Technology

[0002] Hot air drying is the mainstream drying method in rice noodle production, and accurate prediction of the moisture ratio (MR) is a key prerequisite for determining the endpoint of hot air drying and for stable process control. Traditional mathematical models are difficult to accurately describe the dynamic characteristics of the drying process, thus increasing energy consumption fluctuations and quality risks. Moreover, moisture changes during the drying process are affected by a variety of drying factors, and fitting a single model can easily lead to defects and constraints in the prediction results.

[0003] Artificial neural networks are computational models that mimic the human brain's information processing. They consist of numerous simple processing units called neurons. A neuron receives input signals from other neurons, weights these signals, and then uses an activation function to determine its output. Traditional backpropagation (BP) neural network models are highly dependent on initial weight and threshold settings, making them prone to getting trapped in local optima, having limited convergence speed, and struggling to determine structural parameters. Furthermore, they exhibit insufficient stability when sample errors and drying processes differ significantly, limiting their reliability in process prediction and online control scenarios.

[0004] Therefore, it is necessary to design a rice noodle drying system and method based on artificial neural network prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a rice noodle drying system and method based on artificial neural network prediction, so as to solve the problems in the prior art.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] This relates to the field of rice noodle drying technology, specifically a rice noodle drying system based on artificial neural network prediction, including:

[0008] S1. Data Acquisition: Collect the drying data of rice noodles and perform normalization processing;

[0009] S2. Artificial Neural Network Design: Adaptive learning and training are performed using a BP neural network selection function, and regression prediction with mean squared error as the performance function is adopted.

[0010] S3. Optimization Algorithm: Firefly algorithm, particle swarm algorithm and genetic algorithm in evolutionary computation are introduced. All three types of algorithms effectively reduce the uncertainty caused by random initialization through group iterative search to improve the prediction effect of BP neural network.

[0011] S4. Specific parameter settings for the network model: The control variable method was adopted to keep the basic network parameters of all models completely consistent, and the population size and number of generations of the optimization algorithm were also kept uniform.

[0012] S5. Data Statistical Analysis: Origin 2024 statistical software was used for plotting, and Excel was used for data processing.

[0013] S6. Results and Analysis: The BP neural network showed good prediction results for the moisture ratio of rice noodles after hot air drying, with a model R² of 0.99175.

[0014] Furthermore, the specific operation of S2 is as follows: the BP neural network implements weight correction based on the gradient descent strategy. It generally consists of three parts: input layer, hidden layer and output layer. The layers are fully interconnected, but there is no mutual connection between units in the same layer.

[0015] Network structure design: Temperature, humidity, load and drying time that affect the moisture ratio are used as the input layer of the neural network, that is, the input layer has 4 neurons, and the moisture ratio is used as the output layer, that is, the output layer has 1 neuron.

[0016] The performance of the number of nodes in a single hidden layer of the network was tested using software. After multiple experiments, the network topology was finally determined.

[0017] Furthermore, step S2 also includes a method for determining the hidden layers, and the formula for calculating the number of hidden layers is as follows:

[0018] k < m + n + b (6-1)

[0019] k = 2n + 1 (6-2)

[0020] Where n is the number of neurons in the input layer; k is the number of neurons in the hidden layer; m is the number of neurons in the output layer; and b is a constant between 0 and 10.

[0021] Furthermore, the firefly algorithm optimization operation in S3 includes:

[0022] When constructing the FA-BP optimization model based on the firefly algorithm, the BP network weights and threshold vectors are used as the "positions" of the fireflies, and the training set MSE is used as the fitness evaluation metric. Iterative updates of FA yield parameter combinations with smaller errors, replacing the random initialization of parameters in the BP network. This reduces the risk of BP training getting trapped in local optima and improves training stability.

[0023] Furthermore, the genetic algorithm optimization operation in S3 includes:

[0024] In the genetic algorithm optimization model GA-BP, the weights and thresholds of the BP network are encoded as real numbers to form chromosomes. The training set MSE is used as the fitness function. The chromosome with the highest fitness is obtained through GA search and decoded into the initial parameters of the BP network. Then, Adam is used to complete local fine training to improve the final prediction accuracy and generalization performance.

[0025] Furthermore, the particle swarm optimization operation in S3 includes:

[0026] When constructing the Particle Swarm Optimization (PSO-BP) model, the BP network weights and threshold vectors are used as particle positions, and the training set MSE is used as the fitness function. PSO first performs a global search to obtain optimal parameters, and then uses these parameters as initial values ​​for the BP network. Adam is then used for local training, thereby improving convergence stability and reducing the risk of local optima.

[0027] Furthermore, the specific operations of S6 include: model training and testing, neural network regression analysis, and relative error analysis;

[0028] Based on the dynamics experiment of hot air drying of rice noodles, an artificial neural network was introduced to conduct intelligent prediction of the moisture ratio of rice noodles. An improvement model was constructed by building a basic BP neural network and adding three intelligent optimization algorithms, and comparative analysis was carried out.

[0029] The rice noodle drying method based on artificial neural network prediction is applied to the aforementioned rice noodle drying system based on artificial neural network prediction. The drying method includes the following steps:

[0030] The rice noodles are first aged, then placed in a drying device to remove the dried rice noodles, and the drying characteristics of the rice noodles are tested before being packaged into finished products.

[0031] The beneficial effects of this invention are:

[0032] 1. This invention discloses a rice noodle drying system and method based on artificial neural network prediction. Based on the dynamic experiment of hot air drying of rice noodles, an artificial neural network is introduced to conduct intelligent prediction of the moisture ratio of rice noodles. By constructing a basic BP neural network and adding three intelligent optimization algorithms (FA-BP, GA-BP, PSO-BP) to build an improved model and conduct comparative analysis, the BP neural network has better prediction results for the moisture ratio of rice noodles during hot air drying. The model R2 is 0.99175, which is very close to the results of the traditional model.

[0033] 2. This invention discloses a rice noodle drying system and method based on artificial neural network prediction. The introduction of intelligent optimization strategies can effectively improve the training convergence and prediction accuracy of the BP network. Among them, the genetic algorithm (GA-BP) and particle swarm optimization (PSO-BP) performed best, with test set determination coefficients (R²) of 0.99685 and 0.99648, respectively, and average relative errors as low as 6.5% and 6.2%, which are significantly better than the traditional BP model (R²=0.99464, error 9.1%). The performance of the firefly algorithm (FA-BP) is between that of the basic BP and the former two.

[0034] 3. The present invention provides a rice noodle drying system and method based on artificial neural network prediction. Each model has good generalization ability and no overfitting. GA-BP and PSO-BP have comprehensive advantages in fitting accuracy, convergence speed and robustness, and can be used as intelligent models for predicting the moisture ratio of hot air drying, with a wide range of applications. Attached Figure Description

[0035] The invention will now be further described with reference to the accompanying drawings.

[0036] Figure 1 This is a schematic diagram of the rice noodle drying system based on artificial neural network prediction of the present invention;

[0037] Figure 2 This is a schematic diagram of the BP neural network model structure of the present invention;

[0038] Figure 3 This is a flowchart of the FA-BP neural network of the present invention;

[0039] Figure 4 This is a flowchart of the GA-BP neural network of the present invention;

[0040] Figure 5 This is a flowchart of the PSO-BP neural network of the present invention;

[0041] Figure 6 This is a graph showing the loss value of the BP neural network of this invention;

[0042] Figure 7 This is a fitness curve diagram of different optimization algorithms of this invention;

[0043] Figure 8 This is a schematic diagram of the BP neural network training regression of the present invention;

[0044] Figure 9 This is a schematic diagram of the BP neural network training regression of the present invention;

[0045] Figure 10 This is the regression graph of the FA-BP neural network training in this invention;

[0046] Figure 11This is the regression graph of the FA-BP neural network training in this invention;

[0047] Figure 12 This is the regression graph of the GA-BP neural network training in this invention;

[0048] Figure 13 This is the regression graph of the GA-BP neural network training in this invention;

[0049] Figure 14 This is the regression graph of the GA-BP neural network training in this invention;

[0050] Figure 15 This is the regression graph of the PSO-BP neural network training in this invention;

[0051] Figure 16 This is the regression graph of the PSO-BP neural network training in this invention;

[0052] Figure 17 This is a comparison chart of the predicted and output values ​​of the neural network models BP(A), FA-BP(B), GA-BP(C), and PSO-BP(D) of this invention;

[0053] Figure 18 This is a relative error analysis diagram of the four artificial neural network models of this invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Example

[0056] A rice noodle drying system and method based on artificial neural network prediction, wherein the rice noodle raw material is indica rice with an amylose content of about 30%, which is provided by Anhui Wangrenhe Rice Noodle Food Co., Ltd.

[0057] The rice noodle drying equipment includes: DHG-9003 electric thermostatic drying oven (Shanghai Heng Scientific Instruments Co., Ltd.); and JA5003 electronic balance (Shanghai Puchun Measurement Instruments Co., Ltd.).

[0058] The process for preparing rice noodles is as follows: Indica rice → crushing → sieving → adding water to make a paste → steaming and gelatinizing → extrusion molding → aging → drying → packaging → finished product.

[0059] The drying method includes the following steps:

[0060] The rice noodles are first aged, and then placed in a drying device. The drying conditions include a drying temperature of 40-70℃, a drying humidity of 10%-50%, and drying for 30-60 minutes. After drying, the dried rice noodles are removed, and their drying characteristics, such as dry basis moisture content, are tested before packaging.

[0061] Based on the analysis of the hot air drying principle and preliminary experiments, the material properties of rice noodles were obtained, and the final experimental parameter range was determined. Drying humidity (10%, 20%, 30%, 40%, 50%), sample load (20%, 40%, 60%, 100g), and drying temperature (40%, 50%, 60%, 70℃) were used as experimental factors to study the effects of humidity, load, and temperature on the hot air drying characteristics of rice noodles. The experimental design scheme is shown in Table 2-1.

[0062] Table 2-1 Experimental Design of Hot Air Drying Characteristics

[0063]

[0064] The specific steps are as follows:

[0065] Before drying, adjust the equipment to the preset parameters and preheat for 10 minutes. Spread the rice noodle sample flat on the sample sieve and maintain a constant airflow speed of 1 m / s. Measure and record the mass change of the sample every 50 minutes. Stop the experiment when the rice noodles are dried to a dry basis moisture content ≤10%, remove the dried sample, seal and store it. Repeat each experiment three times and take the average value.

[0066] Nine common drying kinetic models were selected (see Table 2-2). The changes in moisture ratio and drying rate during the hot air drying process of rice noodles were investigated, and the model most suitable for describing the process was selected. The goodness of fit of the models was evaluated using the coefficient of determination (R²), chi-square test (χ²), sum of squared residuals (SSE), and root mean square error (RMSE).

[0067] Table 2-2 Hot air drying kinetic model

[0068]

[0069] In the table above, t represents the drying time; a, b, c, d, and k are the parameters to be determined for each model equation.

[0070] In this embodiment, the drying characteristics of rice noodles within the drying system are as follows:

[0071] Dry basis moisture content: The calculation method for the dry basis moisture content of rice noodles is as follows.

[0072] (2-1)

[0073] In the formula: Ma is the dry basis moisture content of the rice noodle sample, g / g; mt is the mass of the rice noodle sample dried to time t, g; mb is the mass of the rice noodles at the end of drying, g.

[0074] Moisture ratio: The calculation method for the moisture ratio of rice noodles is as follows:

[0075] (2-2)

[0076] In the formula: M0 is the initial dry basis moisture content of the rice noodles, g / g; Mt is the dry basis moisture content of the rice noodles at time t after drying, g / g; Mb is the dry basis moisture content at the end of drying, g / g.

[0077] Drying rate: The drying rate refers to the amount of moisture lost from the inside of rice noodles per unit time. The calculation method is as follows:

[0078] (2-3)

[0079] In the formula: DR is the drying rate of rice noodles, g·g-1·min-1; Wc is the dry basis moisture content of rice noodles at time tc, g / g; Wd is the dry basis moisture content of rice noodles at time td, g / g; td-tc represents the time interval between adjacent sampling points, min.

[0080] Effective moisture diffusion coefficient: The calculation of the effective moisture diffusion coefficient Deff can refer to Miao Junwei's method and be modified. Assuming that the physical model of dried rice noodles is a cuboid, moisture will diffuse along the length, width, and height directions. Moreover, the diffusion that occurs during the drying process of rice noodles is all one-dimensional axial diffusion. Then, the simplified solution of Fick's law is as follows:

[0081] (2-4)

[0082] In the formula: D is the effective moisture diffusion rate, m2 / s; L is half the material thickness, m; n is an infinite series; t is time, min.

[0083] Since the width and height of the sample are much smaller than its length, it can be assumed that the diffusion of moisture mainly occurs in the width and height directions. Therefore, to simplify the calculation equations, when n = 0, equations (2-2) and (2-4) can be transformed into equation (2-5).

[0084] (2-5)

[0085] In the formula: Ly and Lz are half the width and half the thickness of the rice noodle sample, respectively, in meters; t is the drying time, in minutes.

[0086] Taking the natural logarithm of both sides of equation (2-5) yields equation (2-6).

[0087] (2-6)

[0088] From formula (2-6), we can see that lnMR is linearly related to time t. After linear fitting, we can obtain the slope K and then calculate the moisture diffusion coefficient Deff according to formula (2-7).

[0089] (2-7)

[0090] Drying activation energy: The relationship between drying activation energy and drying temperature can be expressed by the Arrhenius equation, as follows:

[0091] (2-9)

[0092] In the formula: D0 is the diffusion constant of the material, m2 / s; Ea is the drying activation energy, kJ / mol; R is the gas molar constant, 8.314 × 10⁻³ kJ / (mol·K); T is the drying temperature, ℃. Taking the natural logarithm of both sides of equation (2-9) yields equation (2-10). From equation (2-10), it can be seen that lnDeff has a linear relationship with 1 / (T+273.15). Fitting the linear equation, the slope k = -Ea / R can be obtained, and thus Ea can be calculated.

[0093] (2-10)

[0094] Thermodynamic parameters: enthalpy change refers to the amount of heat absorbed or released during a chemical reaction under constant temperature and pressure conditions; Gibbs free energy is a thermodynamic state function introduced to explore the direction and limits of thermodynamic processes; and entropy change refers to the measure of the disorder of a system. The three are calculated according to formulas (2-11) to (2-14) respectively.

[0095]

[0096] In the above formula: A: pre-exponential constant, min⁻¹; R: gas molar constant, 8.314 × 10⁻³ kJ / (mol·K); T: drying temperature, K; kᵦ: Boltzmann constant, 1.381 × 10⁻²³ J / K; h: Planck constant, 6.626 × 10⁻³ 4 J.s.

[0097] The factor effect analysis is as follows:

[0098] Between-subjects effects tests are typically used to assess the magnitude of a factor's influence on an indicator. This test calculates the F-value as the ratio of between-group variation to within-group error variation, and provides the corresponding P-value and the partial eta squared (η²) of the effect size. When P < 0.05, it indicates that the factor has a significant effect on drying time. The partial eta squared (η²) quantifies the magnitude of the factor's effect; its value is typically between 0 and 1, with values ​​closer to 1 indicating a stronger actual impact on drying.

[0099] In this embodiment, the rice noodle drying system based on artificial neural network prediction includes the following steps:

[0100] S1. Data Acquisition:

[0101] A total of 156 sets of data were collected under the hot air drying conditions of rice noodles, from the initial moisture ratio to the completion of drying. 80% of the data (124 sets) was randomly selected from these 156 sets as the training set for the neural network, and the remaining 20% ​​(32 sets) was used as the test set. Because the input variables (humidity, load, temperature) and output variable (moisture ratio) of the drying experiment data have different physical dimensions and significant numerical differences, directly inputting them into the network could lead to model convergence difficulties or errors due to numerical overflow. Therefore, this study used the min-max normalization method to process the data. The MinMaxScaler function from Python's Scikit-learn library was used to normalize the data, limiting it to the range [-1, 1].

[0102] S2. Artificial Neural Network Design:

[0103] The BP neural network uses the backpropagation algorithm to correct weights based on the gradient descent strategy. It generally consists of three parts: an input layer, a hidden layer, and an output layer. The layers are fully interconnected, but there is no connection between units in the same layer.

[0104] Network Structure Design: During the hot air drying experiment of rice noodles, temperature, humidity, load, and drying time all affect the change in moisture ratio. Therefore, these four factors are used as the input layer of the neural network, i.e., the input layer has 4 neurons, and the moisture ratio is used as the output layer, i.e., the output layer has 1 neuron. The number of hidden layer nodes in a BP neural network has a significant impact on the prediction accuracy of the BP neural network, but there is currently no fixed method to determine the number of hidden layer neurons. Too few hidden layer neurons will lead to an excessively long machine learning process or the learning effect will not meet expectations; too many hidden layer neurons will lead to an excessively long machine learning time or overfitting of learning accuracy, resulting in a decrease in learning accuracy. The number of hidden layers can be determined with reference to the following formula:

[0105] k < m + n + b (6-1)

[0106] k = 2n + 1 (6-2)

[0107] Where n is the number of neurons in the input layer; k is the number of neurons in the hidden layer; m is the number of neurons in the output layer; and b is a constant between 0 and 10.

[0108] In this embodiment, Python 3.10 64-bit software was used to test the performance of the network with 3 to 12 hidden layer nodes. Referring to the above formulas and multiple test results, the impact of different hidden layer node numbers on the model's prediction performance was compared. After multiple experiments, the final network topology was determined, as follows: Figure 2 As shown.

[0109] The selection of functions in BP neural networks includes transfer (activation function), training and learning (solution function), network performance, and simulation functions. Typically, the transfer function determines the network's nonlinear expressive power. Commonly used transfer functions for BP networks include log-sigmoid (logsig), hyperbolic tangent (tanh / tansig), rectified linear function (ReLU), and purelin.

[0110] The output values ​​of the logsig function are (0, 1); the output value of the tanh function is (-1, 1); the output value of the linear rectifier function is [0, +∞); and the output value of the linear function purelin can take any value.

[0111] In this embodiment, tanh is used as the activation function for both the hidden and output layers of the neural network. This function is a non-linear curve function, which is often more natural than linear functions when fitting continuously changing physical processes, such as water drying. During the training phase, an adaptive moment estimation optimization algorithm is used to update the weights and thresholds. Adam achieves adaptive adjustment of the learning rate by adaptively estimating the first and second moments of the gradient, thereby improving convergence speed and training stability. The mean squared error is used as the performance function for regression prediction.

[0112] S3, Optimization Algorithm:

[0113] To improve the globality and training stability of parameter optimization, this study introduces three typical intelligent optimization algorithms to optimize the initial weights and thresholds of the BP neural network: the Firefly Algorithm (FA) and Particle Swarm Optimization (PSO) in swarm intelligence algorithms, and the Genetic Algorithm (GA) in evolutionary computation. All three algorithms effectively reduce the uncertainty caused by random initialization through swarm iterative search, thereby improving the prediction performance of the BP neural network.

[0114] S3.1 FA-BP Neural Network Construction Process

[0115] The Firefly Algorithm (FA) is a stochastic optimization method based on swarm intelligence. Its core idea is to use "brightness" to represent the quality of individuals and to achieve optimal movement of the swarm in the solution space through "brightness attraction." Individuals with better fitness have higher brightness, and other individuals move closer to them under the attraction.

[0116] Firefly movement follows these rules: 1. Regardless of sex, every firefly is attracted to other brightly glowing fireflies; 2. Fireflies with lower brightness are attracted to brighter fireflies and move towards them, and their light sensitivity decreases as the distance between them increases; 3. If a firefly has no brighter fireflies nearby, it will move randomly. Because attraction decreases with distance, FA (Firefly Analysis) can achieve a relative balance between global exploration and local exploitation. Therefore, under multi-modal objective function conditions, it has a good ability to escape local optima and typically exhibits fast convergence characteristics.

[0117] In constructing the FA-BP model, this study uses the BP network weights and threshold vectors as the "positions" of the fireflies, and the training set MSE as the fitness evaluation metric. Iterative updates of the FA model yield parameter combinations with smaller errors, replacing the random initialization of parameters in the BP network. This reduces the risk of BP training getting trapped in local optima and improves training stability. The FA-BP modeling process is as follows: Figure 3 As shown.

[0118] S3.2 GA-BP Neural Network Construction Process

[0119] Genetic algorithms (GA) fall under the category of evolutionary computation. Their optimization process simulates the natural process of "survival of the fittest" in genetic evolution. Through operations such as selection, crossover, and mutation, the population iterates generation by generation, gradually approaching the global optimum. GA does not rely on the differentiability or strict mathematical form of the objective function and is highly adaptable to nonlinear, non-convex, and multimodal problems. Furthermore, the crossover and mutation mechanisms maintain population diversity, thereby reducing the possibility of premature convergence and getting trapped in local optima. Its drawbacks include relatively high computational cost per generation iteration, and the fact that parameter settings have a certain impact on convergence efficiency.

[0120] In the GA-BP model, the weights and thresholds of the BP network are encoded as real numbers to form chromosomes. The training set MSE is used as the fitness function. The chromosome with the highest fitness is obtained through GA search and decoded into the initial parameters of the BP network. Subsequently, Adam is used for local fine-tuning to improve the final prediction accuracy and generalization performance. The GA-BP modeling process is as follows: Figure 4 As shown.

[0121] S3.3, PSO-BP Neural Network Construction Process

[0122] Particle Swarm Optimization (PSO) is a swarm intelligence algorithm that simulates the collaborative search behavior of a group. PSO represents candidate solutions as particles, which have position and velocity in the search space. Their position updates are guided by both their individual historical best position and the swarm's global best position. PSO algorithms have a simple structure and relatively few parameters, and typically exhibit fast convergence speeds in continuous high-dimensional parameter optimization, making them suitable for optimizing high-dimensional continuous variables such as neural network weights and thresholds.

[0123] In constructing the PSO-BP model, this study uses the BP network weights and threshold vectors as particle positions, and the training set MSE as the fitness function. PSO first performs a global search to obtain optimal parameters, then uses these parameters as initial values ​​for the BP network, and employs Adam for local training, thereby improving convergence stability and reducing the risk of local optima. The PSO-BP modeling process is as follows: Figure 5 As shown.

[0124] S4. Specific parameter settings for the network model

[0125] To ensure a fair performance comparison of the four network models (ordinary BP, GA-BP, FA-BP, and PSO-BP), this study employed a controlled variable method. All models maintained identical basic network parameters (learning rate, number of iterations, and optimizer), and the population size and generation number of the optimization algorithms were also kept consistent. Specific parameter settings are shown in Table 2-3 below.

[0126] Table 2-3 Key Parameter Settings for Artificial Neural Networks and Optimization Algorithms

[0127]

[0128] S5, Data Statistical Analysis

[0129] The experiment used Origin 2024 statistical software for plotting and Excel for data processing. Python was used to write the neural network code. The coefficient of determination (R²) and mean squared error (MSE) were selected as evaluation metrics. A higher R² value and a smaller MSE value indicate higher prediction accuracy and better fitting performance.

[0130] S6. Results and Analysis

[0131] S6.1 Model Training and Testing

[0132] The input layer of a BP neural network has 4 neurons and 10 hidden neurons, and the output layer has 1 neuron. It has a total of 4×10+10×1=50 weights and 10+1=11 thresholds. Therefore, the individual encoding length of the BP neural network is 50+11=61.

[0133] Figure 6 , Figure 7 These are the loss curve of the BP neural network and the fitness iteration curve of the optimization algorithm, derived from... Figure 6 It can be seen that the BP neural network converges after about 10 training iterations. The closer the loss value is to 0, the better the model fit and the more successful the data prediction. During the training process, the loss of the BP neural network decreases rapidly with the number of iterations and gradually approaches stability without significant oscillations or divergence, indicating a relatively stable model training process; the optimal MSE reaches 0.00032. Figure 7 It can be seen that different intelligent optimization strategies can continuously reduce the fitness value of the BP network during the iteration process, and the curve shows an overall step-like decline that tends to stabilize in the later stage. This indicates that the optimization algorithm can continuously refresh better combinations of weights and thresholds, eventually entering the convergence interval. Comparing the three models, GA-BP shows a larger decline and the lowest convergence level in the early stage, and its final fitness is better than PSO-BP and FA-BP. This shows that the genetic algorithm has a stronger global search capability and the advantage of escaping local optima in this research problem, and can provide a better starting point for BP training parameters. PSO-BP also shows a relatively fast early convergence feature, but it is more prone to a plateau in the middle and late stages. Subsequently, it achieves another decrease in fitness through group collaborative information, and its final convergence level is between GA-BP and FA-BP. FA-BP has a relatively slow overall decline and a longer plateau period. Although it can still achieve some improvement in the later stage, its final fitness is still higher than the former two. Based on the results of the two figures, all three optimized BP models can improve the sensitivity of BP to random initialization and enhance training convergence. However, in terms of the overall performance of convergence speed and convergence accuracy, GA-BP has a more prominent advantage, followed by PSO-BP, while FA-BP is relatively weaker.

[0134] S6.2 Neural Network Regression Analysis

[0135] Regression from training set and test set Figures 8-16It can be seen that the predicted and measured values ​​of the four artificial neural network models are generally distributed along the 45° reference line (y=x), and the linear fitting line highly coincides with the reference line, indicating that the model output has a strong consistency with the target value. The determination coefficients of the basic BP neural network on the training set and the test set are 0.99464 and 0.99175, respectively. Although the overall fitting level is high, there are still a few discrete points in the low value region, indicating that the fitting stability of the traditional BP to local samples under random initialization conditions still has room for improvement. After introducing intelligent optimization, the bandwidth of the scatter plot was further narrowed, the linear fit was enhanced, and the fitting accuracy was improved simultaneously: FA-BP's R² on the training and test sets increased to 0.99632 and 0.99402, respectively, indicating that the global search of the Firefly algorithm can effectively improve the initial configuration of weights and thresholds, thereby reducing local errors and improving regression consistency. GA-BP and PSO-BP performed best, with R² on the training and test sets reaching 0.99553 and 0.99504, and 0.99685 and 0.99648, respectively. The scatter plot distribution was compact and had the highest overlap with the fitted line, indicating that the genetic algorithm and particle swarm optimization algorithm have strong global optimization capabilities in the parameter space, which can further reduce the sensitivity of BP to initial weights and the risk of getting trapped in local optima. Overall, the R² on the test set of each model was relatively close to the R² on the training set, and the phenomenon of "high fitting during training and significant degradation during testing" did not occur, indicating that the model has good generalization ability under the current network structure and data partitioning conditions. Considering both the regression fit and the R² level, the model prediction performance was best for GA-BP and PSO-BP, followed by FA-BP, while the basic BP was relatively weak.

[0136] Figure 17The results comparing the expected and predicted outputs of four models—BP, FA-BP, GA-BP, and PSO-BP—on the sample sequence are presented. Overall, the predicted curves of each model can track the fluctuations of the expected output well, with peak and trough positions and inflection points generally consistent, indicating that the network can effectively learn the nonlinear mapping relationship between samples and has a good trend characterization ability. Among them, the traditional BP model exhibits slight amplitude deviations in some high-value peak segments, and the overlap between the predicted curve and the expected output is relatively weak, with a determination coefficient of R² = 0.9905. After introducing intelligent optimization, the fit between the predicted curve and the expected output is further improved, especially with varying degrees of convergence near peaks and troughs: the R² of FA-BP, GA-BP, and PSO-BP increases to 0.99349, 0.99543, and 0.99426, respectively, indicating that the optimization algorithm, by improving the initial configuration of weights and thresholds and the search path, helps to enhance the model's response to complex fluctuations and its fitting consistency. Regarding error metrics, the MSE of each model was in the range of 10⁻² (0.00871 to 0.023734). Among them, GA-BP achieved the highest R² while maintaining a low error level (MSE=0.015524), demonstrating the best overall fitting accuracy and stability. PSO-BP was second (MSE=0.019908), while FA-BP had a relatively large MSE (0.023734), indicating that its error was mainly affected by the amplified bias of a few sample points.

[0137] In summary, all four models can achieve high-precision predictions, but the intelligent optimization strategy as a whole improves the degree to which the predicted output fits the expected output and the quality of parameter finding. Among them, GA-BP has a greater advantage in the overall performance of fitting consistency and error control.

[0138] S6.3 Relative Error Analysis

[0139] From the relative error distribution Figure 18It can be seen that the prediction errors of the four models on the test samples are generally at a low level, and the intelligent optimization strategy can significantly reduce the average relative error of the BP network and improve error fluctuation. The average relative error is 9.1% for traditional BP, 7.6% for FA-BP, 6.5% for GA-BP, and 6.2% for PSO-BP. Compared with BP, the average error of the three optimized models is reduced by about 1.5% to 2.9%, with PSO-BP being the lowest, followed by GA-BP, and FA-BP being relatively higher but still significantly better than BP. This indicates that the optimization algorithm improves the quality of network parameter configuration by globally optimizing the weights and thresholds, thereby reducing the sensitivity of BP to random initialization and reducing the prediction bias caused by local optima. From the distribution characteristics of the error on the sample sequence, most sample points of the four error curves are concentrated in the low error range, and there is no persistently high systematic error, indicating that the model has good adaptability to the overall pattern of the samples. In the early part of the sample, the BP model shows a more prominent error peak (reaching nearly 60%), while the FA-BP, GA-BP, and PSO-BP show significantly lower error peaks at the same positions, and their overall curves are smoother with less dispersion. This reflects that the optimized model not only reduces the average error but also improves prediction stability to some extent. Considering both the average error level and the error fluctuation amplitude, the model prediction performance can be summarized as: PSO-BP ≈ GA-BP > FA-BP > BP; among them, PSO-BP performs best in error control, followed closely by GA-BP, and both are more suitable for scenarios with high prediction accuracy requirements.

[0140] In summary, this embodiment is based on the hot air drying kinetics experiment of rice noodles, and introduces artificial neural networks to conduct intelligent prediction of the moisture ratio of rice noodles. An improvement model is constructed by building a basic BP neural network and adding three intelligent optimization algorithms (FA-BP, GA-BP, PSO-BP) and comparative analysis is carried out.

[0141] The results showed that the ordinary BP neural network predicted the moisture ratio of rice noodles after hot air drying better, with an R2 of 0.99175, which was very close to the results of the traditional model.

[0142] Introducing intelligent optimization strategies can effectively improve the training convergence and prediction accuracy of BP networks. Among them, the Genetic Algorithm (GA-BP) and Particle Swarm Optimization (PSO-BP) performed best, with test set determination coefficients (R²) of 0.99685 and 0.99648, respectively, and average relative errors as low as 6.5% and 6.2%, significantly outperforming the traditional BP model (R²=0.99464, error 9.1%). The Firefly Algorithm (FA-BP) performed between the basic BP and the former two. All models showed good generalization ability and no overfitting. Due to their comprehensive advantages in fitting accuracy, convergence speed, and robustness, GA-BP and PSO-BP can be used as intelligent models for predicting the moisture ratio in hot air drying.

[0143] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A rice noodle drying system based on artificial neural network prediction, characterized in that, include: S1. Data Acquisition: Collect the drying data of rice noodles and perform normalization processing; S2. Artificial Neural Network Design: Adaptive learning and training are performed using a BP neural network selection function, and regression prediction with mean squared error as the performance function is adopted. S3. Optimization Algorithm: Firefly algorithm, particle swarm algorithm and genetic algorithm in evolutionary computation are introduced. All three types of algorithms effectively reduce the uncertainty caused by random initialization through group iterative search to improve the prediction effect of BP neural network. S4. Specific parameter settings for the network model: The control variable method was adopted to keep the basic network parameters of all models completely consistent, and the population size and number of generations of the optimization algorithm were also kept uniform. S5. Data Statistical Analysis: Origin 2024 statistical software was used for plotting, and Excel was used for data processing. S6. Results and Analysis: The BP neural network showed good prediction results for the moisture ratio of rice noodles after hot air drying. Model R 2 It is 0.99175.

2. The rice noodle drying system based on artificial neural network prediction according to claim 1, characterized in that, The specific operation of S2 is as follows: The BP neural network implements weight correction based on the gradient descent strategy. It generally consists of three parts: input layer, hidden layer and output layer. The layers are fully interconnected, but there is no mutual connection between units in the same layer. Network structure design: Temperature, humidity, load, and drying time, which affect the moisture ratio, are used as the input layer of the neural network, i.e., the input layer has 4 neurons. Moisture ratio is used as the output layer, i.e., the output layer has 1 neuron. The performance of the number of nodes in a single hidden layer of the network is tested using software. After multiple experiments, the network topology is finally determined.

3. The rice noodle drying system based on artificial neural network prediction according to claim 2, characterized in that, in, S2 also includes a method for determining hidden layers, and the formula for calculating the number of hidden layers is as follows: k < m + n + b (6-1) k = 2n + 1 (6-2) Where n is the number of neurons in the input layer; k is the number of neurons in the hidden layer; m is the number of neurons in the output layer; and b is a constant between 0 and 10.

4. The rice noodle drying system based on artificial neural network prediction according to claim 1, characterized in that, The firefly algorithm optimization operation in S3 includes: When constructing the FA-BP optimization model based on the Firefly Algorithm, the BP network weights and threshold vectors are used as the "positions" of the fireflies, and the training set MSE is used as the fitness evaluation metric. Iterative updates of FA yield parameter combinations with smaller errors, replacing the random initialization of BP network parameters. This reduces the risk of BP training getting trapped in local optima and improves training stability.

5. The rice noodle drying system based on artificial neural network prediction according to claim 4, characterized in that, The genetic algorithm optimization operation in S3 includes: In the genetic algorithm optimization model GA-BP, the weights and thresholds of the BP network are encoded as real numbers to form chromosomes. The training set MSE is used as the fitness function. The chromosome with the highest fitness is obtained through GA search and decoded into the initial parameters of the BP network. Then, Adam is used to complete local fine training to improve the final prediction accuracy and generalization performance.

6. The rice noodle drying system based on artificial neural network prediction according to claim 5, characterized in that, The particle swarm optimization operation in S3 includes: When constructing the Particle Swarm Optimization (PSO-BP) model, the BP network weights and threshold vectors are used as particle positions, and the training set MSE is used as the fitness function. PSO first performs a global search to obtain optimal parameters, and then uses these parameters as initial values ​​for the BP network. Adam is then used for local training, thereby improving convergence stability and reducing the risk of local optima.

7. The rice noodle drying system based on artificial neural network prediction according to claim 1, characterized in that, The specific operations of S6 include: model training and testing, neural network regression analysis, and relative error analysis; Based on the dynamics experiment of hot air drying of rice noodles, an artificial neural network was introduced to conduct intelligent prediction of the moisture ratio of rice noodles. An improvement model was constructed by building a basic BP neural network and adding three intelligent optimization algorithms, and comparative analysis was carried out.

8. A rice noodle drying method based on artificial neural network prediction, applied to the rice noodle drying system based on artificial neural network prediction as described in any one of claims 1-7, wherein the drying method comprises the following steps: The rice noodles are first aged, then placed in a drying device to remove the dried rice noodles, and the drying characteristics of the rice noodles are tested before being packaged into finished products.

Images