A multi-strategy wolf pack algorithm ELM tobacco dynamic weighing method

Abstract

The application discloses a tobacco dynamic weighing method based on a multi-strategy wolf pack algorithm ELM, and improves ELM neural network prediction of a tobacco dynamic weight through a multi-strategy wolf pack algorithm (MWPA), and the method comprises the following steps: collecting tobacco weight data and influence factor data; pre-processing the data; dividing the pre-processed data into a training set and a test set; applying the multi-strategy WPA algorithm to parameter optimization of an ELM model; training the MWPA-ELM model by using the training set to obtain a tobacco dynamic weighing prediction model. Compared with traditional dynamic weighing technology, the application has the advantages of high precision, fast learning speed, strong generalization ability and the like.

Description

Technical Field

[0001] This invention pertains to dynamic weighing methods for tobacco leaves, specifically a dynamic weighing method for tobacco leaves based on a multi-strategy wolf pack algorithm (ELM). Background Technology

[0002] Tobacco is one of my country's economic crops. Dynamic weighing of tobacco leaves mainly uses weighing belt conveyors. However, tobacco leaves are relatively light, and the dynamic weight of tobacco leaves may differ significantly from the static weight during the weighing process. Furthermore, during dynamic weighing, tobacco leaves are easily affected by temperature, humidity, and wind speed as they move on the conveyor belt. In addition, the dynamic weighing equipment generates mechanical vibrations during operation, which is also one of the reasons for the large error in the dynamic weighing measurement.

[0003] Swarm intelligence optimization algorithms, an emerging computing technology that evolves by simulating the rules of nature, have become a focus of increasing attention from scholars, leading to the proposal of various swarm intelligence algorithms. In 2013, Wu Husheng et al. proposed a new swarm intelligence algorithm—the Wolf Pack Algorithm (WPA)—based on the cooperative hunting characteristics of wolf packs. Compared to other algorithms, it has advantages such as fast convergence speed and strong robustness. However, as a swarm intelligence algorithm based on the cooperative hunting characteristics of wolf packs, the Wolf Pack Algorithm also suffers from disadvantages compared to other swarm intelligence algorithms, such as low search accuracy, susceptibility to local optima, and weak global search capabilities. Summary of the Invention

[0004] To address the problems existing in the dynamic weighing of tobacco leaves in the prior art, this invention provides a dynamic weighing method for tobacco leaves based on the multi-strategy wolf pack algorithm ELM. By using an improved wolf pack algorithm ELM neural network to predict the dynamic weight of tobacco leaves, the problem of the ELM neural network easily getting trapped in local optima is improved, effectively enhancing the accuracy of dynamic weighing of tobacco leaves.

[0005] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution:

[0006] A dynamic weighing method for tobacco leaves based on the multi-strategy wolf pack algorithm ELM includes the following steps:

[0007] S1: Collect data on the length, width, dynamic weight, and static weight of tobacco leaves;

[0008] S2: Fit the dynamic weight of the tobacco leaves to the static weight to obtain a fitting function to compensate for the dynamic weight;

[0009] S3: Normalize the data;

[0010] S4: Initialize the weights and thresholds of the ELM neural network, set the number of input layers, hidden layers, and output layers of the ELM network, and select the activation function;

[0011] S5: Optimize the weights and thresholds of the ELM neural network using a multi-strategy wolf pack algorithm;

[0012] S6: Applying a multi-strategy wolf pack algorithm-based ELM neural network to a dynamic weighing method for tobacco leaves.

[0013] Preferably, in step S1, data on the length, width, dynamic weight, and static weight of tobacco leaves at different placement angles are collected.

[0014] Preferably, in step S2, the dynamic weight and static weight of tobacco leaves under different placement angles are fitted using the least squares method to obtain fitting functions under different placement angles to compensate for the dynamic weight.

[0015] Preferably, in step S3, due to the different dimensions of data weight, the data needs to be normalized. The normalization formula is as follows:

[0016]

[0017] In the formula, and These are the maximum and minimum values ​​in the data, respectively. The value of the i-th sample;

[0018] Preferably, in step S4, the extracted features affecting the dynamic weight of tobacco leaves are used as a training set and fed into an ELM-based network. Since the number of influencing factors is 3, the number of input neurons in the ELM network is set to 3. The output layer represents the static weight of tobacco leaves, so the number of output layer neurons is 1. The optimal number of hidden layer neurons is determined experimentally, and the number of hidden layer nodes is L. The specific method is as follows:

[0019] S4.1: Activation function f x Use the swish function, that is:

[0020] f(x) = x . g(γx), γ = 1

[0021] In the formula, g(x) is the Sigmoid function;

[0022] S4.2: Set up a sample set of n samples. , Where i = 1, 2, 3, …, n, Let represent the input vector of the i-th sample. Let represent the output vector of the i-th sample; the goal of training the ELM network algorithm is to minimize the error between the network output value and the true value, then:

[0023]

[0024] In the formula This represents the value of the actual label. This represents the predicted value of the neural network, where N is the number of training data points.

[0025] S4.3: The ELM prediction model for a single hidden layer node is represented as follows:

[0026]

[0027] Where f(x) is the activation function, Here, represents the connection weights between the input layer and the hidden layer, and l represents the number of neurons in the hidden layer. The threshold of the hidden layer neurons. These are the connection weights between the hidden layer and the output layer;

[0028] The above formula can be expressed in matrix form as follows:

[0029] Hβ = T

[0030] When obtaining β, since the H matrix is ​​not invertible, the least squares method is used to solve for the weights β between the hidden layer units and the output layer units. The calculation expression is as follows:

[0031]

[0032] In the formula, Denotes the generalized inverse of the output matrix H;

[0033] S4.4: After determining the initial β, substitute... This allows us to obtain the predicted value based on the ELM regression model;

[0034] Preferably, the specific method of S5 is as follows:

[0035] S5.1: The Tent chaotic map is used to initialize the algorithm population, and the relevant parameters of the wolf pack algorithm are set. The wolf with the smallest fitness value is selected as the alpha wolf. Random variables are introduced into the original Tent map model. To avoid it falling into small or unstable periodic points, the improved Tent chaotic mapping expression is:

[0036]

[0037] Where a and b are constants, taking values ​​of 2 and 0.5 respectively, E is the number of particles in the sequence, and rand(0,1) is a random number generated within [0, 1].

[0038] After generating the chaotic sequence, it is mapped to the solution space, thereby initializing the wolf population:

[0039] S = lb + (ub — lb) . s

[0040] Where ub and lb are the upper and lower bounds of the solution space, respectively;

[0041] S5.2: In the roaming behavior phase of the wolf pack algorithm, the position update strategy of the individual random movement of the coot optimization algorithm is adopted. The position update of the individual random movement of the coot is as follows:

[0042]

[0043]

[0044]

[0045] Where rand(0, d) takes a random number between [0, d], and d is a constant 1. A random number between [0, 1] This represents the current iteration number. This represents the maximum number of iterations.

[0046] S5.3: To balance global and local search capabilities, and to coordinate global and local search capabilities with the roaming behavior, an inertia weight strategy is used to update the step size of the siege behavior in a non-linear form. The update formula for the inertia weight is:

[0047]

[0048] in, and These are the maximum and minimum values ​​of the inertia weight, respectively.

[0049] The step size update formula is:

[0050]

[0051] in, For the first The step size of the next iteration;

[0052] S5.4: The algorithm terminates when the improved wolf pack algorithm reaches the required accuracy or the maximum number of iterations; the optimal weights and thresholds obtained by ELM;

[0053] Preferably, in step S6, the output matrix H of the hidden layer neurons is calculated, and the output weight β is calculated using the generalized inverse algorithm, thereby realizing the dynamic weighing method of tobacco leaves based on the multi-strategy wolf pack algorithm ELM neural network.

[0054] The beneficial effects of this invention are:

[0055] 1) This invention uses an ELM neural network to predict the dynamic weight of tobacco leaves, which has the advantages of fewer training parameters, faster learning speed, and stronger generalization ability compared with traditional feedforward neural networks.

[0056] 2) This invention uses the wolf pack algorithm to optimize the weights and thresholds of the ELM neural network, thereby improving the accuracy of the model;

[0057] 3) This invention uses a strategy of initializing the population using Tent chaotic mapping, which enhances the diversity of the initial population;

[0058] 4) In the wolf pack roaming phase, this invention introduces the individual random movement mode of the coot algorithm to improve the global search capability and avoid the algorithm getting stuck in local optima;

[0059] 5) This invention introduces an inertial weight strategy during the siege phase to improve the siege step size and balance the algorithm's global search and local exploration capabilities. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 is a system flowchart of the present invention for predicting the dynamic weight of tobacco leaves based on the multi-strategy wolf pack algorithm ELM model;

[0062] Figure 2 is a flowchart of the algorithm for predicting the dynamic weight of tobacco leaves based on the multi-strategy wolf pack algorithm ELM model of the present invention;

[0063] Figure 3 shows the convergence curves of the mean absolute error of the training set based on the WPA-ELM and MWPA-ELM networks of this invention. Detailed Implementation

[0064] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0065] Example 1

[0066] As shown in Figure 1, this invention is a tobacco leaf state weighing method based on the multi-strategy wolf pack algorithm (ELM) applicable to tobacco leaf dynamic weighing systems. The method includes the following steps:

[0067] S1: Set monitoring points:

[0068] The monitoring point is located directly above the weighing module of the weighing belt conveyor, and a camera is installed at the monitoring point. The monitoring point is then put into use.

[0069] S2: Get Data

[0070] Under normal temperature conditions, the same tobacco leaf is placed on the conveyor belt at 15-degree angles (180 degrees in total). When the tobacco leaf passes through the dynamic weighing module, it is photographed to obtain the dynamic weight of the tobacco leaf at different angles and the images of the tobacco leaf passing through the weighing conveyor belt. The length and width data of the tobacco leaf are obtained from the images of the tobacco leaf using the minimum bounding rectangle method.

[0071] S3: Sample data preprocessing:

[0072] Outliers and missing values ​​were removed from the data. When the placement angle was between 45 degrees and 135 degrees, the dynamic weight error of the tobacco leaves was relatively large, and the difference from the actual static weight was more than 2 grams. Therefore, the dynamic weight data within the range of 0 degrees to 45 degrees and 135 degrees to 180 degrees were retained. The dynamic weight and static weight of the tobacco leaves obtained from different placement angles were fitted with the least squares method to obtain the fitting function, thereby correcting the dynamic weight of the tobacco leaves. The static weight data of the tobacco leaves was used as the output data of the ELM algorithm model, and the length, width and dynamic weight of the tobacco leaves were used as the input data of the ELM algorithm model.

[0073] Because the data have different dimensions, normalization is required to preprocess the data. The input data is processed using the min-max normalization method to process the initial data.

[0074] The maximum value selected from the input data samples is denoted as... The minimum value is denoted as The value of the i-th sample is denoted as Then the normalized value of the predicted data is:

[0075]

[0076] S4: Model Training and Prediction:

[0077] The extracted features affecting the dynamic weight of tobacco leaves were used as the training set and fed into the MWPA-ELM network. Since the number of influencing factors was 3, the number of input neurons in the ELM network was set to 3. The output layer was the static weight of tobacco leaves, so the number of output layer neurons was 1. The number of hidden layer neurons was determined experimentally to be optimal, so the number of hidden layer nodes was L.

[0078] Activation function f x Use the swish function, that is:

[0079] f(x) = x . g(γx), γ = 1

[0080] In the formula, g(x) is the Sigmoid function;

[0081] Set n samples to form a sample set , Where i = 1, 2, 3, …, n, Let represent the input vector of the i-th sample. Let represent the output vector of the i-th sample; the goal of training the ELM network algorithm is to minimize the error between the network output value and the true value, then:

[0082]

[0083] In the formula This represents the value of the actual label. This represents the predicted value of the neural network, where N is the number of training data points.

[0084] The ELM prediction model for a single hidden layer node is then expressed as:

[0085]

[0086] Where f(x) is the activation function, Here, represents the connection weights between the input layer and the hidden layer, and l represents the number of neurons in the hidden layer. The threshold of the hidden layer neurons. These are the connection weights between the hidden layer and the output layer;

[0087] The above formula can be expressed in matrix form as follows:

[0088] Hβ = T

[0089] In the formula:

[0090]

[0091]

[0092]

[0093] m represents the number of neurons in the output layer, L represents the number of hidden layer nodes, N represents the number of samples, and H is the hidden layer output matrix of the ELM neural network, where the i-th column of H is the input of the i-th hidden layer node. , …, The output is processed by the activation function. , and H can be calculated;

[0094] When obtaining β, since the H matrix is ​​not invertible, the least squares method is used to solve for the weights β between the hidden layer units and the output layer units. The calculation expression is as follows:

[0095]

[0096] In the formula, Represents the generalized inverse of the output matrix H

[0097] After determining the initial β, substitute it into This allows us to obtain the predicted value based on the ELM regression model;

[0098] The unmodified ELM model does not require updated weights and thresholds during iteration, but this may prevent the model from obtaining the optimal solution, increasing the uncertainty in the fitting process.

[0099] This invention uses the butterfly algorithm to optimize the weights and thresholds of ELM.

[0100] In ELM neural networks , It was obtained through training with an improved wolf pack algorithm.

[0101] The wolf pack algorithm is a swarm intelligence algorithm based on the cooperative hunting characteristics of wolf packs. This algorithm categorizes wolves into three types: alpha wolves, scout wolves, and predators, and abstracts three behaviors: wandering, calling, and attacking, as well as a "survival of the fittest" wolf pack update mechanism. It utilizes Tent chaotic mapping to generate an initial population, increasing population diversity. Furthermore, it improves upon the wandering and attacking behaviors of traditional wolf pack algorithms by introducing individual random movement patterns from the coot algorithm to enhance global search capabilities and avoid getting trapped in local optima. In the attacking behavior, it introduces an inertial weight strategy to improve the attack step size, balancing the algorithm's global search and local exploration capabilities. Finally, the improved wolf pack algorithm is used to train the weights and thresholds of an ELM neural network, minimizing the mean absolute error between the neural network's predicted output and the true value.

[0102] The initial parameter settings for the improved wolf pack algorithm are as follows: number of artificial wolves N=40, number of iterations 500, wolf probing factor α=4, and maximum number of walks. =5, distance decision factor ω =5, step size factor s = 30, update scale factor β =6.

[0103] The improved WPA algorithm includes the following steps:

[0104] Step 1: Initialize the algorithm population using the Tent chaotic map, set the relevant parameters of the wolf pack algorithm, and select the wolf with the smallest fitness value as the alpha wolf;

[0105] Random variables were introduced into the original Tent mapping model. To avoid it falling into small or unstable periodic points, the improved Tent chaotic mapping expression is:

[0106]

[0107] Where a and b are constants, taking values ​​of 2 and 0.5 respectively, E is the number of particles in the sequence, and rand(0,1) is a random number generated within [0, 1].

[0108] After generating the chaotic sequence, it is mapped to the solution space, thereby initializing the wolf population:

[0109] S = lb + (ub — lb) . s

[0110] Where ub and lb are the upper and lower bounds of the solution space, respectively;

[0111] The optimization process uses the mean absolute error as the fitness function to calculate the individual fitness value, and the formula is as follows:

[0112]

[0113] In the formula, This represents the static weight of the tobacco leaves. The dynamic weight of tobacco leaves predicted by ELM.

[0114] Step 2: Select S_num artificial wolves other than the alpha wolf to perform roaming behavior. Update their positions using the individual random movement strategy of the coot algorithm until a scout wolf finds a prey concentration greater than that of the alpha wolf, and can replace the alpha wolf to perform summoning behavior, or until the maximum number of roaming times is reached.

[0115] In the roaming phase of the wolf pack algorithm, the position update strategy of individual random movement, similar to that of the coot optimization algorithm, is adopted. This strategy explores different parts of the search space. If the algorithm gets stuck in a local optimum, this strategy will allow it to escape the local optimum. The position update of the individual coot during random movement is as follows:

[0116]

[0117]

[0118]

[0119] Where rand(0, d) takes a random number between [0, d], and d is a constant 1. A random number between [0, 1] This represents the current iteration number. This represents the maximum number of iterations.

[0120] Step 3: The alpha wolf summons the wolves by howling. The wolves then charge towards the alpha wolf with a longer stride. During the charge, if a wolf finds that the concentration of prey is greater than that of the alpha wolf, it will become the alpha wolf and initiate the summoning action. Alternatively, the wolf may reach a distance from the alpha wolf that is less than the detection distance or reach the maximum number of charge attempts.

[0121] Step 4: Update the positions of the wolves participating in the siege using the siege step size based on inertia weights.

[0122] To balance global and local search capabilities, and to coordinate these capabilities with the roaming behavior, an inertia weight strategy is used to update the step size of the siege behavior in a non-linear manner. The update formula for the inertia weight is as follows:

[0123]

[0124] in, and These are the maximum and minimum values ​​of the inertia weight, respectively.

[0125] The step size update formula is:

[0126]

[0127] in, For the first The step size of the next iteration.

[0128] Step 5: Update the group according to the update mechanism;

[0129] Step 6: The algorithm terminates when the wolf pack algorithm reaches the required accuracy or the maximum number of iterations.

[0130] The optimal weights and thresholds obtained from ELM are used to calculate the output matrix H of the hidden layer neurons. The output weights β are calculated using generalized inverse calculation, thereby realizing the dynamic weight prediction of tobacco leaves.

[0131] In this embodiment, the number of hidden neurons in the ELM layer is 10.

[0132] To verify the effectiveness of the method for predicting the dynamic weight of tobacco leaves based on the multi-strategy wolf pack algorithm ELM model, this method was compared with methods based on ELM neural network and WPA-optimized ELM neural network for predicting the dynamic weight of tobacco leaves. The experimental prediction results on the test set were used to verify its accuracy using three error analysis methods: Mean Square Error (MSE), Mean Absolute Deviation (MAE), and the model's coefficient of determination. Record them in Table 1.

[0133] Table 1. Comparison of Test Set Errors

[0134] model MSE MAE <![CDATA[R 2 ]]> ELM 0.27439 0.47351 0.94033 WPA-ELM 0.19327 0.38016 0.95828 MWPA-ELM 0.07625 0.20100 0.98026

[0135] From Table 1, the MSE, MAE, and The comparison shows that the ELM neural network based on multi-strategy WPA optimization outperforms both the standard ELM neural network and the WPA-optimized ELM neural network. The multi-strategy WPA-optimized ELM neural network overcomes the shortcomings of both the standard ELM neural network and the WPA-optimized ELM neural network, which are prone to getting trapped in local optima, thus improving the prediction accuracy of tobacco leaf dynamic weight.

Claims

1. A dynamic weighing method for tobacco leaves based on a multi-strategy wolf pack algorithm (ELM), characterized in that, Includes the following steps: S1: Collect data on the length, width, dynamic weight, and static weight of tobacco leaves; S2: Fit the dynamic weight of the tobacco leaves to the static weight to obtain a fitting function to compensate for the dynamic weight; S3: Normalize the data; S4: Initialize the weights and thresholds of the ELM neural network, set the number of input layers, hidden layers and output layers of the ELM network, select the activation function, and use the static weight data of tobacco leaves as the output data of the ELM algorithm model, and the length, width and dynamic weight of tobacco leaves as the input data of the ELM algorithm model; S5: Optimize the weights and thresholds of the ELM neural network using a multi-strategy wolf pack algorithm; S6: Applying a multi-strategy wolf pack algorithm-based ELM neural network to a dynamic weighing method for tobacco leaves.

2. The method for dynamic weighing of tobacco leaves using the multi-strategy wolf pack algorithm ELM according to claim 1, characterized in that, In S1, data on the length, width, dynamic weight, and static weight of tobacco leaves at different placement angles are collected.

3. The method for dynamic weighing of tobacco leaves using the multi-strategy wolf pack algorithm ELM according to claim 1, characterized in that, In step S2, the dynamic weight and static weight of tobacco leaves under different placement angles are fitted using the least squares method to obtain the fitting function under different placement angles to compensate for the dynamic weight.

4. The method for dynamic weighing of tobacco leaves using the multi-strategy wolf pack algorithm ELM according to claim 1, characterized in that, In step S3, due to the different dimensions of data weight, data normalization is required. The normalization formula is as follows: In the formula, and These are the maximum and minimum values ​​in the data, respectively. Let be the value of the i-th sample.

5. The method for dynamic weighing of tobacco leaves using the multi-strategy wolf pack algorithm ELM according to claim 1, characterized in that, In step S4, the extracted features affecting the dynamic weight of tobacco leaves are used as a training set and fed into the ELM network. Since the number of influencing factors is 3, the number of input neurons in the ELM network is set to 3. The output layer represents the static weight of tobacco leaves, so the number of output layer neurons is 1. The optimal number of hidden layer neurons is determined experimentally, and the number of hidden layer nodes is L. The specific method is as follows: S4.1: Activation function f x Use the swish function, that is: f(x) = x . g(γx), γ = 1 In the formula, g(x) is the Sigmoid function; S4.2: Set up a sample set of n samples. , Where i = 1, 2, 3, …, n, Let represent the input vector of the i-th sample. Let represent the output vector of the i-th sample; the goal of training the ELM network algorithm is to minimize the error between the network output value and the true value, then: In the formula This represents the value of the actual label. This represents the predicted value of the neural network, where N is the number of training data points. S4.3: The ELM prediction model for a single hidden layer node is represented as follows: Where f(x) is the activation function, Here, represents the connection weights between the input layer and the hidden layer, and l represents the number of neurons in the hidden layer. The threshold of the hidden layer neurons. These are the connection weights between the hidden layer and the output layer; The above formula can be expressed in matrix form as follows: Hβ = T When obtaining β, since the H matrix may be non-invertible, least squares is used. The method for solving the weight β between hidden layer units and output layer units is expressed as follows: In the formula, Denotes the generalized inverse of the output matrix H; S4.4: After determining the initial β, substitute... The predicted value based on the ELM regression model can then be obtained.

6. The method for dynamic weighing of tobacco leaves using the multi-strategy wolf pack algorithm ELM according to claim 1, characterized in that, The specific method of S5 is as follows: S5.1: The Tent chaotic map is used to initialize the algorithm population, and the relevant parameters of the wolf pack algorithm are set. The wolf with the smallest fitness value is selected as the alpha wolf. Random variables are introduced into the original Tent map model. To avoid it falling into small or unstable periodic points, the improved Tent chaotic mapping expression is: Where a and b are constants, taking values ​​of 2 and 0.5 respectively, E is the number of particles in the sequence, and rand(0,1) is a random number generated within [0, 1]. After generating the chaotic sequence, it is mapped to the solution space, thereby initializing the wolf population: S = lb + (ub — lb) . s Where ub and lb are the upper and lower bounds of the solution space, respectively; S5.2: In the roaming phase of the wolf pack algorithm, the coot optimization algorithm is adopted. The position update strategy for random movement of a common coot is as follows: Where rand(0, d) takes a random number between [0, d], and d is a constant 1. A random number between [0, 1] This represents the current iteration number. This represents the maximum number of iterations. S5.3: To balance global and local search capabilities, and to coordinate global and local search capabilities with the roaming behavior, an inertia weight strategy is used to update the step size of the siege behavior in a non-linear form. The update formula for the inertia weight is: in, and These are the maximum and minimum values ​​of the inertia weight, respectively. The step size update formula is: in, For the first The step size of the next iteration; S5.4: The algorithm terminates when the improved wolf pack algorithm reaches the required accuracy or the maximum number of iterations; the optimal weights and thresholds are obtained from ELM.

7. The method for dynamic weighing of tobacco leaves using the multi-strategy wolf pack algorithm ELM according to claim 1, characterized in that, In step S6, the output matrix H of the hidden layer neurons is calculated, and the output weight β is calculated using the generalized inverse algorithm, thereby realizing the dynamic weighing method of tobacco leaves based on the multi-strategy wolf pack algorithm ELM neural network.

Images