Cut stem dryer cylinder wall temperature control method based on neural network

Abstract

The invention discloses a method for accurately controlling the temperature of a cylinder wall of a cut stem dryer. The method comprises the following steps of: 1, acquiring parameters of the cut stem dryer at an ith moment and taking the parameters as an ith cylinder temperature node; 2, outputting a production stage category label to which a parameter at the moment belongs according to a cylinder temperature node input at the ith moment by adopting a graph convolutional neural network; 3, entering a proper temperature prediction network channel according to a prediction result of the graph convolutional neural network; and 4, according to the feed-forward and feed-back weight data trained by the multi-layer perceptron, the cylinder wall temperature of the cut stem drying machine required in production is generated in real time through the multi-channel temperature prediction network, so that the cylinder wall temperature control is more accurate and reasonable. According to the method, the problem that the feed-forward and feed-backward adjustment weights of the cylinder wall temperature of the cut stem dryer cannot be adaptively adjusted according to inlet moisture can be solved, so that the stable moisture control capability of the cut stem dryer outlet is ensured, and the process production quality is improved.

Description

Technical Field

[0001] The present invention relates to artificial intelligence algorithm technology, industrial communication field, control theory, signal processing technology, and industrial Internet technology. Background Art

[0002] Drying the cut stems is an important process in the production of tobacco products. It is mainly used to evaporate the moisture in the cut stems so that the cut stems reach a dryness suitable for processing and production. If the temperature of the barrel wall is set unreasonably during the processing, it will not only change the outlet moisture of the cut stems and affect the production quality of the cut stems, but also affect the moisture content of the cut tobacco in the subsequent blending process, directly affecting the compliance rate of the batch of cut tobacco.

[0003] The main way to adjust the drum wall temperature is: before production, a drum wall temperature is preset based on experience, and the moisture at the outlet of the wire dryer is controlled by the preset drum wall temperature at the material head. When the moisture at the outlet of the wire dryer reaches the preset value, the drum wall temperature is controlled by the PID model to achieve the purpose of adjusting the moisture at the outlet of the wire dryer.

[0004] However, there are two problems in actual production: 1. The change of drum wall temperature depends on manual adjustment. When the inlet moisture content fluctuates frequently between batches, the post-feedback adjustment of the PID model cannot quickly control the outlet moisture content of the dryer.

[0005] 2. Manual adjustment relies on prior experience, and each adjustment needs to be tested and observed. Improper operation can easily cause large fluctuations in the moisture content of the dried shredded stems, which is not conducive to ensuring the qualified rate of material quality. Summary of the invention

[0006] The present invention aims to solve the deficiencies of the above-mentioned prior art and proposes a method for controlling the barrel wall temperature of a shredded stem drying machine based on a neural network, in order to solve the problem of low outlet moisture control accuracy of the shredded stem drying machine caused by large inlet moisture fluctuations, thereby ensuring the stable control capability of the outlet moisture of the shredded stem drying machine and improving the process production quality.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: The present invention is characterized in that a method for controlling the barrel wall temperature of a shredded stem drying machine based on a neural network comprises the following steps:

[0008] Step 1: When the batch of materials passes through the stem shred flash steamer, the parameters sent from the lower computer PLC and MES system are obtained. Standard water removal at each moment (i), rated process flow (i) Rated moisture increase after tunnel heating and humidification (i), rated outlet moisture , rated cylinder wall temperature (i), actual inlet moisture (i), actual export moisture (i) and constitute the Parameters of the drying machine at this moment ,right After normalization, Barrel temperature nodes , thus obtaining the barrel temperature node set ,in, represents the number of tube temperature nodes; Barrel temperature nodes The actual production stage category label is , { }, where n represents the total number of stages in the production process;

[0009] calculate The Euclidean distance between each tube temperature node in is used as the similarity between the edges of the corresponding nodes; Barrel temperature nodes With Barrel temperature nodes The similarity between the edges is recorded as ,like , then it means Barrel temperature nodes With Barrel temperature nodes There is an edge between , and let the adjacency matrix Middle Line Column Elements =1, otherwise, it means Barrel temperature nodes With Barrel temperature nodes There is no edge between them, and =0, where Indicates the set threshold;

[0010] Step 2: Build a graph convolutional neural network for the barrel temperature node set Processing is performed to obtain the production status prediction category label to which each barrel temperature node belongs , thus establishing the cross entropy loss function ;

[0011] Step 2.1: Set the total number of layers of the graph convolutional neural network to L and define the current number of layers of the graph convolutional neural network to , and initialize =0; define The feature matrix of the layer graph convolutional neural network is , and initialize = , T represents transpose;

[0012] Step 2.2: Use formula (1) to calculate the Feature matrix output by layer graph convolutional neural network , thus Feature matrix output by layer graph convolutional neural network :

[0013] (1)

[0014] In formula (1), For the The weight matrix to be trained in the layer graph convolutional neural network; is the activation function;

[0015] Step 2.3: After being processed in the fully connected layer, the Softmax layer outputs the classified production status prediction category label vector ={ },in, Indicates Barrel temperature nodes The production status prediction category label;

[0016] Step 2.4: Use formula (2) to construct the cross entropy loss function of the graph convolutional neural network ;

[0017] (2)

[0018] Step 3: Calculate the temperature of the wall of the drying tube of the shredded stems according to the principle of the shredded stem drying machine;

[0019] Step 3.1: Use formula (3) to calculate the The actual drum temperature of the dried shredded silk after feedforward adjustment at each moment :

[0020] (3)

[0021] In formula (3), is the initial weight of feedforward regulation, For the The actual water removal of the leaf silk at a certain moment is:

[0022] (4)

[0023] Step 3.2: Use formula (5) to calculate the The actual drum temperature of the dried silk after feedback adjustment at the moment :

[0024] (5)

[0025] In formula (5), is the proportional gain, is the integral gain, is the differential gain, For the The control error at each moment is ; For the The error differential at each moment is is the initial weight of the feedback adjustment;

[0026] Step 4: Build a multi-channel temperature prediction network , and based on the predicted label output by the graph convolutional network ,right and Processing, output the predicted temperature of the cylinder wall , thus establishing the loss function ;

[0027] Step 4.1: Build a multi-channel temperature prediction network ,in, represents the cylinder wall temperature prediction network of the s-th production stage, and A multi-layer perceptron consisting of an input layer, a hidden layer, and an output layer. ;

[0028] Set the total number of layers of each multilayer perceptron to F and define the current number of layers as , and initialize =1;

[0029] Step 4.2: When When Barrel temperature nodes The production stage category of the forecast is , then (i) and (i) As the Moment Input data of the layer , and enter The fth layer in is processed, and then the fth layer is obtained by using formula (7) Moment The feature matrix of the layer output :

[0030] (6)

[0031] In formula (6), Indicates Moment -1 layer output feature matrix, when =1, let = ; For the Moment -1 layer inter-layer dimension transformation matrix, For the Moment -1 bias vector for layer, is the activation function;

[0032] Step 4.3: Moment The feature matrix of the layer output After inputting into a fully connected layer for processing, we get Predicted temperature of cylinder wall at time ;

[0033] Step 4.4: Use formula (7) to establish the loss function :

[0034] (7)

[0035] In formula (7), For the The actual temperature of the cylinder wall during normal production at a certain moment;

[0036] Step 5: Use formula (8) to construct the total loss function :

[0037] (8)

[0038] Step 6: Jointly train the graph convolutional neural network and the multi-channel temperature prediction network through back propagation and gradient descent algorithms to calculate To update the network parameters until the maximum number of iterations is reached or The method converges until the wall temperature is self-adapted, thus obtaining a wall temperature self-adaptive generation model, which is used to predict the wall temperature of the shredded straw drying machine parameters input in actual production.

[0039] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the method for controlling the barrel wall temperature of the shredded stem drying machine, and the processor is configured to execute the program stored in the memory.

[0040] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the method for controlling the barrel wall temperature of the shredded stem drying machine when the computer program is run by a processor.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. The present invention adopts a neural network algorithm learning method, which can adaptively control the barrel wall temperature according to different environmental conditions and process parameters to cope with different production conditions and tobacco characteristics, solves the problem of inefficiency and lack of flexibility in manual parameter adjustment, and ensures the drying quality of the cut stems and the consistency of the product.

[0043] 2. The present invention adopts graph convolutional neural network to predict labels in the production stage, which effectively solves the problem of poor model fitting effect in different production stages and reduces the difficulty of prediction of subsequent modules;

[0044] 4. The present invention adopts a multi-channel cylinder wall temperature prediction network, which can automatically select the learning channel according to the production stage label predicted by the graph convolutional neural network, solving the problem that the single model predicts the cylinder wall temperature with poor fitting effect due to the large differences in the original data of different production states, and improves the accuracy of cylinder wall temperature prediction.

[0045] 5. The present invention adopts a multi-layer perceptron to learn the optimal weights of feedforward regulation and feedback regulation in the normal production process, which solves the problems of uncontrollable temperature of the barrel wall of the stalk drying machine due to unreasonable weight setting and overly complicated manual adjustment, and reduces the number of human interventions in production. DETAILED DESCRIPTION

[0046] In this embodiment, a method for controlling the barrel wall temperature of a stalk drying machine based on a neural network is applied to a stalk drying production line composed of an inlet moisture detector and a drum stalk drying machine. To ensure the accuracy of the stalk drying machine barrel wall temperature control, it is necessary to set reasonable weight values ​​for the feedforward and feedback adjustment methods of the stalk drying machine under different production conditions to generate a suitable barrel temperature. Therefore, a data reading device is set on one side of the inlet moisture meter, and a set of data acquisition and output control systems are constructed using Siemens 1500 series PLC and Python language to realize data reading. The input data is simulated in real time through a pre-trained graph convolutional neural network and a multi-channel temperature prediction network joint model to generate the barrel temperature parameters that need to be adjusted, and finally the barrel wall temperature of the stalk drying machine is changed in real time with the predicted output. Specifically, the method for controlling the barrel wall temperature of the stalk drying machine includes the following steps:

[0047] Step 1: When the batch of materials passes through the stem shred flash steamer, the parameters sent from the lower computer PLC and MES system are obtained. Standard water removal at each moment (i), rated process flow (i) Rated moisture increase after tunnel heating and humidification (i), rated outlet moisture , rated cylinder wall temperature (i), actual inlet moisture (i), actual export moisture (i) and constitute the Parameters of the drying machine at this moment ,right After normalization, Barrel temperature nodes , thus obtaining the barrel temperature node set ,in, represents the number of tube temperature nodes; Barrel temperature nodes The label is , { }, indicating the Barrel temperature nodes The actual production stage to which it belongs, where n represents the maximum number of stages in the production process; in the specific example, n=2, namely the pre-production stage and the formal production stage.

[0048] Calculated by formula (1) The Euclidean distance between each tube temperature node in is used as the similarity between the edges of the corresponding nodes; Barrel temperature nodes With Barrel temperature nodes The similarity between the edges is recorded as ,like , then it means Barrel temperature nodes With Barrel temperature nodes There is an edge between , and let the adjacency matrix Middle Line Column Elements =1, otherwise, it means Barrel temperature nodes With Barrel temperature nodes There is no edge between them, and =0, where represents the set threshold, as shown in formula (2);

[0049] (1)

[0050] (2)

[0051] Step 2: Build a graph convolutional neural network for the barrel temperature node set Processing is performed to obtain the predicted brand and production status prediction category to which each barrel temperature node belongs;

[0052] Step 2.1: Set the total number of layers of the graph convolutional neural network to L and define the current number of layers of the graph convolutional neural network to , and initialize =0; define The feature matrix of the layer graph convolutional neural network is , and initialize = , T represents transpose; in the specific example, L=2. This is because after using multi-layer GCN, the distinguishability of nodes becomes worse and worse, and their representation vectors tend to be consistent, resulting in a series of problems such as gradient disappearance, over-smoothing, and over-fitting, which makes it difficult to continue related learning tasks.

[0053] Step 2.2: Use formula (3) to calculate the Feature matrix output by layer graph convolutional neural network , thus Feature matrix output by layer graph convolutional neural network :

[0054] (3)

[0055] In formula (3), For the The weight matrix to be trained in the layer graph convolutional neural network; is the activation function; in the specific example, the activation function is Sigmoid. This is because the Sigmoid function converts the output value into probability by mapping the input to (0,1), and performs better in the binary classification problem.

[0056] Step 2.3: After being processed in the fully connected layer, the Softmax layer outputs the classified state vector { },in, Indicates Barrel temperature nodes The predicted label; each component in the state vector represents the predicted label probability of the predicted brand and production status of the corresponding node; in the specific example, the classification function in the Softmax classification layer is shown in formula (4);

[0057] S=1∕(1+e -d )(4)

[0058] In formula (4), d is the input of the classification function, S is the output of the classification function, and through formula (4), the node features are mapped to the interval [0, 1].

[0059] Step 2.4: Use formula (5) to construct the cross entropy loss function of the graph convolutional neural network ;

[0060] (5)

[0061] In formula (5), is the predicted label of the i-th barrel temperature node.

[0062] Step 3: Calculate the temperature of the wall of the drying tube of the shredded stems according to the principle of the shredded stem drying machine;

[0063] Step 3.1: Use equation (6) to calculate The actual drum temperature of the dried shredded silk after feedforward adjustment at each moment :

[0064] (6)

[0065] In formula (6), is the initial weight of feedforward regulation, For the The actual water removal of the leaf silk at a certain moment is:

[0066] (7)

[0067] In the specific example, The calculation method of Similar, except for rated inlet moisture With rated outlet is the set value.

[0068] Step 3.2: Use formula (8) to calculate the The actual drum temperature of the dried silk after feedback adjustment at the moment :

[0069] (8)

[0070] In formula (8), is the proportional gain, is the integral gain, is the differential gain, For the The control error at each moment is ; For the The error differential at each moment is is the initial weight of the feedback adjustment.

[0071] Step 4: Build multiple neural network channels based on the predicted labels output by the graph convolutional network , that is, predict the cylinder wall temperature;

[0072] Step 4.1: Due to the large differences in the original data of different production states, the fitting effect of using a single model to predict the barrel wall temperature is poor. , and obtain the unique cylinder wall temperature prediction model , thus forming a multi-neural network ensemble , where n represents the maximum number of production states, .gather The models in are all multi-layer perceptrons consisting of an input layer, a hidden layer, and an output layer. In the specific example, there is one hidden layer, that is, the perceptron model has three layers. The actual cylinder wall temperature during normal production is labeled as , .

[0073] Prediction labels based on graph convolutional network output ,when When Barrel temperature nodes The forecast production stage is , choose to enter the perceptron model Calculate and output.

[0074] Step 4.2: Set the total number of layers of the multilayer perceptron model to F and define the current number of layers as , and initialize =0;

[0075] Will (i) or (i) As the Input data at time , where the input data For the The feedforward cylinder wall temperature calculated according to the drying principle of stalks at a certain moment (i) and the temperature of the backfeed tube wall (i), and , using formula (9) to output the model In the Moment Layer feature matrix :

[0076] (9)

[0077] In formula (9), For the Moment Model No. The inter-layer dimension transformation matrix obtained by training, For the Moment Model No. The bias vector of the layer, is the activation function. In the specific example, It is the LeakyReLU function. This is because LeakyReLU avoids the problem of neuron death in ReLU and ensures that the output predicted cylinder wall temperature is not always 0.

[0078] Step 4.3: After the output of the last fully connected layer, the predicted cylinder wall temperature vector { },in, Indicates The final predicted cylinder wall temperature is obtained at the moment.

[0079] Step 4.4: Use formula (10) to establish the loss function :

[0080] (10)

[0081] Step 5: Use formula (11) to construct the total loss function .

[0082] (11)

[0083] Step 6: Jointly train the graph convolutional neural network layer and the corresponding channel perceptron model through back propagation and gradient descent algorithms to calculate the loss function and The sum is used to update the graph network model weight matrix and perceptron model parameters and , until the maximum number of iterations or the convergence position of the loss function value is reached, thereby obtaining the barrel wall temperature adaptive generation model, which is used to accurately predict the barrel wall temperature in actual production.

[0084] Step 5: Model application:

[0085] Apply the trained model to the actual equipment use, and output the barrel temperature value that needs to be adjusted according to the actual value of the dryer input parameters in the batch production stage. When the outlet moisture reaches a certain value, the cylinder wall temperature run.

[0086] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0087] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A method for controlling the wall temperature of a shredded stem drying machine based on a neural network, characterized in that: The steps include: Step 1: When the batch of materials passes through the stem shred flash steamer, the parameters sent from the lower computer PLC and MES system are obtained. Standard water removal at each moment (i), rated process flow (i) Rated moisture increase after tunnel heating and humidification (i), rated outlet moisture , rated cylinder wall temperature (i), actual inlet moisture (i), actual export moisture (i) and constitute the Parameters of the drying machine at this moment ,right After normalization, Barrel temperature nodes , thus obtaining the barrel temperature node set ,in, represents the number of tube temperature nodes; Barrel temperature nodes The actual production stage category label is , { }, where n represents the total number of stages in the production process; calculate The Euclidean distance between each tube temperature node in is used as the similarity between the edges of the corresponding nodes; Barrel temperature nodes With Barrel temperature nodes The similarity between the edges is recorded as ,like , then it means Barrel temperature nodes With Barrel temperature nodes There is an edge between , and let the adjacency matrix Middle Line Column Elements =1, otherwise, it means Barrel temperature nodes With Barrel temperature nodes There is no edge between them, and =0, where Indicates the set threshold; Step 2: Build a graph convolutional neural network for the barrel temperature node set Processing is performed to obtain the production status prediction category label to which each barrel temperature node belongs , thus establishing the cross entropy loss function ; Step 2.1: Set the total number of layers of the graph convolutional neural network to L and define the current number of layers of the graph convolutional neural network to , and initialize =0; define The feature matrix of the layer graph convolutional neural network is , and initialize = , T represents transpose; Step 2.2: Use formula (1) to calculate the Feature matrix output by layer graph convolutional neural network , thus Feature matrix output by layer graph convolutional neural network : (1) In formula (1), For the The weight matrix to be trained in the layer graph convolutional neural network; is the activation function; Step 2.3: After being processed in the fully connected layer, the Softmax layer outputs the classified production status prediction category label vector ={ },in, Indicates Barrel temperature nodes The production status prediction category label; Step 2.4: Use formula (2) to construct the cross entropy loss function of the graph convolutional neural network ; (2) Step 3: Calculate the temperature of the wall of the drying tube of the shredded stems according to the principle of the shredded stem drying machine; Step 3.1: Use formula (3) to calculate the The actual drum temperature of the dried silk after feedforward adjustment at each moment : (3) In formula (3), is the initial weight of feedforward regulation, For the The actual water removal of the leaf silk at a certain moment is: (4) Step 3.2: Use formula (5) to calculate the The actual drum temperature of the dried silk after feedback adjustment at each moment : (5) In formula (5), is the proportional gain, is the integral gain, is the differential gain, For the The control error at each moment is ; For the The error differential at each moment is is the initial weight of the feedback adjustment; Step 4: Build a multi-channel temperature prediction network , and based on the predicted label output by the graph convolutional network ,right and Processing, output the predicted temperature of the cylinder wall , thus establishing the loss function ; Step 4.1: Build a multi-channel temperature prediction network ,in, represents the cylinder wall temperature prediction network of the s-th production stage, and A multi-layer perceptron consisting of an input layer, a hidden layer, and an output layer. ; Set the total number of layers of each multilayer perceptron to F and define the current number of layers as , and initialize =1; Step 4.2: When When Barrel temperature nodes The production stage category of the forecast is , then (i) and (i) As the Moment Input data of the layer , and enter The fth layer in is processed, and then the fth layer is obtained by using formula (7) Moment The feature matrix of the layer output : (6) In formula (6), Indicates Moment -1 layer output feature matrix, when =1, let = ; For the Moment -1 layer inter-layer dimension transformation matrix, For the Moment -1 bias vector for layer, is the activation function; Step 4.3: Moment The feature matrix of the layer output After inputting into a fully connected layer for processing, we get Predicted temperature of cylinder wall at time ; Step 4.4: Use formula (7) to establish the loss function : (7) In formula (7), For the The actual temperature of the cylinder wall during normal production at a certain moment; Step 5: Use formula (8) to construct the total loss function : (8) Step 6: Jointly train the graph convolutional neural network and the multi-channel temperature prediction network through back propagation and gradient descent algorithms to calculate To update the network parameters until the maximum number of iterations is reached or The method converges until the wall temperature is self-adapted, thus obtaining a wall temperature self-adaptive generation model, which is used to predict the wall temperature of the shredded straw drying machine parameters input in actual production.

2. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the method for controlling the barrel wall temperature of the shredded stem drying machine according to claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the method for controlling the barrel wall temperature of the shredded stem drying machine according to claim 1 are executed.

Images