A method, medium and system for determining drying parameters in the production process of HIPS boards

Abstract

The present invention provides a method, medium and system for determining drying parameters in the production process of HIPS board, belonging to the technical field of HIPS board production, including: obtaining multiple sets of historical records of HIPS board drying process, including material parameters, process parameters and measured drying quality vectors; establishing a mathematical model, fitting with historical records, and obtaining a drying prediction model; augmenting and expanding historical data, and constructing a training data set; establishing a bidirectional neural network model, including precise and fuzzy branches, training based on the augmented data set, and obtaining a prototype of a drying parameter model; optimizing the model with historical records, and obtaining a final drying parameter model; inputting the current parameters of the HIPS board to be dried and the target drying quality, and outputting the optimal drying process parameters. The present invention solves the technical problem that the existing technology often relies on manual experience for drying parameters, and it is difficult to comprehensively consider the influence of multiple factors, resulting in the determined drying parameters often being inaccurate.

Description

Technical Field

[0001] The invention belongs to the technical field of HIPS board production, and in particular relates to a method, a medium and a system for determining drying parameters in a HIPS board production process. Background Art

[0002] HIPS (high impact polystyrene) is an important plastic material, widely used in home appliance housings, automotive parts, packaging containers and other fields. In the production process of HIPS products, the drying process is a key intermediate link, which directly affects the effect of subsequent molding processing. Reasonable and effective drying can ensure that HIPS resin reaches the ideal moisture content and color and other indicators, thereby ensuring the appearance quality and mechanical properties of the product.

[0003] The existing HIPS board drying mainly adopts the hot air circulation oven. The process personnel set the drying time, temperature and hot air circulation parameters based on production experience and simple rules. However, this empirical method has many shortcomings and it is difficult to fully meet the increasingly high quality requirements.

[0004] 1. Different batches of HIPS sheets have differences in raw material sources, initial processing links, etc., resulting in different physical and chemical properties. Simply relying on empirical parameters is difficult to adapt to different raw materials, thus affecting the drying effect.

[0005] 2. The hot air transfer process is affected by many factors, such as ambient temperature and humidity, material packaging method, plate shape and size, etc. Fixed parameters alone cannot fully consider the impact of actual working conditions.

[0006] 3. Drying quality indicators are diverse, including not only basic indicators such as HIPS moisture content and color difference, but also related qualities such as subsequent processing, plasticization, and molding. Existing empirical parameters lack globality and systematicity.

[0007] 4. HIPS products of different specifications and types have different drying requirements, and different drying target indicators should be set according to the quality requirements of the end products. The existing methods lack flexibility and pertinence.

[0008] 5. The traditional experience setting method cannot monitor and guide the drying process in real time, resulting in certain fluctuations in product quality and inability to ensure batch-to-batch consistency.

[0009] That is to say, the existing technology for the drying parameters of the HIPS board production process often relies on manual experience, and it is difficult to comprehensively consider the influence of multiple factors, resulting in that the determined drying parameters are often not accurate enough. Summary of the invention

[0010] In view of this, the present invention provides a method, medium and system for determining the drying parameters of the HIPS board production process, which can solve the technical problem in the prior art that the drying parameters of the HIPS board production process often rely on manual experience, and it is difficult to comprehensively consider the influence of multiple factors, resulting in the determined drying parameters often being inaccurate.

[0011] The present invention is achieved in that:

[0012] A first aspect of the present invention provides a method for determining drying parameters in a HIPS board production process, comprising the following steps:

[0013] S10, obtaining a plurality of historical records of HIPS board drying processes, wherein the historical records include HIPS board parameters, drying process parameters, and drying quality, wherein the drying quality is represented by a vector, recorded as a drying vector, and the drying quality is obtained by actual detection;

[0014] S20, establishing a drying mathematical model, and fitting the mathematical model with the historical records of the drying process of the plurality of groups of HIPS boards to obtain a drying prediction model;

[0015] S30, for each group of historical records, the HIPS board parameters and drying process parameters are augmented and expanded by an augmentation method, and input into the drying prediction model to obtain a drying quality vector, and form an augmented data set;

[0016] S40, establishing a bidirectional neural network model, including a precise branch and a fuzzy branch, and using the augmented data set to train the bidirectional neural network model to obtain a prototype of a drying parameter model;

[0017] S50, using the historical records as a verification set, optimizing and updating the prototype of the drying parameter model to obtain a drying parameter model;

[0018] S60, obtaining the plate parameters of the HIPS plate to be dried and the required drying quality, inputting them into the drying parameter model, obtaining the drying process parameters as drying parameters and outputting them.

[0019] The precise branch adopts a traditional neural network structure of a fully connected network or a convolutional network to learn the precise mapping relationship between HIPS board parameters and drying quality.

[0020] The fuzzy branch refers to fuzzy logic theory and is used to learn the fuzzy mapping relationship between HIPS board parameters and drying quality through a fuzzification layer, a fuzzy rule layer, a fuzzy reasoning layer and a defuzzification layer.

[0021] The advantage of fuzzy branching is that it can capture the uncertainty and ambiguity of data well, but the disadvantage is that the prediction results may not be precise enough.

[0022] In order to fully utilize the advantages of the two branches, the present invention combines their outputs together. Specifically, the output vectors of the precise branch and the fuzzy branch are concatenated and then passed through a fusion layer (such as a fully connected layer or a convolutional layer) to obtain the final comprehensive drying quality vector prediction value. During the training process, the loss functions of the precise branch and the fuzzy branch are minimized at the same time, so that the two branches promote each other and achieve a collaborative optimization effect.

[0023] The uniqueness of this bidirectional neural network structure is that it can give a precise prediction result and a fuzzy prediction result at the same time, so as to better capture the complexity and uncertainty of the HIPS board drying process. Through the complementarity of the two branches, the model is expected to improve the accuracy and robustness of the drying parameter prediction.

[0024] Among them, when training the bidirectional neural network model, it is necessary to minimize the loss functions of the precise branch and the fuzzy branch at the same time, so that the two branches can promote each other and achieve the effect of coordinated optimization; the output vectors of the precise branch and the fuzzy branch will be spliced, and then passed through a fusion layer to obtain the final comprehensive drying quality vector prediction value.

[0025] Furthermore, the network structure of the precise branch includes: the input layer receives the HIPS board parameters and the drying process parameters, then undergoes nonlinear transformation of one or more hidden layers, and finally reaches the output layer to output a vector with the same dimension as the drying quality vector.

[0026] Furthermore, the function of the fuzzification layer is to convert the input HIPS board parameters and drying process parameters into fuzzy sets, and for each input feature, multiple fuzzy memberships are calculated according to its numerical value.

[0027] Furthermore, the function of the fuzzy rule layer is to construct a number of fuzzy rules based on expert experience or historical data, each rule corresponds to a precondition and a consequent, the precondition is a combination of fuzzy sets of input features, and the consequent is the fuzzy state of the output.

[0028] Furthermore, the function of the fuzzy reasoning layer is to reason about the intermediate fuzzy state according to fuzzy rules to obtain an output fuzzy set, and the reasoning method adopts the basic operations of fuzzy logic.

[0029] Furthermore, the function of the defuzzification layer is to convert the output fuzzy set into an accurate drying quality vector.

[0030] Commonly used defuzzification methods include the centroid method and the maximum mean method.

[0031] Wherein, the step S50 specifically includes the following steps: using the historical records as a verification set, optimizing and updating the prototype of the drying parameter model to obtain a final drying parameter model;

[0032] Optionally, during the optimization and updating process, the HIPS board parameters and drying process parameters in the historical records are input into the model prototype to obtain a predicted drying quality vector, which is then compared with the actual drying quality vector, a verification loss is calculated, and the verification loss is minimized to continuously adjust the model parameters.

[0033] Wherein, the step S60 specifically includes the following steps: obtaining the plate parameters of the HIPS plate to be dried and the required drying quality target, substituting these input data into the drying parameter model, and the model outputs the corresponding drying process parameters as the determined drying parameters.

[0034] Optionally, the number of historical records includes at least several thousand records, and these records cover as many combinations of different HIPS board parameters and process parameters as possible to improve the applicability of the model.

[0035] Optionally, the mathematical model in step S20 includes an empirical model, a mechanism model and a data-driven model. For the mechanism model, it is necessary to conduct in-depth research on the physical and chemical properties of the HIPS board and establish a control equation in combination with the principles of heat and mass transfer.

[0036] Optionally, for the augmentation of numerical features, a method of adding limited Gaussian noise is adopted; for categorical features, a random permutation or replacement based on semantic similarity is adopted, and the augmentation ratio is 2-5 times the amount of data.

[0037] Optionally, the precise branch is a 2-4 layer fully connected network or a small convolutional network, the number of neurons in the hidden layer is 128-512, and the activation function is ReLU or Tanh.

[0038] Optionally, the membership function of the fuzzification layer adopts a triangular function or a Gaussian function, and each input feature corresponds to 3-5 fuzzy sets.

[0039] Optionally, the drying process parameters output by the model need to match the actual production equipment and process conditions, and reasonable parameter range constraints should be set.

[0040] Optionally, an attention mechanism is introduced into the bidirectional neural network model to analyze the contribution of each input feature to the final drying quality vector prediction result, thereby improving the model interpretability.

[0041] Optionally, the fully connected network of the precise branch adopts Dropout technology to prevent overfitting, and the convolutional network of the precise branch adopts a maximum pooling layer after the convolution layer to reduce the feature dimension.

[0042] Optionally, the fuzzification layer adopts a generalized bell-shaped membership function with adjustable parameters.

[0043] Optionally, the rules of the fuzzy rule layer are automatically learned from historical data through optimization algorithms such as genetic algorithms or evolutionary strategies.

[0044] Optionally, the fuzzy reasoning adopts Mamdani reasoning or Sugeno reasoning method.

[0045] Optionally, the defuzzification adopts an improved centroid method and introduces an adaptive adjustment factor to balance the importance of different rules.

[0046] Optionally, the loss function adopts a weighted summation method to assign different weights to the losses of the precise branch and the fuzzy branch.

[0047] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for determining drying parameters in the production process of HIPS board.

[0048] A third aspect of the present invention provides a system for determining drying parameters in a HIPS board production process, which includes the above-mentioned computer-readable storage medium.

[0049] Compared with the prior art, the method, medium and system for determining drying parameters in the HIPS board production process provided by the present invention have the following beneficial effects:

[0050] 1. High prediction accuracy: The model of the present invention combines the advantages of automated modeling of machine learning and the expert knowledge expression ability of fuzzy logic. The precise branch can automatically mine nonlinear laws from historical data; the fuzzy branch introduces expert experience to reasonably describe the process and target requirements. The two branches complement each other, making the overall model have strong fitting and prediction capabilities, which is significantly better than traditional single machine learning models or shallow fuzzy models.

[0051] 2. Strong model interpretability: The rule system of the fuzzy rule branch is constructed by a human-understandable fuzzy logic description language, which clearly explains the expert knowledge and enhances the interpretability of the model. In addition, by adding an attention mechanism, the importance of each input feature on the final prediction result can be analyzed. This is not only conducive to human-computer collaborative interaction, but also provides theoretical support for further optimization of the rule knowledge base.

[0052] 3. Strong generalization ability: The generalization ability of the model of the present invention is mainly reflected in the following aspects:

[0053] (1) The augmentation technology of the expanded data set generates a large number of reliable samples from limited historical data, expands the size of the training set, and improves the robustness of the model.

[0054] (2) Taking into account various conditions such as component characteristics and environmental conditions, it becomes a universal modeling framework.

[0055] (3) The fuzzy rule system can be flexibly edited and automatically learned with the help of optimization algorithms, making it easy to migrate and expand.

[0056] In summary, the solution of the present invention solves the technical problem that the drying parameters of the HIPS board production process in the prior art often rely on manual experience, and it is difficult to comprehensively consider the influence of multiple factors, resulting in that the determined drying parameters are often not accurate enough. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0058] Figure 1 A flow chart of the method provided by the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0060] like Figure 1 FIG. 1 is a flow chart of a method for determining drying parameters in a HIPS board production process provided by the present invention. The method comprises the following steps:

[0061] S10, obtaining a plurality of historical records of HIPS board drying processes, the historical records including HIPS board parameters, drying process parameters and drying quality, the drying quality is represented by a vector, recorded as a drying vector, and the drying quality is obtained by actual detection;

[0062] S20, establishing a drying mathematical model, and fitting the mathematical model with historical records of multiple groups of HIPS board drying processes to obtain a drying prediction model;

[0063] S30, for each group of historical records, the HIPS board parameters and drying process parameters are augmented and expanded by using an augmentation method, and input into a drying prediction model to obtain a drying quality vector, and form an augmented data set;

[0064] S40, establishing a bidirectional neural network model, including a precise branch and a fuzzy branch, and using an augmented data set to train the bidirectional neural network model to obtain a prototype of a drying parameter model;

[0065] S50, using historical records as a verification set, optimizing and updating the prototype of the drying parameter model to obtain a drying parameter model;

[0066] S60, obtaining the plate parameters of the HIPS plate to be dried and the required drying quality, inputting them into the drying parameter model, obtaining the drying process parameters as drying parameters and outputting them.

[0067] The specific implementation methods of the above steps are described in detail below:

[0068] The specific implementation method of step S10 is: first, it is necessary to collect multiple sets of historical records of the HIPS board drying process, which include HIPS board parameters, drying process parameters and drying quality. Among them, HIPS board parameters may include board thickness, density, filler content, etc.; drying process parameters may include temperature, time, humidity, etc.; and drying quality is obtained through actual detection methods, such as detecting indicators such as moisture content and deformation rate of the board. In order to facilitate data processing, the drying quality needs to be represented by a vector, that is, multiple quality indicators are combined into a high-dimensional vector, which is called a drying vector. The purpose of collecting historical records is to provide sufficient data support for subsequent data modeling and model training.

[0069] The specific implementation of step S20 is: based on the historical records collected in step S10, a drying mathematical model needs to be established to mathematically describe the relationship between the HIPS board parameters, drying process parameters and drying quality. This mathematical model may take the form of an empirical model, a mechanism model or a data-driven model. For example, a mechanism model can be constructed based on the physical and chemical properties of the HIPS board and combined with the heat and mass transfer equations to describe the drying process; or a data-driven model can be fitted based on historical data using techniques such as multivariate nonlinear regression. Regardless of the model form used, the historical records collected in step S10 need to be used to perform parameter fitting or model optimization on the mathematical model to obtain the final drying prediction model. The function of this prediction model is to predict the corresponding drying quality vector based on the given HIPS board parameters and drying process parameters.

[0070] The specific implementation method of step S30 is: based on the drying prediction model obtained in step S20, it is necessary to further expand the sample space to provide sufficient training data for the subsequent neural network model training. Specifically, for each set of HIPS board parameters and drying process parameters in the historical records, the augmentation method is used to transform them, such as adding some Gaussian noise to the numerical features, randomly permuting the categorical features, etc. The new parameter combination obtained in this way is then input into the drying prediction model of step S20 to obtain the corresponding drying quality vector. By repeating this process for all historical records, an augmented data set can be finally formed, which contains a large number of HIPS board parameters, drying process parameters and corresponding drying quality vectors. The role of this augmented data set is to provide sufficient training samples, which is helpful for the generalization ability of the subsequent neural network model.

[0071] The specific implementation method of step S40 is: design and train a bidirectional neural network model, which includes two sub-networks, a precise branch and a fuzzy branch. The precise branch adopts a traditional neural network structure, such as a fully connected network or a convolutional network, etc., to learn the precise mapping relationship between the HIPS board parameters and the drying quality. The fuzzy branch draws on fuzzy logic theory and learns the fuzzy mapping relationship between the two through components such as a fuzzification layer, a fuzzy rule layer, a fuzzy reasoning layer, and a defuzzification layer.

[0072] Specifically, the network structure of precise branching can be in the following form: the input layer receives HIPS board parameters and drying process parameters, then undergoes nonlinear transformation in one or more hidden layers, and finally reaches the output layer, outputting a vector with the same dimension as the drying quality vector. Each hidden layer can be a fully connected layer or a convolutional layer, and the activation function can be selected from ReLU, Tanh, etc. During training, it is necessary to minimize the loss function such as the mean square error between the network output and the true drying quality vector.

[0073] The network structure of fuzzy branch includes the following key components:

[0074] 1. Fuzzification layer: Convert the input HIPS board parameters and drying process parameters into fuzzy sets. Specifically, for each input feature, multiple fuzzy memberships are calculated according to its numerical value. These memberships indicate the degree to which the feature belongs to different fuzzy sets. Commonly used membership functions include triangular function, Gaussian function, etc.

[0075] 2. Fuzzy rule layer: Based on expert experience or historical data, construct several fuzzy rules in the form of "IF-THEN" to map the input fuzzy set to the intermediate fuzzy state. Each rule corresponds to an antecedent (IF part) and a consequent (THEN part). The antecedent is a combination of the fuzzy set of input features, and the consequent is the output fuzzy state.

[0076] 3. Fuzzy reasoning layer: According to fuzzy rules, the intermediate fuzzy state is reasoned to obtain the output fuzzy set. The reasoning method can use the basic operations of fuzzy logic, such as union, intersection, complement, etc.

[0077] 4. Defuzzification layer: convert the output fuzzy set into an accurate drying quality vector. Commonly used defuzzification methods include centroid method, mean maximum method, etc.

[0078] During the training process, it is necessary to minimize the loss functions of both the precise branch and the fuzzy branch so that the two branches can promote each other and achieve the effect of collaborative optimization. In addition, the output vectors of the two branches will be concatenated and then passed through a fusion layer (such as a fully connected layer) to obtain the final comprehensive drying quality vector prediction value. The advantage of this bidirectional neural network structure is that it can balance precise prediction and fuzzy prediction, and better capture the complexity and uncertainty of the HIPS board drying process.

[0079] The specific implementation method of step S50 is: using the historical records collected in step S10 as a validation set, optimizing and updating the drying parameter model prototype obtained in step S40, and obtaining the final drying parameter model. Specifically, the HIPS board parameters and drying process parameters in the historical records are input into the model prototype to obtain a predicted drying quality vector, which is then compared with the actual drying quality vector to calculate a validation loss. During the optimization process, the validation loss needs to be minimized, and the model parameters are continuously adjusted until the model reaches the desired performance level on the validation set.

[0080] The specific implementation of step S60 is: obtaining the board parameters of the HIPS board to be dried and the required drying quality target (such as the expected moisture content, deformation rate, etc.), substituting these input data into the drying parameter model obtained in step S50, and the model will output the corresponding drying process parameters, such as temperature, time, humidity, etc. These process parameters are the determined drying parameters, which can guide the actual HIPS board production process.

[0081] In the above steps, some key points and technical details need further explanation:

[0082] 1. Number of historical records in step S10: To ensure the generalization ability of the model, the number of historical records should be sufficient, and it is recommended to contain at least several thousand records. At the same time, these records should cover as many combinations of different HIPS board parameters and process parameters as possible to improve the applicability of the model.

[0083] 2. Mathematical model selection in step S20: Depending on the specific situation, different types of mathematical models can be selected, including empirical models, mechanism models, and data-driven models. For the mechanism model, it is necessary to conduct in-depth research on the physical and chemical properties of HIPS sheets and establish control equations in combination with heat and mass transfer principles; for the data-driven model, multiple nonlinear regression, Gaussian process regression and other technologies can be used. The selection of the model should balance the computational complexity and prediction accuracy.

[0084] 3. Augmentation method in step S30: The purpose of augmentation is to expand the sample space, but at the same time, it is also necessary to ensure that the newly generated samples are still reasonable and in line with the actual situation. For numerical features, limited Gaussian noise can be added; for categorical features, random permutation or replacement based on semantic similarity can be performed. The augmentation ratio can refer to 2-5 times the amount of data.

[0085] 4. Network structure details in step S40:

[0086] The precise branch can use a 2-4 layer fully connected network or a small convolutional network. The number of neurons in the hidden layer can be selected between 128 and 512, and the activation function can use ReLU or Tanh.

[0087] The membership function of the fuzzification layer can be a triangular function or a Gaussian function, and each input feature corresponds to 3-5 fuzzy sets.

[0088] The number of rules in the fuzzy rule layer can be determined based on the number of features and expert experience, usually ranging from dozens to hundreds.

[0089] Fuzzy reasoning can use basic operations such as fuzzy intersection and union, and defuzzification can use the centroid method or the mean maximum method.

[0090] The fusion layer can be a fully connected layer, which concatenates the outputs of the precise branch and the fuzzy branch and inputs them.

[0091] 5. Model optimization in step S50:

[0092] In addition to minimizing the validation set loss, regularization terms (such as L1 / L2 regularization) can also be introduced to prevent overfitting.

[0093] The optimization algorithm can be stochastic gradient descent (SGD), Adam, etc., and the learning rate can be selected between 0.01-0.001.

[0094] The early stopping strategy can monitor the model performance on the validation set and stop training when the performance does not improve within a certain number of training generations.

[0095] 6. Implementation details in step S60:

[0096] When setting drying quality targets, you can refer to industry standards or customer requirements, and usually use indicators such as moisture content and deformation rate as optimization targets.

[0097] The output process parameters need to match the actual production equipment and process conditions, and reasonable parameter range constraints can be set.

[0098] In order to improve interpretability, technologies such as attention mechanism can be introduced into the model to analyze the contribution of each input feature to the final result.

[0099] Through the above detailed step-by-step implementation, the technical solution of the invention can be described in a relatively comprehensive manner. This method combines mathematical modeling, data augmentation and bidirectional neural network, which can effectively capture the complexity and uncertainty of the HIPS board drying process and provide an innovative technical route for optimizing drying parameters. At the same time, corresponding suggestions and reference values ​​are also given for the specific implementation details of each step, providing guidance for engineering practice.

[0100] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the above-mentioned method for determining drying parameters in the production process of HIPS board.

[0101] A third aspect of the present invention provides a system for determining drying parameters in a HIPS board production process, which includes the above-mentioned computer-readable storage medium.

[0102] In order to better understand and implement the present invention, a specific embodiment of the present invention running on a computer is provided below in combination with the formula. In this embodiment, step S10: obtaining the historical records of the drying process of multiple groups of HIPS boards: assuming that Group History, A group record can be represented as:

[0103] ;

[0104] in: is the HIPS sheet parameter vector, with dimension ,like ;

[0105] is the drying process parameter vector, with dimension ,like ;

[0106] is the actual detected drying quality vector, with dimension ,like ;

[0107] Step S20: Establish a drying mathematical model and perform fitting. It can be assumed that the drying quality vector And plate parameters and process parameters Satisfies certain mathematical relations:

[0108] ;

[0109] in is the unknown model function, is the model parameter vector. The goal is to use the given historical data To fit the model parameters , get the optimal model The common fitting methods are:

[0110] 1. For linear models, the least squares method can be used:

[0111] ;

[0112] 2. For nonlinear models, nonlinear least squares method, maximum likelihood estimation and other methods can be used.

[0113] 3. If the model form is unknown, non-parametric regression can be used, such as kernel regression, Gaussian process regression, etc.

[0114] The best model obtained after fitting This is the required drying prediction model.

[0115] Step S30: augment the data set. Group Data , can be disturbed by and Expand and obtain new plate parameters and process parameters :

[0116] ;

[0117] in and is the perturbation function, and is a random noise vector. For continuous features, Gaussian noise can be added:

[0118] ;

[0119] For discrete features, random permutation or similar replacement can be performed.

[0120] The new data Substitute into the prediction model , the corresponding prediction quality vector can be obtained :

[0121] ;

[0122] This constitutes an augmented dataset .

[0123] Step S40: Design a bidirectional neural network model, which consists of a precise branch and a fuzzy branch. The input is and , the output is the predicted quality vector .

[0124] The precise branch uses a traditional fully connected network or convolutional network structure:

[0125] ;

[0126] in is the precise branch network function, is the weight parameter of the precise branching neural network.

[0127] The fuzzy branch draws on the idea of ​​fuzzy logic and consists of a fuzzification layer, a rule layer, a reasoning layer, and a defuzzification layer:

[0128] Fuzzification layer: The input and Mapping to fuzzy sets:

[0129] ;

[0130] in and is a membership function, such as a Gaussian function:

[0131] ;

[0132] in, :Fuzzy layer for plate parameters The fuzzy membership of the components; :The fuzzy layer has a great influence on the process parameters The fuzzy membership of the components; : Plate parameters The membership function of the components; :Process parameters The membership function of the components; : independent variable of membership function; : mean parameter of Gaussian membership function; : standard deviation parameter of Gaussian membership function;

[0133] Rule layer: built based on expert knowledge IF-THEN rules:

[0134] ;

[0135] in, : No. Fuzzy rules; : No. The antecedent fuzzy set of plate parameters in the rules; : No. The antecedent fuzzy set of process parameters in the rules; : No. The consequent fuzzy set of the quality vector in the rules.

[0136] Reasoning layer: Fuzzy logic operation is used to obtain the fuzzy set of quality vectors:

[0137] ;

[0138] in is the fuzzy AND operation, is the fuzzy reasoning operation, : No. Membership function of the plate parameter antecedents in the rules; : No. Membership function of the process parameter antecedents in the rules; : No. The mass vector in the rule The membership function of the component consequences; : Prediction quality vector fuzzy set of components.

[0139] Defuzzification layer: convert fuzzy sets into exact values:

[0140] ;in, represents the defuzzification function, is the output of the fuzzy branch, i.e., the fuzzy prediction vector;

[0141] Commonly used defuzzification methods include the centroid method:

[0142] ;

[0143] Finally, the outputs of the precise branch and the fuzzy branch are combined through a fusion layer:

[0144] ;

[0145] in is the fusion function for exact branching and fuzzy branching, is the corresponding weight parameter.

[0146] The goal of model training is to minimize the precision branch loss and fuzzy branch loss The weighted sum of:

[0147] ;

[0148] in and is the weight coefficient, where : loss function of the precise branch; : The loss function of the fuzzy branch. The loss function can be the mean square error:

[0149] ;

[0150] in, : The number of samples in the training data set; : No. Accurate branch prediction vector for samples; : No. The fuzzy branch prediction vector of samples; : No. The actual drying quality vector of the samples;

[0151] Step S50: Optimize the model using the validation set; As a validation set, define the validation loss:

[0152] ;

[0153] By minimizing the validation loss , you can optimize and adjust the model parameters , and , and obtain the final drying parameter model.

[0154] Step S60: predict new process parameters and give the parameters of the HIPS board to be dried and the target mass vector , substitute it into the optimized model:

[0155] ;

[0156] Solve to obtain the corresponding optimal drying process parameters .

[0157] Optionally, some fuzzy rules of fuzzy branches in the present invention are provided below and expressed using formulas.

[0158] Assume the following fuzzy rules:

[0159] If (the initial moisture content of the board is high and the board temperature is low and the oven temperature is medium and the drying time is high);

[0160] Then (the final moisture content is low, the heat uniformity is medium, and the bending deformation is medium);

[0161] If (the initial moisture content of the board is medium and the board temperature is medium and the oven temperature is high and the drying time is medium);

[0162] Then (the final moisture content is low, the heat uniformity is high, and the bending deformation is medium);

[0163] If (the initial moisture content of the board is low and the board temperature is high and the oven temperature is low and the drying time is low);

[0164] Then (the moisture content at the end is high, the heat uniformity is low, and the bending deformation is high);

[0165] In the inference layer, for each output quality indicator , through fuzzy reasoning operation, we get its fuzzy membership function:

[0166] ;

[0167] in Represents fuzzy reasoning operations, usually using Sherbet multiplication operations:

[0168] ;

[0169] Finally, the fuzzy output is converted into an exact value through a defuzzification layer, such as using the centroid method:

[0170] ;

[0171] Through the above-mentioned fuzzy layer design, the uncertainty and ambiguity in the drying process of HIPS boards can be better captured, providing strong support for downstream decision-making.

[0172] Specifically, the principle of the present invention is:

[0173] The key technical principle of the model of the present invention is to skillfully combine the deep learning model with the fuzzy logic system to achieve the organic integration of data-driven and knowledge-driven. Its basic principles are as follows:

[0174] 1. Architecture design principles

[0175] The entire model consists of two parallel subsystems: precise branch and fuzzy branch, and the final output is obtained through the fusion layer.

[0176] The precise branch is a typical deep neural network that can automatically mine implicit rules and features from a large amount of data and learn parameter models. However, it has black box characteristics and poor explanatory power.

[0177] The fuzzy branch is composed of "fuzzification layer-rule reasoning layer-defuzzification layer". The fuzzification layer converts the input continuous quantity into a fuzzy set; the rule reasoning layer performs fuzzy reasoning on the fuzzy input through artificially designed rules to obtain the fuzzy output of the target quantity; the defuzzification layer converts the fuzzy set into an accurate output. This branch fully incorporates human expert knowledge and has strong interpretability.

[0178] The two branches are coupled through the fusion layer and output the prediction results in a coordinated manner. It not only takes advantage of the strong interpretability of fuzzy systems, but also leverages the powerful data mining capabilities of deep learning. This parallel integration architecture maximizes the model capabilities and the strengths of different branches, achieving the concept of "integration optimization, fusion innovation".

[0179] 2. Model training and optimization principles

[0180] Both precise branches and fuzzy branches need to be optimized through training data to achieve better prediction results. The precise branch uses the back propagation algorithm of traditional deep learning to optimize the weight parameters of the branch network; the fuzzy branch uses evolutionary algorithms and data-driven technology to automatically build a rule base and adjust rule parameters. In addition, the fusion layer is also included in the joint optimization target to ensure the collaborative output quality of the two branches.

[0181] In addition, the present invention also proposes a data augmentation strategy to generate a large number of new reliable samples from limited historical data to expand the size of the training set. This technology not only improves the model's fitting and generalization capabilities, but also avoids the high cost of collecting a large amount of dense data in reality.

[0182] 3. Model interpretation and analysis principles

[0183] Compared with the black box deep learning model, the interpretability of the model of the present invention mainly comes from the introduction of expert knowledge by fuzzy branches. Experts can build a fuzzy rule system to describe the HIPS drying rules based on their understanding of the field and accumulated experience, making the model decision-making process explainable to humans. For example, rules such as "if the moisture content of the board is high, the temperature is moderate, and the time is extended" can intuitively understand the requirements of the drying process.

[0184] In addition, the present invention also designs an explanation analysis module to study the importance of input variables to the prediction results through the attention mechanism. The analysis results can not only identify the importance of features and review the rationality of the model, but also provide data support for in-depth optimization of the rule base.

[0185] 4. Principle of process parameter optimization

[0186] Most existing technologies can only predict the quality results of HIPS board drying, but cannot optimize process parameters to achieve the target requirements. One of the core contributions of this invention is to propose an end-to-end optimization method from input to output, which uses a constrained optimization algorithm to directly solve the optimal process parameters from the model output prediction data.

[0187] Specifically, given the HIPS raw material feature vector and the expected drying quality target, these two are used as model inputs, and the objective function is set as the difference between the output quality vector and the expected quality. By optimizing the input process parameter vector, the model output achieves the quality target while minimizing the error. In this way, the optimal process parameters can be directly optimized at one time without the need for manual analysis and repeated experiments.

[0188] 5. Multi-objective balance principle

[0189] HIPS drying quality detection involves multiple indicators, such as moisture content, color, and post-processing quality. Different products have different requirements for each indicator. Considering only a single indicator can easily lead to the problem of "over-optimization". In order to balance the coordination between various indicators, the present invention adopts a decision layer based on fuzzy reasoning, which integrates the multi-objective prediction results into a unified judgment and optimizes it as the overall objective function. This process itself is a fuzzy decision-making process, which conforms to human decision-making habits and has a strong theoretical explanatory power.

[0190] In order to better understand the present invention, an embodiment of a specific application scenario of the present invention is provided below in combination with a specific application scenario. This embodiment introduces how to optimize the drying parameters for a batch of HIPS boards using the bidirectional neural network model of the present invention, thereby achieving better drying quality.

[0191] Step 1: Get historical data.

[0192] First, the drying records of a batch of HIPS panels in the past two years were extracted from the company's production database. The main characteristics of this batch of HIPS panels include:

[0193] Polymer type: HIPS (high impact polystyrene);

[0194] Production batch: B20200315;

[0195] Average molecular weight: 220000 g / mol;

[0196] Density:1.05g / ;

[0197] Glass transition temperature: 105°C;

[0198] Each set of history records contains the following information:

[0199] Plate parameters (dimensions ): Initial moisture content of the board (%), board temperature (°C), board thickness (mm), board area ( );

[0200] Drying process parameters (dimensions ): Oven temperature (°C), drying time (h), calcining gas flow (m^3 / h), vacuum degree (Pa);

[0201] Drying quality (dimension ): Moisture content (%), heating uniformity, bending deformation;

[0202] Among them, the heating uniformity and bending deformation are both subjective scores of 0-10 points, which are determined by skilled workers based on actual conditions. The database contains a total of 2,000 sets of historical records, which are used as training data after cleaning and standardization.

[0203] Step 2: Establish a drying mathematical model and perform fitting. Since the drying process of HIPS sheets involves complex phenomena such as heat conduction, moisture diffusion, and phase change, it is difficult to establish a strict theoretical model. Therefore, this embodiment adopts a semi-empirical and semi-data-driven hybrid model.

[0204] Firstly, based on the HIPS material properties and heat transfer theory, the following basic form of drying prediction model is proposed:

[0205] ;

[0206] in:

[0207] It is the drying quality index, including the final moisture content, heating uniformity and bending deformation;

[0208] is the initial moisture content of the board;

[0209] is the oven temperature;

[0210] is the calcining gas flow rate;

[0211] is the drying time;

[0212] are the parameters that need to be fitted;

[0213] Using the given 2000 sets of historical records, the parameters of the above model corresponding to the final moisture content, heating uniformity and bending deformation were fitted by nonlinear least squares method, and the following prediction model was obtained:

[0214] Final moisture content model:

[0215] ;

[0216] Thermal uniformity model:

[0217] ;

[0218] Bending deformation model:

[0219] ;

[0220] in , , They represent final moisture content, heating uniformity and bending deformation respectively.

[0221] Step 3: Augmenting the data set. In order to make full use of limited historical data and improve the generalization ability of the model, this embodiment adopts data augmentation to generate 20,000 sets of new data samples based on the existing 2,000 sets of data.

[0222] The specific approach is: for each set of original data , through and Gaussian noise is superimposed on each dimension to generate new samples :

[0223] ;

[0224] ;

[0225] in and are the noise variances of the plate parameters and process parameters respectively, which can be specified according to actual conditions. , .

[0226] For the newly generated , we can use the prediction model established earlier to calculate the corresponding quality indicators as pseudo labels:

[0227] ;

[0228] This constructs the augmented dataset , used for subsequent neural network training.

[0229] Step 4: Design a bidirectional neural network model. According to the present invention, a bidirectional neural network model is designed, which includes a precise branch and a fuzzy branch.

[0230] The precise branch uses a standard fully connected neural network structure, with the input being the plate parameters and process parameters The hidden layer uses the ReLU activation function, and the final output is the predicted quality vector :

[0231] ;

[0232] ;

[0233] ;

[0234] ;

[0235] in and are the weight matrix and bias vector of the i-th layer respectively.

[0236] The fuzzy branch includes the fuzzification layer, the rule layer, the reasoning layer and the defuzzification layer.

[0237] Fuzzification layer: For each input feature and , set 3 Gaussian membership functions:

[0238] ;

[0239] in and are the mean and standard deviation of the Gaussian function, determined by the data.

[0240] 2. Rule layer: contains M fuzzy rules:

[0241] ;

[0242] 3. Reasoning layer: The output fuzzy set is obtained through fuzzy AND operation and fuzzy reasoning operation:

[0243] ;

[0244] 4. Defuzzification layer: Use the centroid method to convert the fuzzy output into an exact value:

[0245] ;

[0246] Output vectors of exact branch and fuzzy branch and Fusion through concatenation and fully connected layers:

[0247] ;

[0248] in and are the weights and biases of the fusion layer.

[0249] The bidirectional neural network is used in the augmented dataset The goal is to minimize the exact branch loss and fuzzy branch loss The weighted sum of:

[0250] ;

[0251] in and is the weight coefficient, and the loss function can be the mean square error:

[0252] ;

[0253] The model performance is evaluated on the validation set, and parameters are tuned and optimized.

[0254] For a new product, given the quality target vector And plate parameters , solve the optimal process parameters that meet the target :

[0255] ;

[0256] This can be done by traversing the possible range of process parameters, inputting them into the model to calculate the loss function, and finally finding the value that minimizes the loss function. value.

[0257] In the model optimization and verification stage, the original 2000 sets of historical records are used to calculate the root mean square error (RMSE) of the bidirectional neural network model on the verification set:

[0258] ;

[0259] in is the validation set size, , , are the predicted final moisture content, heating uniformity and bending deformation of the i-th sample, , , is the corresponding actual value.

[0260] In order to further reduce the validation error, the model was tuned in many aspects, including:

[0261] Network structure adjustment: try to deepen or widen the network, add residual connections, etc.

[0262] Hyperparameter optimization: Optimize learning rate through grid search and random search , regularization coefficient wait;

[0263] Loss function design: According to the characteristics of the task, a composite loss function is constructed to balance various indicators:

[0264] ;

[0265] Component replacement: Explore other fuzzy layer designs, such as triangular membership functions, etc.

[0266] After several rounds of iterative tuning, the final RMSE results on the validation set are shown in Table 1:

[0267] Table 1 RMSE results on the validation set

[0268] index Final moisture content Heating uniformity Bending deformation RMSE 0.087% 0.31 points 0.45 points

[0269] It can be seen that the bidirectional neural network model proposed in the present invention performs well in predicting the final moisture content, while the prediction accuracy of heating uniformity and bending deformation is relatively low due to subjective scoring factors. But overall, the model has reached a practical level and can provide a reference for optimizing drying process parameters.

[0270] Step 6: Apply to new product development

[0271] The company is about to launch a new HIPS sheet product, the main features are as follows:

[0272] Polymer type: HIPS;

[0273] Estimated production batch: B20240522;

[0274] Target molecular weight: 250000 g / mol;

[0275] Density:1.03g / ;

[0276] Glass transition temperature: 110°C;

[0277] According to customer requirements, the product must meet:

[0278] Final moisture content <0.05%;

[0279] Thermal uniformity score>8 points;

[0280] Bending deformation score>7 points;

[0281] For this new product, the above quality indicators can be used as targets Input into the trained bidirectional neural network model and specify the plate parameters , solve the optimal drying process parameters that meet quality requirements :

[0282] After calculation and search, the optimal drying process parameters output are:

[0283] Temperature: 165°C;

[0284] Duration: 11 hours;

[0285] Calcining gas flow: 1400 / h;

[0286] Vacuum degree: 1600Pa;

[0287] According to the model prediction, under this process parameter, the quality index of the new HIPS board is:

[0288] Final moisture content: 0.043%;

[0289] Heating uniformity: 8.1 points;

[0290] Bending deformation: 7.3 points;

[0291] It can be seen that this set of process parameters can basically meet the quality requirements of new products.

[0292] In order to further verify the accuracy of the model prediction, the company organized a small batch trial production of HIPS boards. Drying was carried out according to the above optimal process parameters, and the product quality was tested. The results are shown in Table 2 below:

[0293] Table 2 Product quality test results

[0294] index Final moisture content Heating uniformity Bending deformation Target value <0.05% >8 points >7 points Measured value 0.041% 8.2 points 7.6 points

[0295] It can be seen that the measured values ​​are consistent with the model prediction values ​​within the allowable range, which effectively verifies the effect of the method of the present invention.

[0296] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for determining drying parameters in a HIPS board production process, characterized in that: The following steps are involved: S10, obtaining a plurality of historical records of HIPS board drying processes, wherein the historical records include HIPS board parameters, drying process parameters, and drying quality, wherein the drying quality is represented by a vector, recorded as a drying vector, and the drying quality is obtained by actual detection; S20, establishing a drying mathematical model, and fitting the mathematical model with the historical records of the drying process of the plurality of groups of HIPS boards to obtain a drying prediction model; S30, for each group of historical records, the HIPS board parameters and drying process parameters are augmented and expanded by an augmentation method, and input into the drying prediction model to obtain a drying quality vector, and form an augmented data set; S40, establishing a bidirectional neural network model, including a precise branch and a fuzzy branch, and using the augmented data set to train the bidirectional neural network model to obtain a prototype of a drying parameter model; S50, using the historical records as a verification set, optimizing and updating the prototype of the drying parameter model to obtain a drying parameter model; S60, obtaining the plate parameters of the HIPS plate to be dried and the required drying quality, inputting them into the drying parameter model, obtaining the drying process parameters as drying parameters and outputting them; The precise branch adopts a traditional neural network structure of a fully connected network or a convolutional network to learn the precise mapping relationship between HIPS board parameters and drying quality; The fuzzy branch refers to fuzzy logic theory, and is used to learn the fuzzy mapping relationship between HIPS board parameters and drying quality through fuzzification layer, fuzzy rule layer, fuzzy reasoning layer and anti-fuzzification layer; The network structure of the precise branch includes: the input layer receives the HIPS board parameters and the drying process parameters, then undergoes nonlinear transformation of one or more hidden layers, and finally reaches the output layer to output a vector with the same dimension as the drying quality vector; The function of the fuzzification layer is to convert the input HIPS board parameters and drying process parameters into fuzzy sets, and for each input feature, multiple fuzzy memberships are calculated according to its numerical value; The function of the fuzzy rule layer is to construct a number of fuzzy rules based on expert experience or historical data. Each rule corresponds to a precondition and a consequent. The precondition is a combination of the fuzzy set of input features, and the consequent is the fuzzy state of the output. The function of the fuzzy reasoning layer is to reason about the intermediate fuzzy state according to the fuzzy rules to obtain the output fuzzy set, and the reasoning method adopts the basic operation of fuzzy logic; The function of the defuzzification layer is to convert the output fuzzy set into an accurate drying quality vector; Among them, the fuzzy branch draws on the idea of ​​fuzzy logic and consists of a fuzzification layer, a rule layer, a reasoning layer, and a defuzzification layer: Fuzzification layer: The input and Mapping to fuzzy sets: ； in and As the membership function, the Gaussian function is used: ; in, :Fuzzy layer for plate parameters The fuzzy membership of the components; :The fuzzy layer has a great influence on the process parameters The fuzzy membership of the components; : Plate parameters The membership function of the components; :Process parameters The membership function of the components; : independent variable of membership function; : mean parameter of Gaussian membership function; : standard deviation parameter of Gaussian membership function; Rule layer: built based on expert knowledge IF-THEN rules: ; in, : No. Fuzzy rules; : No. The antecedent fuzzy set of plate parameters in the rules; : No. The antecedent fuzzy set of process parameters in the rules; : No. The fuzzy set of consequents of the quality vectors in the rules; Reasoning layer: Fuzzy logic operation is used to obtain the fuzzy set of quality vectors: ; in is the fuzzy AND operation, is the fuzzy reasoning operation, : No. Membership function of the plate parameter antecedents in the rules; : No. Membership function of the process parameter antecedents in the rules; : No. The mass vector in the rule The membership function of the component consequences; : Prediction quality vector A fuzzy set of components; Defuzzification layer: convert fuzzy sets into exact values: ;in, represents the defuzzification function, is the output of the fuzzy branch, i.e., the fuzzy prediction vector; The defuzzification method uses the centroid method: ; Finally, the outputs of the precise branch and the fuzzy branch are combined through a fusion layer: ; in is the fusion function for exact branching and fuzzy branching, is the corresponding weight parameter.

2. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed, they are used to execute the method for determining drying parameters in the HIPS board production process according to claim 1.

3. A system for determining drying parameters in the production process of HIPS boards, characterized in that: The computer-readable storage medium of claim 2 is included.

Images