Artificial-intelligence-based performance prediction processing method for carbon-fiber carbonization process

Abstract

Disclosed in the present invention is an artificial-intelligence-based performance prediction processing method for a carbon-fiber carbonization process. The method comprises: preprocessing experimental data under test, so as to obtain said experimental data that has been subjected to data cleaning; then, using a sliding window processing method to slide on time series data, extracting data within a window at each position and using the extracted data as an input sample, and determining an input feature and an output variable feature of each input sample, so as to convert the time series data into a plurality of experimental data samples under test in the format of a target model input; performing random data set division on said plurality of experimental data samples, so as to obtain some training sets and some test sets; and constructing a target model, and inputting said experimental data samples into the target model. The target model can implement a relatively accurate mechanical-performance prediction for a carbon-fiber-precursor carbonization process, and the model has an optimal performance in all aspects and has a relatively good generalization capability.

Description

Artificial Intelligence-Based Performance Prediction and Processing Method for Carbon Fiber Carbonization Process Technical Field

[0001] This invention relates to the field of carbon fiber precursor carbonization process production, and in particular to an artificial intelligence-based method for predicting and processing the performance of carbon fiber carbonization process. Background Technology

[0002] PAN-based carbon fiber is a high-performance carbon fiber prepared from polyacrylonitrile (PAN). It is one of the most widely used and highest-performing types of carbon fiber. Its preparation process includes polymerization, spinning, oxidation, and carbonization. Through high-temperature treatment, heteroatoms in the PAN molecular structure are removed, ultimately forming carbon fibers with a highly ordered structure. PAN-based carbon fiber exhibits excellent high-temperature resistance and maintains stable performance under high-temperature environments, thus finding important applications in aerospace, aero-engines, and other fields.

[0003] However, research has found that currently, the core technologies of carbon fiber are still mainly controlled by a few developed countries globally. Domestic carbon fiber production technology urgently needs in-depth research to improve the performance level of my country's carbon fiber products. The carbon fiber production process is complex, with the later carbonization process playing a crucial role and being a critical link in producing high-quality carbon fiber. Therefore, it is essential to intelligently and accurately grasp and understand the performance of the carbonization process to effectively control its production quality.

[0004] Currently, many studies have proposed constructing neural network models such as BP neural networks, support vector regression (SVR), long short-term memory networks (LSTM), and gate recurrent units (GRU) to map carbonization process parameters to the mechanical properties of carbon fibers—especially tensile strength and tensile modulus. However, BP neural networks are prone to overfitting and have low prediction accuracy, while current machine learning models offer significantly better performance, generalization ability, interpretability, and efficiency than the aforementioned models. The carbon fiber carbonization process is a continuous time series, requiring control of parameters such as time and temperature to ensure the quality and performance of the carbon fibers.

[0005] Therefore, machine learning models can serve as an alternative to map process parameters to carbon fiber properties. Developing a method that can quickly and accurately predict carbon fiber properties is both a key focus and a challenge.

[0006] Summary of the Invention:

[0007] The purpose of this invention is to provide an artificial intelligence-based method for predicting the performance of carbon fiber carbonization process. This method can accurately predict the performance of carbon fiber carbonization process, thereby obtaining the most accurate prediction model and improving the performance of carbon fiber precursor.

[0008] To achieve the above-mentioned objectives, embodiments of the present invention provide a method for predicting the performance of carbon fiber carbonization process based on artificial intelligence, comprising the following steps:

[0009] S1. Preprocess the experimental data to be tested to obtain cleaned experimental data; then use the sliding window processing method to slide on the time series data, extract the data in the window at each position as an input sample, and determine the input features and output variable features of each input sample, thereby converting the time series data into multiple experimental data samples in the format of the target model input.

[0010] S2. Randomly divide the dataset of multiple experimental data samples to be tested to obtain a training set and a test set; normalize the training set and the test set to obtain a normalized dataset;

[0011] S3. Construct the target model by inputting the experimental data samples to be detected into the target model; the target model includes an input layer, a CNN layer, two LSTM layers, a GPT-Decoder layer, a fully connected layer, and an output layer;

[0012] S4. The CNN layer of the target model first performs extraction processing and outputs the input sequence consisting of real number vectors; the output of the CNN layer of the target model is passed to the LSTM layer; the LSTM layer is then used to process the temporal relationships of the new time series data and output all hidden states.

[0013] S5. Subsequently, all hidden states of the LSTM layer are used as input to the GPT-Decoder layer, which is used to multiply the attention weights by the value matrix.

[0014] S6. After calculating and weighting the output by multiplying the attention weights by the value matrix, save the target model;

[0015] S7. Set the loss function, optimizer, and hyperparameters of the learning rate for the target model;

[0016] S8. Load the saved target model, use the test set to make predictions, and use preset evaluation indicators to evaluate the target model, finally obtaining a target model with good fitting effect;

[0017] The input features include input time, temperature in different furnace zones, fiber tension, and gas flow rate; the output variable features include the tensile strength and tensile modulus of the carbon fiber.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] This invention proposes an artificial intelligence-based method for predicting the performance of carbon fiber carbonization. It designs a CLGD (CNN+LSTM+GPT-Decoder) neural network model, which combines the advantages of CNN, LSTM and GPT-Decoder, and can better process time series data.

[0020] The CNN layer can extract local features and spatial relationships between different feature values, compensating for the LSTM layer's inability to capture spatial components of data. The LSTM layer can effectively handle long-term dependencies in time series data. The GPT-Decoder layer further enhances the model's ability to process sequence data. Simultaneously, a multi-layer LSTM structure is used to enhance the target model's memory capacity. The self-attention mechanism of the GPT-Decoder improves the model's understanding of the input sequence; data normalization is employed to improve model training efficiency and optimize the training process.

[0021] After the target model converges, multiple evaluation metrics (MSE, MAE, R) are used. 2 A comprehensive evaluation of the model's performance was conducted to ensure its predictive accuracy and generalization ability. This model can more accurately predict the performance of the carbon fiber carbonization process, meeting actual production requirements and significantly improving the accuracy and efficiency of carbon fiber carbonization process performance prediction. It possesses strong practical value and technological innovation.

[0022] In summary, the embodiments of the present invention utilize an innovative CLGD (CNN+LSTM+GPTDecoder) neural network model for transfer learning to construct a more accurate model for predicting the performance of carbon fiber carbonization processes, thus meeting actual production requirements. Attached image description:

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 is a schematic diagram of the overall operation steps of a carbon fiber carbonization process performance prediction and processing method based on artificial intelligence provided in Embodiment 1 of the present invention;

[0025] Figure 2 is a schematic diagram of the deep neural network structure involved in the carbon fiber carbonization process performance prediction processing method based on artificial intelligence provided in Embodiment 1 of the present invention. Detailed implementation method:

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] Figure 1 shows a carbon fiber carbonization process performance prediction method based on artificial intelligence (neural network model) proposed in an embodiment of the present invention, which includes the following steps:

[0028] S1. Preprocess the experimental data to be tested, including data cleaning processes such as removing duplicate values ​​and filling in null values ​​to obtain cleaned experimental data to be tested. Then, use the sliding window processing method to generate input features and output variable features, and convert the time series data into a format suitable for the input of the neural network model of this invention, thereby obtaining multiple experimental data samples to be tested.

[0029] The process of using the sliding window method described above is as follows: define a window of fixed size, slide this window over the time series data, and extract the data within the window at each position as a sample. This operation can obtain multiple experimental data samples to be tested, i.e. discrete samples suitable for input to machine learning models.

[0030] S2. Randomly divide the dataset of multiple experimental data samples to be tested to obtain a training set and a test set; normalize the training set and the test set to obtain a normalized dataset;

[0031] After successful conversion, the dataset is randomly divided. 90% of the dataset is used as the training set to train the target model, and 10% is used as the test set to test the generalization performance of the target model (the target model is specifically the neural network model of this embodiment, i.e., the CLGD model). Then, the maximum and minimum values ​​of the training set are calculated, and the maximum and minimum values ​​of the training set are normalized. The test set is also normalized using the maximum and minimum values ​​of the training set, normalized to between 0 and 1. The normalization calculation is shown in formula (1); X new =(XX) min ) / (X max -X min ); formula (1);

[0032] Among them, X newX represents the normalized data, and X represents the original data. min X represents the minimum value of the original data. max This represents the maximum value of the original data.

[0033] The maximum and minimum values ​​in the training set refer to the maximum and minimum values ​​of each feature in the training dataset. Specifically: maximum value: the largest observed value of a feature in the training set; minimum value: the smallest observed value of a feature in the training set. These values ​​are used for data normalization, scaling all features to the same range (between 0 and 1 in this example), which helps improve the efficiency and performance of model training.

[0034] The data obtained in the above experiments (i.e., the experimental data to be tested) are all time-series data, including input time, temperature in different furnace areas, fiber tension, and gas flow rate. The output is the mechanical properties of carbon fiber (specifically, tensile strength and tensile modulus). The data type is numerical, and traditional machine learning methods are typically used for processing, such as Support Vector Machines (SVR) and Backpropagation (BP) neural networks. Deep neural networks, on the other hand, generally perform better in processing unstructured data such as images, speech, and text. Furthermore, deep neural networks have developed rapidly in recent years, and a large number of mature models are available for construction. Among them, deep neural networks, mainly CNN, LSTM, and Transformer, perform exceptionally well in processing time-series data. The GPT series uses the Transformer architecture, and the powerful fitting ability of the Transformer enables the model to converge.

[0035] The target model in this embodiment of the invention is the CLGD neural network model (i.e., a composite model of CNN+LSTM+GPTDecoder);

[0036] S3. Construct the target model by inputting the experimental data samples to be detected into the target model; the target model includes an input layer, a CNN layer, two LSTM layers, a GPT-Decoder layer, a fully connected layer, and an output layer;

[0037] Figure 2 shows the structure of the proposed CLGD model. The CLGD model (i.e., the target model) consists of an input layer, a CNN layer, two LSTM layers, a GPT-Decoder layer, a fully connected layer, and an output layer.

[0038] First, the format of the input data for the input layer is standardized (i.e., the format of the input data is defined, creating it as time-series data). Each time-series data point contains seq_len+1 data points, where seq_len is the time window size, defining the sequence length of the input data. A "sample" refers to a complete input sequence containing data at seq_len+1 time points. Each sample contains a series of time-series data needed to predict carbon fiber properties; seq_len (sequence length) refers to the number of time steps contained in each input sample. In this target model, each sample contains seq_len+1 data points, where seq_len points are used as input, and the last point is used as the prediction target (i.e., the output is the tensile strength and tensile modulus of the carbon fiber). With a default batch size of 1, a time step of t, and a feature dimension of n (meaning "batch size of 1" means processing one sample at a time, "time step of t" means each sequence contains data from t time points, and "feature dimension of n" means each time point has n features or variables), each input sample can be represented as a real number sequence matrix R. t ×n, where x is denoted as x i For R t The vector representation of the data at the i-th time step in ×n. The processed data is input into the CLGD model, and the CNN layer extracts local features and spatial relationships between different feature values ​​in the data, thereby making up for the shortcomings of the LSTM layer in capturing the spatial components of the data. At the same time, the features it extracts still have temporal sequence (that is, the CNN layer can achieve the technical effect of making up for the shortcomings of the LSTM layer in capturing the spatial components of the data by extracting local features).

[0039] In a CNN layer, the input data for the convolutional layer passes through the input layer and then enters the one-dimensional convolutional layer, where the kernel moves only in the time direction. Let X = {x1, x2, ..., x...} T The data is input to the Conv 1d layer (where Conv 1d refers to a one-dimensional convolutional layer, a core component in the CNN (Convolutional Neural Network) structure). Sample data entering the CNN layer undergoes convolution, pooling, and dimensionality reduction operations sequentially. For sequence data, this model uses one-dimensional convolution, with the convolution kernel performing convolution along a single temporal direction. The CLGD model uses r convolution kernels, each with a size of k. When a convolution kernel convolves a sequence of sample data, it produces a feature map of shape (t-k+1)×1. Then, max pooling is performed on the convolution result. The r feature maps are the features extracted by the CNN layer, which are then reduced to a real vector of length r×(t-k+1) / 2. This vector stores the spatial relationships between different feature values ​​in the sample data.

[0040] S4. The CNN layer of the target model first performs extraction processing and outputs the input sequence consisting of real number vectors; the output of the CNN layer of the target model is passed to the LSTM layer; the LSTM layer is then used to process the temporal relationships of the new time series data and output all hidden states.

[0041] It should be noted that the output of the CNN layer of the target model is passed to the first LSTM layer. The LSTM layer introduces an input gate, a forget gate, and an output gate, as well as cell states and hidden states. Cell states are used to store long-term memory, which helps to alleviate the gradient vanishing problem, while hidden states are used to store short-term memory. The neural network model of this embodiment uses a two-layer LSTM (including a first LSTM layer and a second LSTM layer), where the output of the previous LSTM layer (i.e., the first LSTM layer) is used as the input of the next layer, and is passed layer by layer. The output of the second LSTM layer is sent to the GPT-Decoder layer for further processing. The specific formulas are shown in formulas (2)-(5).

[0042] Forgotten Gate:

[0043] Input Gate:

[0044] Output gate:

[0045] Current hidden layer (cell layer):

[0046] Among them, f t i t ,o t These correspond to the on / off states of the input gate, forget gate, and output gate, respectively. It represents the current cell state, h. t It is the hidden state at the current moment, x t is the value of the input sequence at the current time, W and b are the weight and bias respectively, σ represents the Sigmoid function, and Tanh is the hyperbolic tangent function.

[0047] S5. Subsequently, all hidden states of the LSTM layer are used as input to the GPT-Decoder layer, which is used to multiply the attention weights by the value matrix.

[0048] S6. After calculating and weighting the output by multiplying the attention weights by the value matrix, save the target model;

[0049] S7. Set the loss function, optimizer, and hyperparameters of the learning rate for the target model;

[0050] S8. Load the saved target model, use the test set for prediction, and use preset evaluation indicators to evaluate the target model, finally obtaining a target model with good fitting effect; the input features include input time, temperature of different furnace areas, fiber tension, and gas flow rate; the output variable features include the tensile strength and tensile modulus of carbon fiber.

[0051] During S5 execution, all hidden states of the target model's LSTM layer are used as input to the GPT-Decoder layer, through which the importance of these hidden states is learned.

[0052] It should be noted that the GPT-Decoder layer is a structure in which the Encoder-Decoder Attention layer has been removed, while the Masked Multi-Head Attention layer and the Feed forward layer are retained. The Masked Multi-Head Attention layer is used to learn features in multiple subspaces in a unidirectional manner, and the two linear + ReLU layers in the Feed forward layer are used to transform the dimensions and better fit the features. Each Sub-layer is followed by an Add & Norm layer for residual and normalization. Finally, the attention weights are multiplied by the value matrix to obtain the final output. The self-attention mechanism is shown in Equation (6), and the positionally fully connected feedforward network (i.e., the feedforward network layer) is shown in Equation (7).

[0053] Where Q is the query matrix, used to calculate the relevance with other locations; K is the key matrix, used to represent the importance of different locations; V is the value matrix, used to represent the information of different locations; and the Softmax function normalizes the attention weights to a probability distribution. T This represents the dot product of the query matrix and the key matrix. The value of the dot product is used to scale the gradient to prevent gradient explosion or vanishing.

[0054] The location-fully connected feedforward network consists of two Dense layers and provides a nonlinear transformation through the ReLU activation function.

[0055] It should be noted that the self-attention mechanism is described in Equation (6), where: Q (Query): the lookup matrix, representing the representation of the input at the current position in the model, used to query information at other positions. K (Key): the key matrix, corresponding to the lookup matrix, storing the key information for each position in the sequence. V (Value): the value matrix, storing the value information for each position in the sequence. The Softmax function is used to convert the result of the dot product into a probability distribution, ensuring that the sum of all attention weights is 1.

[0056] Position-wise Feed-Forward Network (FFN): Equation (7) describes the feed-forward network, where: FFN(Z): represents the output of the position-wise feed-forward network, Z represents the input vector to the position-wise feed-forward network; W1 and W2: represent the weight matrices in the position-wise feed-forward network, W1 is the weight of the first layer, and W2 is the weight of the second layer; b1 and b2: represent the bias terms in the position-wise feed-forward network, b1 is the bias of the first layer, and b2 is the bias of the second layer; the position-wise feed-forward network includes two Dense layers and a ReLU activation function.

[0057] A positionally fully connected feedforward network typically consists of two linear layers connected by a non-linear activation function (such as ReLU). The input Z first passes through the first linear transformation and the non-linear activation function, then through the second linear transformation to obtain the final output. This process can be represented as: FFN(Z) = max(0, ZW1+b1)W2+b2; here, max(0, ZW1+b1) represents applying the ReLU activation function to the result of the first linear transformation, then multiplying this result by the weights W2 of the second layer and adding the bias b2. This process helps the model capture more complex features.

[0058] The activation function used is ReLU, as shown in equation (8):

[0059] S7. The loss function used is MSELoss, with mean squared error (MSE) as the evaluation metric. The calculation formula is shown in Equation 9. The loss is optimized using the RAdam optimizer, which dynamically turns the adaptive learning rate on or off based on the latent divergence of the variance. The learning rate is set to lr = 0.008, the batch size to batch_size = 64, and the training epochs to 100.

[0060] Where n represents the total number of samples, Y i Represents the true value. This represents the predicted value.

[0061] S8. Load the trained model to make predictions on the test set to test the model's generalization ability. Use Mean Square Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determination (R^2) to evaluate the model. The calculation formulas are shown in Equation 10-12.

[0062] in Represents the predicted value, y i This represents the true value, and n represents the total number of samples. The smaller the value, the better the model performance.

[0063] in Represents the predicted value, y i denoted by , where n represents the total number of samples. The smaller the MAE, the smaller the model error and the closer the predicted value is to the true value.

[0064] in Represents the predicted value, y i Represents the true value. RSSR represents the mean of the true values, and n represents the total number of samples. SSR calculates the sum of squared differences between the predicted and true mean values, reflecting the dispersion of the model data relative to the true mean. SST calculates the sum of squared differences between the true values ​​and their mean values, reflecting the dispersion of the true values ​​relative to the mean. 2 R represents the ratio of SSR to SST. 2 The larger the value, the better the model fits.

[0065] See Figure 2, which shows a comparison of experimental results for various methods. The BP model is a fully connected network with two hidden layers, each with 64 and 32 neurons respectively. The SVR model uses the default Gaussian radial basis function (rbf) for kernel coefficients, with a regularization parameter of 0.01 and an error tolerance parameter of 0.1.

[0066] The LSTM model uses a two-layer LSTM, where each layer has 8 LSTM units, and the activation function is Tanh.

[0067] The CNN+LSTM model uses a 1DCNN with 32 convolutional kernels, and the LSTM model is configured as above; the CNN+GRU model uses the same CNN configuration as above, and the GRU model uses Tanh as its activation function.

[0068] The CNN+LSTM+Attention model uses the same CNN and LSTM models as above, with the only addition being the attention mechanism.

[0069] The CNN+LSTM+GPTDecoder model uses the same CNN and LSTM models as above, except for the addition of 5 GPTDecoder layers, where the activation function is Softmax and the number of heads for the attention mechanism is set to 4.

[0070] Ultimately, it can be seen that the CNN+LSTM+GPTDecoder model has the best performance in all aspects of the experiment, and the model has stronger generalization ability.

[0071] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the performance of carbon fiber carbonization process based on artificial intelligence, comprising the following steps: S1. Preprocess the experimental data to be tested to obtain cleaned experimental data; Then, the sliding window processing method is used to slide across the time series data, extracting the data within the window at each position as an input sample, and determining the input features and output variable features of each input sample, thereby converting the time series data into multiple experimental data samples to be tested in the format of the target model input; S2. Randomly divide the dataset of multiple experimental data samples to be tested to obtain a training set and a test set; normalize the training set and the test set to obtain a normalized dataset; S3. Construct the target model by inputting the experimental data samples to be detected into the target model; the target model includes an input layer, a CNN layer, two LSTM layers, a GPT-Decoder layer, a fully connected layer, and an output layer; S4. The CNN layer of the target model first performs extraction processing and outputs the input sequence consisting of real number vectors; the output of the CNN layer of the target model is passed to the LSTM layer; the LSTM layer is then used to process the temporal relationships of the new time series data and output all hidden states. S5. Subsequently, all hidden states of the LSTM layer are used as input to the GPT-Decoder layer, which is used to multiply the attention weights by the value matrix. S6. After calculating and weighting the output by multiplying the attention weights by the value matrix, save the target model; S7. Set the loss function, optimizer, and hyperparameters of the learning rate for the target model; S8. Load the saved target model, use the test set to make predictions, and use preset evaluation indicators to evaluate the target model, finally obtaining a target model with good fitting effect; The input features include input time, temperature in different furnace zones, fiber tension, and gas flow rate; The output variable characteristics include the tensile strength and tensile modulus of carbon fibers.

2. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 1, characterized in that, The step of randomly dividing the dataset of multiple experimental data samples to be tested into a training set and a test set includes: using 90% of the total dataset as the training set to train the target model, and using 10% of the total dataset as the test set to test the generalization performance of the model.

3. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 2, characterized in that, The training set and the test set are normalized to obtain a normalized dataset; wherein, the normalization is between 0 and 1, and the normalization calculation is shown in formula (1); X new =(XX) min ) / (X max -X min ); Formula (1); Among them, X new Let X represent the normalized data, and let X represent the original data of the training or test set. min X represents the minimum value of the original data in the training or test set. max This represents the maximum value of the original data in the training or test set.

4. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 3, characterized in that, The step of inputting the experimental data sample to be tested into the target model includes: The data format of the input samples of the input layer is standardized to form new time series data; each input sample contains seq_len+1 data points, where seq_len data points are used as input and the last data point is used as the prediction target, that is, the prediction output is the tensile strength and tensile modulus of carbon fiber; Each input sample x i Represented as a real number sequence matrix R t Let x = ×n, and denote x as ×n. i For R t The vector representation of the data at the i-th time step in ×n.

5. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 4, characterized in that, The CNN layer of the target model first performs extraction processing and outputs an input sequence consisting of real number vectors, including: In the CNN layer of the target model, the data from the input samples passes through the input layer and is then output, before entering a one-dimensional convolutional layer. The kernel moves only in the time direction, specifically including: Let the input sample X = {x1, x2, ..., x...} T The Conv 1d layer, which is input into the CNN layer, uses one-dimensional convolution. The convolution kernel performs convolution along a single temporal direction, and then performs convolution, pooling, and dimensionality reduction operations in sequence. The CNN layer of the target model uses r convolutional kernels, each with a size of k. When a convolutional kernel performs convolution and pooling on a new time series data of an input sample, it produces a feature map with a shape of (t-k+1)×1. Then, through all r convolutional kernels, r feature maps are obtained, which are the features extracted by the CNN layer. These features are then reduced to a real vector of length r×(t-k+1) / 2, forming the input sequence of the CNN layer's output.

6. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 5, characterized in that, The input sequence of the output of the CNN layer of the target model is passed to the LSTM layer, which is a two-layer LSTM. The two-layer LSTM includes a first LSTM layer and a second LSTM layer, wherein the output of the first LSTM layer is used as the input of the next layer, namely the second LSTM layer, and so on. The output of the second LSTM layer is sent to the GPT-Decoder layer for further processing. The LSTM layer includes an input gate, a forget gate, and an output gate, as shown in formulas (2)-(5). Forgotten Gate: Input Gate: Output gate: Current hidden layer: Among them, f t i t ,o t These correspond to the on / off states of the input gate, forget gate, and output gate, respectively. It represents the cell state of the input sequence at the current moment, h. t It is the hidden state of the input sequence at the current time step, x. t yes The input sequence is the value at the current time step, W and b are the weights and biases respectively, σ represents the Sigmoid function, and Tanh is the hyperbolic tangent function.

7. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 5, characterized in that, The GPT-Decoder layer calculates and sums the values ​​by multiplying the attention weights by the value matrix, including: Among them, the self-attention mechanism is shown in Equation (6) and the position fully connected feedforward network is shown in Equation (7); Where Q is the query matrix, K is the key matrix, V is the value matrix, and the Softmax function is used to standardize the attention weights to a probability distribution, QK T This represents the dot product of the query matrix and the key matrix. Used to scale the value of the dot product; FFN(Z)=max(0,ZW1+b1)W2+b2; Formula (7); FFN(Z): Represents the output of the positionally fully connected feedforward network, where Z represents the input vector to the positionally fully connected feedforward network; W1 and W2: Represent the weight matrices in the positionally fully connected feedforward network, where W1 is the weight of the first layer and W2 is the weight of the second layer; b1 and b2: Represent the bias terms in the positionally fully connected feedforward network, where b1 is the bias of the first layer and b2 is the bias of the second layer; The positionally fully connected feedforward network includes two Dense layers and a ReLU activation function.

8. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 7, characterized in that, During the execution of S7, the loss function, optimizer, and hyperparameters of the learning rate of the target model are set; The loss function used is MSELoss, and the mean squared error (MSE) is used as the evaluation index. The calculation formula is shown in formula (9): Where n represents the total number of samples, Y i Represents the actual value. Indicates the predicted value; The optimizer is the RAdam optimizer, which is used to optimize the loss. The learning rate was set to lr = 0.008, the batch size to batch_size = 64, and the training epochs to 100.

9. The method for predicting and processing the performance of carbon fiber carbonization process based on artificial intelligence according to claim 7, characterized in that, During the execution of S8, the saved target model is loaded, the test set is used for prediction, and the target model is evaluated using preset evaluation metrics, ultimately obtaining a target model with good fitting performance. S8. Load the trained target model to predict the test set, and use the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R^2) to evaluate the target model. The calculation formulas are shown in formulas (10)-(12). in Represents the predicted value, y i This represents the true value, and n represents the total number of samples. The smaller the value, the better the model performance. in Represents the predicted value, y i The MAE represents the true value, and n represents the total number of samples. The smaller the MAE, the smaller the model error and the closer the predicted value is to the true value. in Represents the predicted value, y i Represents the actual value. The mean of the true values ​​is represented by n, and the total number of samples is represented by n. SSR calculates the sum of squares of the differences between the mean of the predicted and true values, reflecting the degree of dispersion of the target model data relative to the mean of the true values; SST calculates the sum of squares of the differences between the true value and its mean, reflecting the degree of dispersion of the true value relative to the mean. R 2 R represents the ratio of SSR to SST. 2 The larger the value, the better the fit of the target model.

Images