Intelligent development method of preparation process of polyacrylonitrile-based carbon fiber

Abstract

The invention relates to an intelligent development method for a preparation process of polyacrylonitrile-based carbon fibers. The intelligent development method comprises the following steps: constructing a prediction model comprising a first total model, a first type of sub-models, a second total model and a second type of sub-models; the first total model is a machine learning algorithm model which takes whole-line process parameters as input and takes performance data of the polyacrylonitrile-based carbon fibers as output. The first type of sub-model is a machine learning algorithm model which takes local process parameters as input and takes fiber performance data and / or structure data after local process processing as output. The second total model is a machine learning algorithm model which takes the performance data of the polyacrylonitrile-based carbon fiber as input and takes the whole-line process parameters as output. And the second type of sub-model is a machine learning algorithm model which takes the performance data of the polyacrylonitrile-based carbon fiber as input and takes local process parameters as output. By means of the prediction model, intelligent development of the carbon fiber preparation technology is achieved, and the trial and error cost in the carbon fiber development and debugging process is remarkably reduced.

Description

Technical Field

[0001] The invention relates to the technical field of carbon fibers, and in particular to an intelligent development method for a preparation process of polyacrylonitrile-based carbon fibers. Background Art

[0002] Polyacrylonitrile (PAN)-based carbon fiber is a high-tech integrated, high-value-added product with properties such as light weight, high strength, high temperature resistance, and corrosion resistance. It has been widely used in aerospace, aviation, and industrial products. High-performance PAN-based carbon fiber has irreplaceable and wide applications in special fields.

[0003] High-performance polyacrylonitrile-based carbon fiber has many performance indicators and high requirements. Its preparation process is particularly long (there are many coupling and contradictions in the design of the previous and subsequent processes), many process control points (≥3000 control points), multiple technical disciplines are interwoven, and the temperature span of the entire process is extremely large (from room temperature to 3000°C). Therefore, the research and development of high-performance carbon fiber is a complex system engineering problem.

[0004] At present, an important direction of the research and development of a new generation of carbon fiber is the research and development of high-model carbon fiber. If traditional experimental research continues to be used, it will inevitably face many "trial and error" choices, resulting in huge investment in time, energy and funds.

[0005] In recent years, machine learning has made breakthroughs in the fields of modeling / performance prediction and optimization in materials research and development, showing more outstanding advantages than traditional research methods. Figure 1 A new paradigm for the development of carbon fiber materials (see Figure 1 Figure a in Figure 2, characterized by data-driven artificial intelligence) and traditional research paradigms (see Figure 1 Comparison with Figure b).

[0006] At present, machine learning technology is infiltrating the field of carbon fiber production technology, and has formed a cross-integration development trend. Existing researchers have applied specific machine learning algorithms to achieve research results with practical value and guiding significance for carbon fiber scientific research and production in the fields of carbon fiber production process control, carbon fiber local process optimization, and carbon fiber production process energy consumption optimization.

[0007] In summary, there is an urgent need to introduce machine learning algorithms into the research process of the entire process and local processes of polyacrylonitrile-based carbon fiber preparation, and to innovate an intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber, so as to significantly reduce the trial and error costs (time, energy consumption, funds, etc.) of the complex system engineering debugging process. Summary of the invention

[0008] In view of this, the present invention provides an intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber, the main purpose of which is to reduce the trial and error cost of the development of the preparation process of polyacrylonitrile-based carbon fiber.

[0009] In order to achieve the above object, the present invention mainly provides the following technical solutions:

[0010] On the one hand, an embodiment of the present invention provides an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, wherein the intelligent development method comprises the following steps:

[0011] Constructing a prediction model step: based on the process data set of polyacrylonitrile-based carbon fiber, constructing a first general model, a first type of sub-model, a second general model and a second type of sub-model; wherein, the first general model is a machine learning algorithm model that takes the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and the performance data of polyacrylonitrile-based carbon fiber as output; the first type of sub-model is a machine learning algorithm model that takes the local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and the fiber performance data and / or structure data after local process treatment as output; the second general model is a machine learning algorithm model that takes the performance data of polyacrylonitrile-based carbon fiber as input and the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output; the second type of sub-model is a machine learning algorithm model that takes the performance data of polyacrylonitrile-based carbon fiber as input and the local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output;

[0012] The step of setting the criterion conditions: setting the target performance index of the required polyacrylonitrile-based carbon fiber; setting the parameter range of each process point in the full-line process parameters; setting the output criterion conditions of the first general model, the second general model, the first sub-model, and the second sub-model;

[0013] Process search step: according to the set target performance index of polyacrylonitrile-based carbon fiber, first run the second general model to generate a set of full-line process parameters; if the generated full-line process parameters meet the output criterion conditions of the second general model, run the second type of sub-model, otherwise, re-run the second general model in a loop iteration to generate a set of full-line process parameters;

[0014] If the outputs of the second-type sub-models all meet the output criterion conditions of the second-type sub-models, the outputs of the second-type sub-models will replace the corresponding contents of the output process parameters of the second overall model, and a set of full-line process parameters will be re-formed. Next, the first-type sub-models will be run. Otherwise, the second overall model will be re-run in a loop iteration to generate a set of full-line process parameters.

[0015] If the outputs of the first type of sub-models all meet the output criterion conditions of the first type of sub-models, the first overall model is run, otherwise the second overall model is re-run in a loop iteration to generate a set of full-line process parameters;

[0016] If the output of the first overall model meets the output criterion conditions of the first overall model, the corresponding full-line process parameters are recorded as the screened full-line process parameters; if the output of the first overall model does not meet the output criterion conditions of the first overall model, the second overall model is re-run in an iterative loop to generate a set of full-line process parameters.

[0017] Preferably, after the process search step, the method further comprises:

[0018] Test verification step: test and verify the selected process parameters of the entire line;

[0019] Wherein, if the performance of the prepared polyacrylonitrile-based carbon fiber reaches the target performance index, the process development is completed;

[0020] Among them, if the performance of the prepared polyacrylonitrile-based carbon fiber does not meet the target performance indicators, the test process data is added to the process data set of the polyacrylonitrile-based carbon fiber, the first general model, the first type of sub-model, the second general model and the second type of sub-model are retrained, and the process search step and the test verification step are re-performed.

[0021] Preferably, the process data set of the polyacrylonitrile-based carbon fiber includes: full-line process parameter data in the preparation process of polyacrylonitrile-based carbon fiber, fiber structure data, fiber performance data and fiber single filament diameter; preferably, the full-line process parameter data includes: polymerization reaction process parameters, spinning process parameters, pre-oxidation process parameters, low-temperature carbonization process parameters and high-temperature carbonization process parameters; preferably, the full-line process parameters also include graphitization process parameters; further preferably, the polymerization reaction process parameters include: polymerization reaction formula, temperature and time; further preferably, the spinning process parameters include: coagulation molding process parameters, water washing process parameters, hot water drawing process parameters, upper Oil concentration, drying and densification process parameters, steam drawing process parameters, shrinkage heat setting process parameters; further preferably, the pre-oxidation process parameters include pre-oxidation temperature parameters, pre-oxidation time parameters and drawing rate; further preferably, the low-temperature carbonization process parameters include low-temperature carbonization temperature parameters and low-temperature carbonization time parameters; further preferably, the high-temperature carbonization process parameters include high-temperature carbonization temperature parameters and high-temperature carbonization time parameters; preferably, the fiber performance data include performance data of polyacrylonitrile precursor and performance data of polyacrylonitrile-based carbon fiber; preferably, the fiber structure data include: the cross-sectional roundness of the primary fiber and the skin thickness of the skin-core structure of the polyacrylonitrile precursor.

[0022] Preferably, in the step of constructing a prediction model: the machine learning algorithm model is a BP neural network model; and / or based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and the performance data of polyacrylonitrile-based carbon fiber as output, and training is performed to obtain a first overall model; and / or based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with the performance data of polyacrylonitrile-based carbon fiber as input and the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output, and training is performed to obtain a second overall model; and / or the performance data of the polyacrylonitrile-based carbon fiber include: tensile strength, tensile modulus; preferably, the performance data of the polyacrylonitrile-based carbon fiber also include: one or more of the CV value of tensile strength, the CV value of tensile modulus, elongation at break, CV value of elongation at break, bulk density, filament diameter, carbon content, and ash content.

[0023] Preferably, the first type of sub-model includes: a first sub-model, a second sub-model, and a third sub-model; wherein, according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model with polymerization reaction process parameters and coagulation molding process parameters as input and the cross-sectional roundness of the spun fiber as output is constructed, and training is performed to obtain the first sub-model; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model with polymerization reaction process parameters and spinning process parameters as input and the performance data of polyacrylonitrile precursor as output is constructed, and training is performed to obtain the second sub-model; preferably, the performance data of the polyacrylonitrile precursor includes tensile strength data and tensile modulus data of the polyacrylonitrile precursor; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model with spinning process parameters as input and the skin thickness data of the skin-core structure of the polyacrylonitrile precursor is constructed, and training is performed to obtain the third sub-model.

[0024] Preferably, in the step of setting the criterion condition, the output criterion condition of the first sub-model is set as follows: the roundness of the cross section of the spun fiber is in a set range, preferably 0.65-0.95. Wherein, the roundness of the cross section of the spun fiber is the ratio of the short axis to the long axis of the cross section of the spun fiber. Preferably, in the step of setting the criterion condition, the output criterion condition of the second sub-model is set as follows: the tensile strength of the polyacrylonitrile precursor is in a set strength range, preferably 750-1350MPa. The tensile modulus of the polyacrylonitrile precursor is in a set modulus range, preferably 7.5-16GPa.

[0025] Preferably, in the step of setting the criterion condition, the output criterion condition of the third sub-model is set as follows: the cortical thickness of the polyacrylonitrile precursor is a set thickness, preferably 120-430 nm.

[0026] Preferably, the second type of sub-model includes: a fourth sub-model, a fifth sub-model, and a sixth sub-model; wherein, according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the high-temperature carbonization process parameters as output is constructed, and training is performed to obtain the fourth sub-model; preferably, the performance parameters of the polyacrylonitrile-based carbon fiber include tensile strength, tensile modulus and carbon content; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the pre-oxidation process parameters as output is constructed, and training is performed to obtain the fifth sub-model; preferably, the poly The performance parameters of acrylonitrile-based carbon fiber include tensile strength, tensile strength CV value, tensile modulus, tensile modulus CV value, and bulk density; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the solidification molding process parameters as output is constructed and trained to obtain the sixth sub-model; preferably, the performance parameters of the polyacrylonitrile-based carbon fiber include tensile strength, tensile strength CV value, tensile modulus, tensile modulus CV value and bulk density; preferably, in the step of setting the criterion condition, the output criterion condition of the fourth sub-model is set as: the maximum temperature of the high-temperature carbonization process ≤2100°C. Preferably, in the step of setting the criterion condition, the output criterion condition of the fifth sub-model is set as: the maximum temperature of the pre-oxidation process ≤275°C; the total drafting multiple of the pre-oxidation process ≤1.3; the time of the pre-oxidation process ≤80min. Preferably, in the step of setting the criterion conditions, the output criterion conditions of the sixth sub-model are set as follows: the temperature of the first coagulation bath ≤ 62°C; the concentration of the first coagulation bath ≤ 68%; and the draft ratio in the first coagulation bath ≤ 0.95.

[0027] Preferably, in the step of setting the criterion conditions, the output criterion conditions of the second overall model are: making the parameter range of each process point in the full-line process parameters within the set effective range; and / or the criterion conditions of the first overall model are: the deviation between the performance data of the polyacrylonitrile-based carbon fiber output by the first overall model and the target performance index is ≤5%.

[0028] Preferably, in the process search step: when the number of groups of screened process parameters for the entire line is less than the set number of groups, the second overall model is re-run in an iterative loop to generate a set of process parameters for the entire line, and the process search step is performed again until the number of groups of screened process parameters for the entire line reaches the set number of groups; preferably, the set number of groups is 20 groups.

[0029] Preferably, in the test verification step: one or more groups of screened full-line process parameters are tested and verified, if the performance of the polyacrylonitrile-based carbon fiber prepared by one or several groups of the screened full-line process parameters reaches the target value, the process development is completed; and / or one or more groups of screened full-line process parameters are tested and verified, if the performance of the prepared polyacrylonitrile-based carbon fiber does not reach the target value; it is necessary to add the test verification data to the process data set of the polyacrylonitrile-based carbon fiber, retrain and correct the first overall model, the second overall model, the first type of sub-model, and the second type of sub-model; then, the second overall model is re-run in an iterative loop to generate a set of full-line process parameters, and the process search step and the test verification step are re-performed.

[0030] On the other hand, an embodiment of the present invention provides a polyacrylonitrile-based carbon fiber, wherein the polyacrylonitrile-based carbon fiber is prepared by a process developed by the intelligent development method of the preparation process of the polyacrylonitrile-based carbon fiber described in any one of the above items;

[0031] Preferably, the tensile strength of the polyacrylonitrile-based carbon fiber is 2.6-5.1 GPa, the tensile modulus is 560-810 GPa, the diameter is 4.2-5.2 μm, and the bulk density is 1.88-1.95 g / cm 3 ;

[0032] Preferably, the polyacrylonitrile-based carbon fiber is characterized by small-angle X-ray scattering, when the scattering vector When the scattering intensity I(q)∝q -α , α takes a value of 3.25-3.7; when the scattering vector When the scattering intensity I(q)∝q -α , α takes a value of 1.5-2.0.

[0033] Compared with the prior art, the intelligent development method of the preparation process of polyacrylonitrile-based carbon fiber of the present invention has at least the following beneficial effects:

[0034] The embodiment of the present invention provides an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, which mainly includes the following steps: a prediction model construction step: based on a process data set of polyacrylonitrile-based carbon fiber, a first general model, a first type of sub-model, a second general model and a second type of sub-model are constructed; wherein the first general model is a machine learning algorithm model with full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and performance data of polyacrylonitrile-based carbon fiber as output; the first type of sub-model is a machine learning algorithm model with local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and fiber performance data and / or structural data after local process treatment as output; the second general model is a machine learning algorithm model with performance data of polyacrylonitrile-based carbon fiber as input and full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output; the second type of sub-model is a machine learning algorithm model with performance data of polyacrylonitrile-based carbon fiber as input and local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output; the judgment condition setting step: setting the target performance index of the required polyacrylonitrile-based carbon fiber; setting the parameter range of each process point in the full-line process parameters; setting the second type of sub-model; the prediction model is a prediction model of the whole process ... The output criteria of the first general model, the second general model, the first sub-model and the second sub-model are as follows: a process search step: according to the target performance index of the set polyacrylonitrile-based carbon fiber, the second general model is first run to generate a set of full-line process parameters; if the generated full-line process parameters meet the output criteria of the second general model, the second sub-model is run, otherwise the second general model is re-run in a loop iteration; if the outputs of the second sub-models all meet the output criteria of the second sub-models, the outputs of the second sub-models are replaced with the corresponding contents of the output process parameters of the second general model to form a set of full-line process parameters, and then the first sub-models are run, otherwise the second general model is re-run in a loop iteration; if the outputs of the first sub-models all meet the output criteria of the first sub-models, the first general model is run, otherwise the second general model is re-run in a loop iteration; if the outputs of the first general model meet the output criteria of the first general model, the corresponding full-line process parameters are recorded as the screened full-line process parameters; if the output of the first general model does not meet the output criteria of the first general model, the second general model is re-run in a loop iteration. Here, the present invention adopts the above-mentioned scheme, collects and organizes historical process-performance-structure data, uses a specific machine learning algorithm, and constructs a carbon fiber performance and process bidirectional prediction-optimization model based on nested sub-model criteria through training and verification (that is, designs two overall models and two types of sub-models to achieve bidirectional prediction).To further explain, the core idea of ​​the bidirectional prediction-process search model is as follows: the prediction accuracy of the first total model (the model that predicts performance from the process) is much higher than that of the second total model (the model that predicts the process from the performance); the prediction accuracy of the sub-model (two types of sub-models) is much higher than that of the total model (the first total model and the second total model); starting from the target performance, the cyclic iterative process search starts with the output of the second type of total model, and the model with low prediction accuracy (the second total model) is reversely trained through the subsequent high prediction accuracy model (sub-model and the first total model), and at the same time, the real-time correction of the bidirectional model with the help of experimental verification data is achieved to achieve the intelligent design of the target performance fiber preparation process. With the help of the above model, the intelligent design and development of fiber performance and its preparation process can be realized for multiple performance requirements, which significantly reduces the trial and error cost of the complex system engineering debugging process (reducing time, energy consumption, and capital investment).

[0035] Furthermore, an embodiment of the present invention provides an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, which, after the process search step, also includes an experimental verification step: experimental verification is performed on the screened process parameters of the entire line; wherein, if the performance of the prepared polyacrylonitrile-based carbon fiber reaches the target performance index, the process development is completed; wherein, if the performance of the prepared polyacrylonitrile-based carbon fiber does not reach the target performance index, the process data of the experimental process is added to the process data set of the polyacrylonitrile-based carbon fiber, the first overall model, the first type of sub-model, the second overall model and the second type of sub-model are retrained, and the process search step and the experimental verification step are re-performed. Here, the overall concept of the present invention combines the machine learning model and process search model of bidirectional prediction of fiber properties and preparation process with experimental verification, which can effectively reduce the probability of "trial and error" in the development of long and complex process technology, thereby reducing the development cost of high modulus fiber; it can effectively play the advantages of the two parts (model and experimental verification) and make up for their respective shortcomings. The prediction and process search model can continuously output system process parameters that meet the requirements to provide exploration direction or test content for the experiment, greatly avoiding the risk of blind trial and error, and the experimental verification continuously corrects the prediction model through experimental data feedback, timely and real-time avoids the shortcomings of model overfitting or underfitting, improves the prediction accuracy, and improves the quality of the process group output by the cyclic iterative process search; it can provide a research paradigm for the development of complex process systems for carbon fibers, using machine learning models to mine the non-explicit bidirectional relationship between process parameters and performance or structural parameters in past research data sets, and using models to exhaust various combinations of process parameters within the effective range of the entire line, providing early screening and prediction for experimental verification, and continuously correcting the model with the help of real experimental data. Compared with the traditional development method of theoretical modeling + experimental verification, it is more realistic, more operational, more effective, and more in line with the characteristics of the current data-driven-intelligent era.

[0036] Furthermore, an embodiment of the present invention provides an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, by making the first type of sub-model include a first sub-model, a second sub-model, and a third sub-model; wherein, according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with polymerization reaction process parameters and coagulation molding process parameters as input and the cross-sectional roundness of the spun fiber as output, and training is performed to obtain the first sub-model; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with polymerization reaction process parameters and spinning process parameters as input and the performance data of polyacrylonitrile precursor as output, and training is performed to obtain the second sub-model; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with spinning process parameters as input and the skin thickness data of the skin-core structure of the polyacrylonitrile precursor as output, and training is performed to obtain the third sub-model. Here, the present invention sets the first type of sub-model to the above three sub-models, and nests the first type of sub-model with the first overall model to construct a prediction model from process to fiber performance. The reasons are as follows: 1. The sub-model reflects the implicit relationship between process parameters and process fiber performance or structural parameters in the local process; the overall model reflects the overall situation, the amount of effective data for local training is much larger than the amount of global data, and the prediction accuracy of the sub-model is much higher than that of the overall model. The use of sub-models or multiple sub-models can effectively make up for the shortcomings of low prediction accuracy of the overall model and improve the accuracy of process selection; 2. The global process flow is too long and there are too many process parameters. If only one overall model is used, the search volume of the iterative process search process is too huge, which consumes computing resources. With the use of sub-models with higher prediction accuracy, large parameters can be filtered out in advance. 1. The process flow of the whole line is long, and it is necessary to set up multiple sub-models at multiple core process control points along the process flow to screen the process parameter groups in advance around multiple process control points. Under the premise of ensuring the process screening of the core process control points, the process parameters of the whole line are searched iteratively, thereby improving the efficiency of the iterative cycle while maintaining a high accuracy rate; 2. Specifically, if one or two sub-models are selected, the process parameter group output by the iterative process search process is very likely to deviate from the current principled cognition, and the output process parameter group intuitively feels of low quality, which increases the burden of later experimental verification; 3. In theory, more sub-models can be selected, but too many sub-models will lead to a reduction in the effective space for process screening, which is not conducive to the exploration of new process directions. Considering the above, three first-class sub-models are selected along the process flow.

[0037] Further, the embodiment of the present invention provides an intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber, by making the second type of sub-model include a fourth sub-model, a fifth sub-model, and a sixth sub-model; wherein, according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the high-temperature carbonization process parameters as output is constructed, and training is performed to obtain the fourth sub-model; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the pre-oxidation process parameters as output is constructed, and training is performed to obtain the fifth sub-model; according to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the solidification molding process parameters as output is constructed, and training is performed to obtain the sixth sub-model. Here, the present invention sets the second type of sub-model to the above three sub-models, and at the same time nests the second type of sub-model with the second total model to construct a prediction model from fiber performance to preparation process, and constructs a full-line process parameter prediction output from the target performance through the second total model, thereby starting the full-line process cyclic iterative search. The second type of sub-model is connected to the second overall model. The reasons why the second type of sub-model is set to the above three sub-models are as follows: 1. The sub-model reflects the implicit relationship between the carbon fiber performance or structural parameters and the process parameters in the local process; the overall model reflects the overall situation (carbon fiber performance and full-line process parameters). The effective data volume of the sub-model focusing on local training is much larger than the global data volume, and the prediction accuracy of the sub-model is much higher than that of the overall model. The use of sub-models or multiple sub-models can effectively make up for the shortcomings of the low prediction accuracy of the overall model and improve the accuracy of process selection; 2. The global process flow is too long and there are too many process parameters. If only one overall model is used, the search volume of the cyclic iterative process search process is too huge, which consumes computing resources. With the use of sub-models with higher prediction accuracy, the output of the full-line process parameters can be optimized to improve the process search. Search efficiency; 3. The whole process is long, and it is necessary to set up multiple sub-models at multiple core process control points along the process to optimize the process parameter output around multiple process control points and improve the efficiency of loop iteration; 4. Specifically, three sub-models are selected along the core control points of the process flow. The process control points reflected by the process parameters output by the three sub-models have a high correlation with the performance of carbon fiber high modulus; 5. Specifically, if one or two sub-models are selected, there are too few output optimization screening points for the whole process parameters, which is very easy to deviate from the current principled cognition, and the output process parameter group intuitively feels low in quality, which increases the burden of later experimental verification; 6. In theory, more sub-models can be selected, but too many sub-models will lead to too many control points for process optimization output, which is not conducive to the exploration of new process directions. Considering the above, three second-type sub-models are selected along the process flow.

[0038] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a new paradigm for the research and development of carbon fiber materials (see Figure 1 Figure a in Figure 2, characterized by data-driven artificial intelligence) and traditional research paradigms (see Figure 1 (b) in Figure 2;

[0040] Figure 2 It is a flow chart of an intelligent development method for a preparation process of a polyacrylonitrile-based carbon fiber provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the present invention application are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

[0042] The present invention discloses an intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber. By collecting and collating historical process, performance and structural data, with the help of a specific machine learning algorithm, a carbon fiber performance and process bidirectional prediction-optimization model with nested sub-model criteria is constructed through training and verification. With the help of this model, the intelligent design and development of fiber performance and its preparation process can be achieved facing multiple performance requirements.

[0043] like Figure 2 As shown, the intelligent development method of a preparation process of polyacrylonitrile-based carbon fiber of the present invention comprises the following steps:

[0044] Steps for constructing the process data set of the polyacrylonitrile-based carbon fiber:

[0045] Among them, the constructed polyacrylonitrile-based carbon fiber process data set is a standardized polyacrylonitrile-based carbon fiber preparation system data set, which includes: polymerization-spinning-carbonization process data, fiber performance data corresponding to the process, corresponding fiber structure data, and fiber single filament diameter.

[0046] Among them, the polymerization-spinning-carbonization process data include: the formula of the polymerization reaction, the temperature and time of the polymerization reaction; the coagulation molding process parameters (temperature, concentration, time, drawing ratio), water washing process parameters (temperature, time), hot water drawing process parameters (temperature, drawing ratio), oiling concentration, drying densification process parameters (temperature, time), steam drawing process parameters (temperature, drawing ratio), shrinkage heat setting process parameters; pre-oxidation process parameters (temperature, time, drawing ratio), low-temperature carbonization process parameters (temperature, time, preferably including drawing ratio), high-temperature carbonization process parameters (temperature, time, preferably also including drawing ratio, etc.), graphitization process parameters (temperature, time, preferably also including drawing ratio, etc.).

[0047] Among them, the fiber performance data corresponding to the process include: tensile strength, tensile modulus, preferably also including the CV value of tensile strength, the CV value of tensile modulus, elongation at break, CV value of elongation at break, bulk density, single fiber diameter, carbon content, ash content, etc.

[0048] The corresponding fiber structure data include: skin-core structure skin thickness, fiber roundness, and preferably also fiber orientation, etc.

[0049] Specifically, the laboratory to pilot research and development data, engineering research and development data, and industrial production data of different brands of carbon fiber (T300, T700, T800, T1000, T1100, M40J, M46J, M55J, M60J, M65J, etc.) will be standardized and normalized. The data of different research and development stages will be unified and summarized, and the data will be screened / supplemented / standardized and normalized. The data of systematic research results in the past and recent periods around the world (articles / patents containing system processes, fiber properties or fiber structure information, etc.) will be collected, screened, and organized, and standardized and normalized.

[0050] Among them, data is unified and standardized: process parameter data are organized in a unified process flow sequence. The process flow settings may be different in different R&D stages. When the actual data is organized according to the standard process, unified data is supplemented in the blanks. For example: the standard hot water drawing is a four-level setting, and the actual pilot data may be a three-level setting. When the standard is organized, the fourth-level temperature is set to 20°C and the drawing ratio is set to 1.

[0051] Among them, data normalization: the specific data corresponding to each parameter is converted into a number between [0,1], the purpose is to eliminate the order of magnitude difference between the data of each dimension, to avoid large network prediction errors caused by large order of magnitude differences. Specific commonly used methods: maximum and minimum method, mean variance method, can be processed using MATLAB's own function mapminmax.

[0052] Steps for constructing a prediction model: Based on the process data set of polyacrylonitrile-based carbon fiber, a first overall model, a first type of sub-model, a second overall model and a second type of sub-model are constructed.

[0053] According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (the first overall model) of the whole line process parameters (polymerization reaction process parameters, spinning process parameters, pre-oxidation process parameters, low-temperature carbonization process parameters and high-temperature carbonization process parameters) → carbon fiber performance parameters (such as strength and CV value, modulus and CV value, elongation at break and CV value, bulk density, single filament diameter, carbon content, ash content, etc.) is constructed and trained.

[0054] Before training the BP neural network model, the standardized and normalized data set can be randomly divided into a training set (ratio 45-90%), a validation set (ratio 5-20%) and a test set (ratio 5-35%). The sum of the ratios of the training set, validation set and test set is always 1; the training set is used to train the neural network model, the validation set is used to prevent the overfitting of the neural network model, and the training set and validation set are used when training the BP neural network model; the test set is used to test the accuracy of the neural network model and is used after the model training is completed.

[0055] Train the first total model: select the transfer function and training method and train the first total model; the comprehensive accuracy (training, validation and testing) characteristics of the first total model are not less than 0.88; the accuracy is to use the data set as input and obtain the predicted value through the trained model. The linear regression correlation between the predicted value and the actual value is the accuracy of the model.

[0056] The training methods mainly include the steepest descent method, the additional momentum method, the adaptive learning rate method, the momentum-adaptive learning rate adjustment method and the quantized conjugate gradient algorithm. The following embodiments use the steepest descent algorithm, but other algorithms may also be selected.

[0057] The second general model is a machine learning algorithm model that takes the performance data of polyacrylonitrile-based carbon fiber as input and the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output. Specifically, a machine learning algorithm model (the second general model) of carbon fiber performance parameters (such as one or more of strength and CV value, modulus and CV value, elongation at break and CV value, bulk density, filament diameter, carbon content, ash content, etc.) → full-line process parameters (polymerization reaction process parameters, spinning process parameters, pre-oxidation process parameters, low-temperature carbonization process parameters and high-temperature carbonization process parameters) is constructed based on the data set and trained.

[0058] Here, the training method of the second total model refers to the first total model, and finally, the comprehensive accuracy (training, verification and testing) characteristics of the second total model are not less than 0.75.

[0059] The first sub-model is a machine learning algorithm model that takes the local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and takes the fiber performance data and / or structure data after local process treatment as output. Preferably, the first sub-model includes the first sub-model, the second sub-model and the third sub-model, as follows:

[0060] Based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (first sub-model) of polymerization reaction process parameters, spinning process parameters → cross-sectional roundness of nascent fiber is constructed and trained. Here, the cross-sectional roundness of nascent fiber is the ratio of the short axis to the long axis of the cross section measured by optical microscope observation.

[0061] Before model training, the standardized and normalized data set can be randomly divided into a training set (ratio 45-90%), a validation set (ratio 5-20%) and a test set (ratio 5-35%). Among them, the sum of the ratios of the training set, validation set and test set is always 1; the training set is used to train the neural network model, the validation set is used to prevent overfitting of the neural network model, and the training set and validation set are used when training the BP neural network model; the test set is used to test the accuracy of the neural network model and is used after the model training is completed. Train the first sub-model: select the transfer function and training method and train the first sub-model; the comprehensive accuracy of the first sub-model (training, validation and test characteristics are not less than 0.94; accuracy is to use the data set as input, and obtain the predicted value through the trained model. The linear regression correlation between the predicted value and the actual value is the accuracy of the model.

[0062] Based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (second sub-model) of polymerization reaction process parameters, spinning process parameters → properties (strength and modulus) of polyacrylonitrile precursor is constructed and trained. The training of the second sub-model refers to the training method of the first sub-model, which will not be repeated here.

[0063] Based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (third sub-model) of spinning process parameters → cortical thickness of the skin-core structure of polyacrylonitrile precursor was constructed and trained. Among them, the cortical thickness of the skin-core structure of polyacrylonitrile precursor: the radial dissolution rate of the precursor was obtained by the solubility method, and the thickness corresponding to the turning point of the dissolution rate was defined as the cortical thickness. The training of the third sub-model refers to the training method of the first sub-model, which will not be repeated here.

[0064] The second sub-model is a machine learning algorithm model that uses the performance data of polyacrylonitrile-based carbon fiber as input and the local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output. Preferably, the second sub-model includes the fourth sub-model, the fifth sub-model and the sixth sub-model, as follows:

[0065] Based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (fourth sub-model) of the performance parameters (strength, modulus, carbon content) of polyacrylonitrile-based carbon fiber → high-temperature carbonization process parameters is constructed and trained. The training of the fourth sub-model refers to the training method of the first sub-model. The comprehensive accuracy (training, verification and testing) characteristics of the fourth sub-model are not less than 0.81, and the rest are not repeated here.

[0066] Based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (fifth sub-model) of the performance parameters (strength, modulus, elongation at break) of polyacrylonitrile-based carbon fiber → pre-oxidation process parameters (number of temperature steps, temperature at each step, total draft multiple) of the model is constructed and trained. The training of the fifth sub-model refers to the training method of the first sub-model. The comprehensive accuracy (training, verification and testing) characteristics of the fifth sub-model are not less than 0.80, and the rest are not repeated here.

[0067] Based on the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model (sixth sub-model) of the performance parameters (strength, modulus, elongation at break) of polyacrylonitrile-based carbon fiber → solidification molding process parameters (solidification bath level, temperature and concentration at each level) is constructed and trained. The training of the sixth sub-model refers to the training method of the first sub-model. The comprehensive accuracy (training, verification and testing) characteristics of the sixth sub-model are not less than 0.80, and the rest are not repeated here.

[0068] The step of setting the criterion conditions: setting the target performance index of the required polyacrylonitrile-based carbon fiber; setting the output criterion conditions of the first general model, the second general model, the first sub-model, and the second sub-model. Preferably, in this step, it also includes: limiting the effective range of the process control points of the whole line; limiting the criterion range of the polyacrylonitrile precursor performance range, the roundness of the primary fiber, and the thickness of the polyacrylonitrile precursor skin layer.

[0069] Process search step: according to the set target performance index of polyacrylonitrile-based carbon fiber, first run the second general model to generate a set of full-line process parameters; if the generated full-line process parameters meet the output criterion conditions of the second general model, run the second type of sub-model, otherwise, re-run the second general model in a loop iteration to generate a set of full-line process parameters;

[0070] If the outputs of the second-type sub-models all meet the output criterion conditions of the second-type sub-models, the outputs of the second-type sub-models will replace the corresponding contents of the output process parameters of the second overall model, and a set of full-line process parameters will be re-formed. Next, the first-type sub-models will be run. Otherwise, the second overall model will be re-run in a loop iteration to generate a set of full-line process parameters.

[0071] If the outputs of the first type of sub-models all meet the output criterion condition of the first type of sub-models, then the first overall model is run, otherwise the second overall model is re-run in a loop iteration;

[0072] If the output of the first overall model meets the output criterion conditions of the first overall model, the corresponding full-line process parameters are recorded as the screened full-line process parameters; if the output of the first overall model does not meet the output criterion conditions of the first overall model, the second overall model is re-run in an iterative loop.

[0073] In addition, when the number of groups of screened process parameters for the entire line is less than the set number of groups, the process search step is performed again until the number of groups of screened process parameters for the entire line reaches the set number of groups; preferably, the set number of groups is 20 groups.

[0074] Experimental verification step: experimental verification is carried out on the screened process parameters of the entire line; wherein, if the performance of the prepared polyacrylonitrile-based carbon fiber reaches the target performance index, the process development is completed; wherein, if the performance of the prepared polyacrylonitrile-based carbon fiber does not reach the target performance index, the process data of the experimental process is added to the process data set of the polyacrylonitrile-based carbon fiber, the first overall model, the first type of sub-model, the second overall model and the second type of sub-model are retrained, and the process search step and the experimental verification step are re-performed.

[0075] In summary, the prediction model of the embodiment of the present invention is a bidirectional optimization design model. The core idea is as follows: the prediction accuracy of the first total model (predicting performance from process) is much higher than that of the second total model (predicting process from performance); the prediction accuracy of the sub-model is much higher than that of the total model; starting from the target performance, the cyclic iterative process search starts from the output of the second type of total model, and the model with low prediction accuracy (the second total model) is reversely trained through the model with high prediction accuracy (the sub-model and the first total model), and at the same time, the real-time correction of the bidirectional model is carried out with the help of experimental verification data to achieve the intelligent design of the fiber preparation process with target performance.

[0076] The operation process of the prediction model (bidirectional optimization design model) of the present invention is as follows: input the target fiber performance parameters, run the second general model, generate a set of full-line process parameters, if the full-line process parameters are all within the specified effective range, continue to execute the fourth sub-model, the fifth sub-model, and the sixth sub-model, otherwise the second sub-model is re-executed in a loop iteration (adjusting the parameters and re-outputting). When the prediction results of the fourth sub-model, the fifth sub-model, and the sixth sub-model all meet the criteria, the output of the second sub-model is replaced with the corresponding content of the output process parameters of the second general model, and a set of full-line process parameters is re-formed. Next, continue to run the first sub-model, the second sub-model, and the third sub-model, otherwise the second general model is re-executed in a loop iteration (adjusting the parameters and re-outputting). When the prediction results of the first sub-model, the second sub-model, and the third sub-model all meet the criteria, continue to run the first general model, otherwise the second general model is re-executed in a loop iteration (adjusting the parameters and re-outputting). Compare the prediction results of the first general model with the target performance. If the deviation is ≤5%, continue to execute, otherwise re-run the second general model (adjusting the parameters and re-outputting). If the number of predicted process parameter groups that meet the requirements reaches 20, process test verification is carried out. If it is less than 20, the second general model is re-run. Test verification is carried out to measure the fiber performance. If the fiber performance reaches the target, it ends. If the measured fiber performance does not meet the target, the measured data and the corresponding process group parameters are added to the data set, and the first general model, the second general model, and all sub-models are retrained and corrected. After the model correction training is completed, it starts again until the test fiber performance reaches the target performance requirements.

[0077] The following is further described in detail by specific embodiments:

[0078] Example 1

[0079] This embodiment proposes an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, wherein the specific steps are as follows:

[0080] (I) Constructing prediction models (overall model and sub-models)

[0081] 1) Construct a process data set of polyacrylonitrile-based carbon fiber in Excel or Matlab software, including: polymerization-spinning-carbonization process data, fiber performance data corresponding to the process, and corresponding fiber structure data.

[0082] Standardize and normalize the laboratory to pilot research and development data, engineering research and development data, and industrial production data of different brands of carbon fiber (T300, T700, T800, T1000, T1100, M40J, M46J, M55J, M60J, M65J);

[0083] Summarize the data from different R&D stages, screen / supplement the data, standardize and normalize them.

[0084] Collect, screen, and organize past and recent systematic research results data worldwide (articles / patents containing system processes, fiber properties or fiber structure information, etc.), and standardize and normalize them.

[0085] 2) Construction of the overall model and sub-models (constructed in MATLAB software, which has a machine learning algorithm embedded in it)

[0086] The first general model: Establish a BP neural network model of full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber → carbon fiber performance. Take the full-line process parameters as input, carbon fiber performance data (strength, modulus, etc.) as output, set the initial hidden layer number and initial hidden layer node number of the neural network, and establish a neural network model that predicts fiber performance from full-line process data.

[0087] Before training the BP neural network model, the standardized and normalized data set can be randomly divided into a training set (80%), a validation set (10%) and a test set (10%). The sum of the proportions of the training set, validation set and test set is always 1; the training set is used to train the neural network model, and the validation set is used to prevent overfitting of the neural network model. The training set and validation set are used during the training of the BP neural network model; the test set is used to test the accuracy of the neural network model and is used after the model training is completed.

[0088] Train the first total model: select the transfer function and training method and train the total model; the comprehensive accuracy (training, validation and testing) of the first total model is not less than 0.88; the accuracy is the linear regression correlation between the predicted value and the actual value obtained by taking the data set as input through the trained model.

[0089] The training methods mainly include the steepest descent method, the additional momentum method, the adaptive learning rate method, the momentum-adaptive learning rate adjustment method and the quantized conjugate gradient algorithm. The steepest descent algorithm is selected in this embodiment, and other algorithms may also be selected.

[0090] Second general model: Establish a BP neural network model of carbon fiber performance parameters → full-line process parameters. The construction process is the same as the first general model; the comprehensive accuracy (training, verification and testing) characteristics of the second general model are not less than 0.75.

[0091] The first sub-model: Establish a BP neural network model of polymerization reaction process parameters, spinning process parameters → cross-sectional roundness of nascent fibers. With polymerization reaction process parameters and spinning process parameters as input, and cross-sectional roundness of nascent fibers as output, the initial number of hidden layers and the number of nodes of the initial hidden layer of the neural network are set, and a BP neural network model is established to predict the roundness of nascent fibers from polymerization reaction process parameters and spinning process parameters. Cross-sectional roundness of nascent fibers: The ratio of the short axis to the long axis of the cross section is measured by optical microscope observation.

[0092] Before model training, the standardized and normalized data set can be randomly divided into a training set (85%), a validation set (10%), and a test set (5%). The sum of the proportions of the training set, validation set, and test set is always 1; the training set is used to train the neural network model, the validation set is used to prevent overfitting of the neural network model, and the training set and validation set are used when training the BP neural network model; the test set is used to test the accuracy of the neural network model and is used after the model training is completed.

[0093] Train the first sub-model: select the transfer function and training method and train the overall model; the overall model comprehensive accuracy (training, validation and testing) characteristics are not less than 0.94; the accuracy is the linear regression correlation between the predicted value and the actual value obtained by taking the data set as input and the trained model;

[0094] Second sub-model: Based on the process data set of polyacrylonitrile-based carbon fiber, a BP neural network model (second sub-model) of polymerization reaction process parameters, spinning process parameters → properties of polyacrylonitrile precursor (tensile strength and tensile modulus) is constructed and trained (the construction process is the same as the first sub-model).

[0095] The third sub-model: According to the process data set of polyacrylonitrile-based carbon fiber, a BP neural network model of spinning process parameters → cortical thickness of polyacrylonitrile precursor skin-core structure is constructed and trained (the construction process is the same as the first sub-model); Cortical thickness of precursor skin-core structure: the solubility method is used to obtain the radial dissolution rate of the precursor, and the thickness corresponding to the turning point of the dissolution rate is defined as the cortical thickness.

[0096] Fourth sub-model: Establish a BP neural network model of carbon fiber performance parameters (tensile strength, tensile modulus, carbon content) → high-temperature carbonization process parameters, and train it. The comprehensive accuracy (training, verification and testing) characteristics of the fourth sub-model are not less than 0.81, and the construction process is the same as the first sub-model.

[0097] The fifth sub-model: establish a BP neural network model of carbon fiber performance parameters (tensile strength, tensile modulus, elongation at break) → pre-oxidation process parameters (number of temperature steps, temperature at each step, total draft multiple), and train it. The construction process is the same as the fourth sub-model;

[0098] The sixth sub-model: establish a BP neural network model of carbon fiber performance parameters (tensile strength, tensile modulus, elongation at break) → solidification molding process parameters (solidification bath level, temperature and concentration at each level), and train it. The construction process is the same as the fourth sub-model.

[0099] (II) Output criteria of the overall model and sub-models

[0100] The first general model predicts and outputs the tensile strength and tensile modulus of polyacrylonitrile-based carbon fibers. The output criteria (the target range of fiber performance in this embodiment) are as follows: tensile strength 3.2-3.8 GPa, tensile modulus 665-730 GPa, diameter 4.3-5.1 μm, and bulk density 1.89-1.95 g / cm 3 .

[0101] The second overall model predicts the process parameters of the entire line, and the output criteria refer to (III) the range of fiber process parameters of the entire line.

[0102] The first sub-model predicts and outputs the roundness of the as-spun fiber, and its output criterion is: roundness 0.80-0.90.

[0103] The second sub-model predicts and outputs the tensile strength and tensile modulus of polyacrylonitrile precursor, and its output criteria are: tensile strength is 780-1120MPa, and tensile modulus is 8-15Gpa.

[0104] The third sub-model predicts the cortical thickness of the polyacrylonitrile precursor, and its output criterion is: 180-320nm.

[0105] The fourth sub-model predicts the high temperature carbonization process, and its output criterion is: the maximum temperature of the high temperature carbonization process is ≤2100℃.

[0106] The fifth sub-model predicts the pre-oxidation process, and its output criteria are: the maximum temperature of the pre-oxidation process ≤ 275°C; the total drafting multiple of the pre-oxidation process ≤ 1.3; and the time of the pre-oxidation process ≤ 80 min.

[0107] The sixth sub-model predicts the solidification molding process, and its output criteria are: the temperature of the first solidification bath ≤ 62°C; the concentration of the first solidification bath ≤ 68%; the draft ratio in the first solidification bath ≤ 0.95.

[0108] (III) Full-line process parameter range (output criteria of the second overall model)

[0109] 1) Preparation of spinning solution

[0110] The monomer acrylonitrile, the monomer itaconic acid, and the monomer 2-acrylamide-2-methylpropane sulfonic acid are mixed in a mass ratio of (90.0-99.9): (0.1-5): (0.00-5.0), and added into a reactor with a stirring device. The mass percentage of the total monomer mass in the reaction solution is 10-25%. Azobisisobutyl cyanide is used as an initiator, and the amount of the initiator is 0.1-2.0% of the total weight of the monomer. The solvent is selected from one of dimethylformamide, dimethylacetamide, dimethyl sulfoxide, ethylene carbonate, propylene carbonate, and γ-butyrolactone, and solution polymerization is carried out. The polymerization reaction temperature is controlled in the range of 30-85°C, and the polymerization time is 3-36h to obtain a polyacrylonitrile spinning solution.

[0111] 2) Solidification molding

[0112] The wet spinning method is adopted. After the spinning solution is ejected from the spinneret, it is sequentially solidified in 2-8 levels of solidification baths. The solidification bath consists of three parts: solvent, solidifying agent, and regulator. The solvent is selected from one of dimethylformamide, dimethylacetamide, dimethyl sulfoxide, ethylene carbonate, propylene carbonate, and γ-butyrolactone. The solidifying agent is deionized water. The regulator is selected from one of ethanol, n-propanol, n-butanol, isobutanol, ammonium bicarbonate, and ammonia water.

[0113] a) Primary solidification molding

[0114] The temperature of the coagulation bath is 35-65°C; the coagulation stretching rate is 0.4-1.2 times; the circulation flow rate of the coagulation bath is 4-14 mm / s; the coagulation residence time is 0.1-3 min; the mass fraction of the solvent in the coagulation bath is 35-70%; and the mass concentration of the regulator in the coagulation bath is 0.01-0.1 mol / L.

[0115] b) Final stage solidification

[0116] The coagulation bath temperature is 50-95°C; the coagulation drawing rate is 1-2 times; the coagulation bath circulation flow rate is 4-14mm / s; the coagulation residence time is 0.2-1min; the mass fraction of the solvent in the coagulation bath is 0-15%; the mass concentration of the regulator in the coagulation bath is 0.01-0.03mol / L.

[0117] 3) Water washing process

[0118] The washing temperature is 40-80°C; the washing time is 1-3 minutes.

[0119] 4) Hot water drawing process

[0120] The hot water drawing temperature is 70-95°C; the drawing ratio is 1-4 times.

[0121] 5) Oiling process

[0122] Use silicone oil at a concentration of 0.5-3%.

[0123] 6) Drying and densification process

[0124] A low-temperature, long-time gradient drying and densification process is adopted.

[0125] The temperature gradient is 4-18 levels; each level uses a different temperature, and the temperature increases step by step. The temperature of the first level is 80-100℃; the temperature of the last level is 115-140℃; the temperature difference between adjacent temperature gradients is 0-8℃; the drying and densification time of each level is 4-9s.

[0126] 7) Steam drawing and shrinkage heat setting process

[0127] Steam drawing process: saturated water vapor or superheated steam is used as the medium to apply high-multiple drawing to the tow. The steam pressure is 0.15-0.4MPa; the drawing multiple is 2-4 times; the steam drawing residence time is 2-5s.

[0128] Shrinkage heat setting process: saturated water vapor or superheated steam is used as the medium. The steam temperature is 110-180℃; the draft ratio is 0.9-1 times; the shrinkage heat setting residence time is 2-5s.

[0129] 8) Pre-oxidation process

[0130] A 3-8-stage hot air medium pre-oxidation process is adopted, and the pre-oxidation temperature increases step by step. The first stage pre-oxidation temperature is 170-200℃; the last stage pre-oxidation temperature is 240-280℃; the temperature difference between adjacent temperature gradients is 5-35℃; the total pre-oxidation draft multiple is 0.8-1.4 times; the total pre-oxidation residence time is 20-90min.

[0131] 9) Low temperature carbonization process

[0132] High-purity nitrogen protection, low-temperature carbonization temperature is 300-900℃; low-temperature carbonization time is 1-6min.

[0133] 10) High temperature carbonization process

[0134] Under high-purity nitrogen protection conditions, 2-6 levels of high-temperature carbonization treatment. The first level temperature is 900-1000℃; the highest level carbonization temperature is 1300-2100℃; the temperature difference between adjacent temperature gradients is 50-600℃; the high-temperature carbonization time is 0.4-5min.

[0135] 11) Graphitization process

[0136] Under the protection of high-purity argon, 2-4 levels of high-temperature graphitization treatment. The first level temperature is 2050-2400℃; the highest level temperature is 2700-3050℃; the temperature difference between adjacent temperature gradients is 200-500℃; the graphitization draft ratio is 1.02-1.1 times; the graphitization treatment time is 0.35-4min.

[0137] (IV) Fiber test performance and verified optimal full-line process parameters

[0138] Taking carbon fiber strength and modulus as the target (the target range of fiber performance in this embodiment: strength 3.2-3.8Gpa, modulus 665-730Gpa), four rounds of bidirectional optimization model prediction and experimental verification (prediction method, see Figure 2 Finally, 20 sets of process parameters were output. After experimental verification, the fiber performance of 5 sets of full-line process parameters reached the target range.

[0139] The test performance list of the five groups of fibers that achieved the target is shown in Table 1:

[0140] Table 1

[0141]

[0142] Among them, the carbon fiber corresponding to No. 3 has the best comprehensive performance (both strength and modulus), and the full-line process parameters corresponding to No. 3 are as follows:

[0143] 1) Preparation of spinning solution

[0144] The monomer acrylonitrile, the monomer itaconic acid, and the monomer 2-acrylamide-2-methylpropane sulfonic acid are added into a reactor with a stirring device in a ratio of 99.09, 0.9, and 0.01. Among them, in the reaction solution, the mass percentage of the total monomers is 22%, the amount of the initiator azobisisobutyl cyanide is 1.0% of the total weight of the monomers, and dimethyl sulfoxide is used as a solvent to carry out solution polymerization. During the polymerization process, high-purity nitrogen is first bubbled into the reactor for 15 minutes to replace the air in the reaction system, and then stirring is started and the temperature is raised to 68°C. Under the protection of nitrogen, the polymerization is carried out at a constant temperature for 12 hours to obtain a ternary copolymerized polyacrylonitrile-based carbon fiber spinning solution.

[0145] 2) Solidification molding

[0146] The wet spinning method is adopted, and the spinning solution is ejected from the spinneret and then solidified and formed in six stages of solidification baths in sequence. The composition of the solidification bath solution of each stage is as follows: the solvent is dimethyl sulfoxide, the coagulant is water, and the regulator is ammonia water.

[0147] a) First stage solidification molding

[0148] The temperature of the coagulation bath is 50°C; the coagulation drawing ratio is 0.65 times; the circulation flow rate of the coagulation bath is 14 mm / s; the coagulation residence time is 0.2 min; the mass fraction of the solvent in the coagulation bath is 51%; the mass concentration of the regulator is 0.02 mol / L.

[0149] b) Second stage solidification molding

[0150] The temperature of the coagulation bath is 55°C; the coagulation drawing ratio is 1.05 times; the circulation flow rate of the coagulation bath is 4 mm / s; the coagulation residence time is 0.3 min; the mass fraction of the solvent in the coagulation bath is 40%; the mass concentration of the regulator is 0.03 mol / L.

[0151] c) Three-stage solidification molding

[0152] The temperature of the coagulation bath is 68°C; the coagulation stretching ratio is 1.65 times; the circulation flow rate of the coagulation bath is 5 mm / s; the coagulation residence time is 2 min; the mass fraction of the solvent in the coagulation bath is 35%; and the mass concentration of the regulator in the coagulation bath is 0.01 mol / L.

[0153] d) Four-stage solidification molding

[0154] The temperature of the coagulation bath is 72°C; the coagulation stretching ratio is 1.4 times; the circulation flow rate of the coagulation bath is 5 mm / s; the coagulation residence time is 1.6 min; the mass fraction of the solvent in the coagulation bath is 18%; and the mass concentration of the regulator in the coagulation bath is 0.01 mol / L.

[0155] e) Five-stage solidification molding

[0156] The temperature of the coagulation bath is 76°C; the coagulation stretching ratio is 1.4 times; the circulation flow rate of the coagulation bath is 5 mm / s; the coagulation residence time is 1.5 min; the mass fraction of the solvent in the coagulation bath is 10%; and the mass concentration of the regulator in the coagulation bath is 0.01 mol / L.

[0157] f) Six-stage solidification molding

[0158] The temperature of the coagulation bath is 77°C; the coagulation drawing ratio is 1.3 times; the circulation flow rate of the coagulation bath is 8 mm / s; the coagulation residence time is 1.0 min; the mass fraction of the solvent in the coagulation bath is 1%; and the mass concentration of the regulator in the coagulation bath is 0.01 mol / L.

[0159] 3) Water washing process

[0160] Water washing temperature: 65℃; Water washing time: 2min.

[0161] 4) Hot water drawing process

[0162] The filament bundle is drawn in hot water at 80-95°C, and the drawing ratio is 2 times.

[0163] 5) Oiling process

[0164] Use self-prepared silicone oil, the oil concentration is 0.8%.

[0165] 6) Drying and densification process

[0166] Temperature gradient levels: 6 levels, the temperature gradients are: 98℃, 102℃, 108℃, 110℃, 115℃, 120℃; drying time for each level: 4s.

[0167] 7) Steam drawing and shrinkage heat setting process

[0168] Steam drawing process: using saturated water vapor as the medium, high-multiple drawing is applied to the tow, saturated water vapor pressure: 0.35MPa; drawing multiple: 3 times; steam drawing residence time: 3s.

[0169] Shrinkage heat setting process: superheated steam as the medium, superheated steam temperature: 150°C; draft ratio: 0.92 times; shrinkage heat setting residence time: 3s.

[0170] 8) Pre-oxidation process

[0171] Under hot air medium, a 5-stage pre-oxidation process is adopted. The pre-oxidation temperatures are: 198℃, 212℃, 228℃, 238℃, 258℃; the total pre-oxidation drafting multiple is 1.25 times; the total pre-oxidation residence time is 48min.

[0172] 9) Low temperature carbonization process

[0173] Under high-purity nitrogen protection conditions, low-temperature carbonization temperature: 300-900℃; low-temperature carbonization time: 4min.

[0174] 10) High temperature carbonization process

[0175] Under high-purity nitrogen protection conditions, 5 levels of high-temperature carbonization treatment. The temperatures of high-temperature carbonization treatment are: 900℃, 1100℃, 1400℃, 1800℃, 1350℃; high-temperature carbonization time: 5min.

[0176] 11) Graphitization process

[0177] Under the protection of high-purity argon, 4-stage high-temperature graphitization treatment was carried out. The high-temperature graphitization treatment temperatures were 2100℃, 2450℃, 2920℃, and 2700℃, respectively; the graphitization draft ratio was 1.05 times; and the graphitization treatment time was 4 minutes.

[0178] Example 2

[0179] This embodiment proposes an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, wherein the specific steps are as follows:

[0180] (I) Constructing prediction models (overall model and sub-models)

[0181] Same as Example 1.

[0182] (II) Output criteria of the overall model and sub-models

[0183] The first general model predicts the output of carbon fiber strength and modulus, and its output criteria (the target range of fiber performance in this embodiment) are: tensile strength is 4.0-5.0 GPa, tensile modulus is 570-640 GPa; diameter is 4.4-5.1 μm, and bulk density is 1.90-1.93 g / cm 3 .

[0184] The second general model: predicts the process parameters of the entire line, and the output criteria refer to the range of the process parameters of the entire fiber line in Example 1 (III).

[0185] The first sub-model predicts and outputs the cross-sectional roundness of the nascent fiber, and its output criterion condition is: the cross-sectional roundness of the nascent fiber is 0.75-0.85.

[0186] The second sub-model predicts and outputs the original silk strength and modulus, and its output criteria are: tensile strength is 780-1050MPa, and tensile modulus is 8-15Gpa.

[0187] The third sub-model predicts and outputs the skin thickness of the polyacrylonitrile precursor, and its output criterion condition is: the layer thickness of the skin polyacrylonitrile precursor is 180-220nm.

[0188] The fourth sub-model predicts the high temperature carbonization process, and its output criterion is: the maximum temperature of the high temperature carbonization process is ≤2050℃.

[0189] The fifth sub-model predicts the pre-oxidation process, and its output criteria are: the maximum temperature of the pre-oxidation process ≤ 270°C; the total drafting multiple of the pre-oxidation process ≤ 1.25; and the time of the pre-oxidation process ≤ 80 min.

[0190] The sixth sub-model predicts the solidification molding process, and its output criteria are: the temperature of the first solidification bath ≤ 62°C; the concentration of the first solidification bath ≤ 68%; the draft ratio in the first solidification bath ≤ 0.95.

[0191] (III) Full-line process parameter range (output criteria of the second overall model)

[0192] Same as Example 1.

[0193] (IV) Fiber test performance and verified optimal full-line process parameters

[0194] Taking carbon fiber strength and modulus as the target (the target range of fiber performance in this embodiment: strength 4.0-5.0Gpa, modulus 570-640Gpa), three rounds of bidirectional optimization model prediction and experimental verification (prediction method, see Figure 2 Finally, 20 sets of process parameters were output. After experimental verification, the fiber performance of 6 sets of full-line process parameters reached the target range.

[0195] The performance list of the six groups of fibers that achieved the target is shown in Table 2:

[0196] Table 2

[0197]

[0198] From Table 2, we can see that the comprehensive performance of the carbon fiber corresponding to number 1 is the best (both strength and modulus are taken into account). The full-line process parameters corresponding to number 1 are as follows:

[0199] 1) Preparation of spinning solution

[0200] The monomer acrylonitrile, the monomer itaconic acid, and the monomer 2-acrylamide-2-methylpropane sulfonic acid are added into a reactor with a stirring device in a ratio of 99.00:1:0. In the reaction solution, the mass percentage of the total monomer is 22%, the amount of the initiator azobisisobutyl cyanide is 1.0% of the total weight of the monomer, and dimethyl sulfoxide is used as the solvent to carry out solution polymerization. During the polymerization process, high-purity nitrogen is first bubbled into the reactor for 15 minutes to replace the air in the reaction system, and then stirring is started and the temperature is raised to 70°C. Under the protection of nitrogen, the polymerization is carried out at a constant temperature for 12 hours to obtain a ternary copolymerized polyacrylonitrile-based carbon fiber spinning solution.

[0201] 2) Polymerization liquid solidification molding

[0202] The wet spinning method is adopted, and the spinning solution is ejected from the spinneret and then solidified in four-stage coagulation baths in sequence. The composition of each stage of the coagulation bath is as follows: the solvent is dimethyl sulfoxide, the coagulant is water, and the regulator is ammonia water.

[0203] a) First stage solidification molding

[0204] Coagulation bath temperature: 53°C; coagulation draw ratio: 0.65 times; coagulation bath circulation flow rate: 14 mm / s; coagulation residence time: 0.2 min; solvent mass fraction in coagulation bath: 51%; mass concentration of regulator: 0.02 mol / L.

[0205] b) Second stage solidification molding

[0206] Coagulation bath temperature: 55°C; coagulation draw ratio: 1.05 times; coagulation bath circulation flow rate: 4 mm / s; coagulation residence time: 0.3 min; solvent mass fraction in coagulation bath: 40%; regulator mass concentration: 0.03 mol / L.

[0207] c) Three-stage solidification molding

[0208] The temperature of the coagulation bath is 68°C; the coagulation stretching ratio is 1.65 times; the circulation flow rate of the coagulation bath is 5 mm / s; the coagulation residence time is 2 min; the mass fraction of the solvent in the coagulation bath is 20%; the mass concentration of the regulator is 0.015 mol / L.

[0209] d) Four-stage solidification molding

[0210] The temperature of the coagulation bath is 72°C; the coagulation drawing ratio is 1.4 times; the circulation flow rate of the coagulation bath is 5 mm / s; the coagulation residence time is 1.0 min; the mass fraction of the solvent in the coagulation bath is 10%; the mass concentration of the regulator is 0.01 mol / L.

[0211] 3) Water washing process

[0212] Water washing temperature: 65℃; Water washing time: 2min.

[0213] 4) Hot water drawing process

[0214] The filament bundle is drawn in hot water at 80-95°C, and the drawing ratio is 2 times.

[0215] 5) Oiling process

[0216] Use self-prepared silicone oil, the oil concentration is 0.8%.

[0217] 6) Drying and densification process

[0218] Temperature gradient levels: 6 levels, the temperature gradients are: 98℃, 102℃, 108℃, 110℃, 115℃, 120℃; drying time for each level: 4s.

[0219] 7) Steam drawing and shrinkage heat setting process

[0220] Steam drawing process: using saturated water vapor as the medium, high-multiple drawing is applied to the tow, saturated water vapor pressure: 0.35MPa; drawing multiple: 3 times; steam drawing residence time: 3s.

[0221] Shrinkage heat setting process: superheated steam as the medium, superheated steam temperature: 150°C; draft ratio: 0.92 times; shrinkage heat setting residence time: 3s.

[0222] 8) Pre-oxidation process

[0223] Under hot air medium, a 5-stage pre-oxidation process is adopted. The pre-oxidation temperatures are: 198℃, 212℃, 228℃, 238℃, 265℃; the total pre-oxidation drafting multiple is 1.25 times; the total pre-oxidation residence time is 60min.

[0224] 9) Low temperature carbonization process

[0225] Under high-purity nitrogen protection conditions, low-temperature carbonization temperature: 300-900℃; low-temperature carbonization time: 4min.

[0226] 10) High temperature carbonization process

[0227] Under high-purity nitrogen protection conditions, 5-level high-temperature carbonization treatment. The temperatures of high-temperature carbonization treatment are: 900℃, 1100℃, 1500℃, 2100℃, 1550℃; high-temperature carbonization time: 3.5min.

[0228] 11) Graphitization process

[0229] Under the protection of high-purity argon, 4-stage high-temperature graphitization treatment was carried out. The high-temperature graphitization treatment temperatures were 2050℃, 2450℃, 2850℃, and 2700℃, respectively; the graphitization draft ratio was 1.1 times; and the graphitization treatment time was 4 minutes.

[0230] Example 3

[0231] This embodiment proposes an intelligent development method for a preparation process of polyacrylonitrile-based carbon fiber, wherein the specific steps are as follows:

[0232] (I) Constructing prediction models (overall model and sub-models)

[0233] Same as Example 1.

[0234] (II) Output criteria of the overall model and sub-models

[0235] The first general model predicts the output of carbon fiber strength and modulus, and its output criteria (the target range of fiber performance in this embodiment) are: tensile strength is 2.6-3.2Gpa, tensile modulus is 735-810Gpa; diameter is 4.2-4.4μm, and bulk density is 1.93-1.95g / cm 3 .

[0236] The second general model: predicts the process parameters of the entire line, and the output criteria refer to the range of the process parameters of the entire fiber line in Example 1 (III).

[0237] The first sub-model predicts and outputs the cross-sectional roundness of the nascent fiber, and its output criterion condition is: the cross-sectional roundness of the nascent fiber is 0.65-0.80.

[0238] The second sub-model predicts and outputs the original silk strength and modulus, and its output criteria are: tensile strength is 900-1350MPa, and tensile modulus is 11-16Gpa.

[0239] The third sub-model predicts and outputs the skin thickness of the polyacrylonitrile precursor, and its output criterion condition is: the layer thickness of the skin polyacrylonitrile precursor is 210-430nm.

[0240] The fourth sub-model predicts the high temperature carbonization process, and its output criterion is: the maximum temperature of the high temperature carbonization process is ≤1950℃.

[0241] The fifth sub-model predicts the pre-oxidation process, and its output criteria are: the maximum temperature of the pre-oxidation process ≤ 265°C; the total drafting multiple of the pre-oxidation process ≤ 1.25; and the time of the pre-oxidation process ≤ 70 min.

[0242] The sixth sub-model predicts the solidification molding process, and its output criteria are: the temperature of the first solidification bath ≤ 55°C; the concentration of the first solidification bath ≤ 65%; the draft ratio in the first solidification bath ≤ 0.85.

[0243] (III) Full-line process parameter range (output criteria of the second overall model)

[0244] Same as Example 1.

[0245] (IV) Fiber test performance and verified optimal full-line process parameters

[0246] Taking the strength and modulus of carbon fiber as the target (the target range of fiber performance in this embodiment: strength 2.6-3.2Gpa, modulus 735-810Gpa), a round of two-way optimization model prediction and experimental verification (prediction method, see Figure 2 Finally, 20 sets of process parameters were output. After experimental verification, the fiber performance of 2 sets of full-line process parameters reached the target range.

[0247] The test performance of the two groups of fibers that achieved the target is shown in Table 3:

[0248] Table 3

[0249]

[0250] From Table 3, we can see that the comprehensive performance of carbon fiber corresponding to number 2 is the best (both strength and modulus are taken into account). The full-line process parameters corresponding to number 2 are as follows:

[0251] 1) Preparation of spinning solution

[0252] The monomer acrylonitrile, the monomer itaconic acid, and the monomer 2-acrylamide-2-methylpropane sulfonic acid are added into a reactor with a stirring device in a ratio of 99.50:0.5:0. In the reaction solution, the mass percentage of the total monomer is 19.8%, the amount of the initiator azobisisobutyl cyanide is 0.8% of the total weight of the monomer, and dimethyl sulfoxide is used as the solvent for solution polymerization. During the polymerization process, high-purity nitrogen is first bubbled into the reactor for 15 minutes to replace the air in the reaction system, and then stirring is started and the temperature is raised to 70°C. Under the protection of nitrogen, the polymerization is carried out at a constant temperature for 10 hours to obtain a ternary copolymerized polyacrylonitrile-based carbon fiber spinning solution.

[0253] 2) Polymerization liquid solidification molding

[0254] The wet spinning method is adopted, and the spinning solution is ejected from the spinneret and then solidified and formed in three stages of solidification baths in sequence. The composition of the solidification bath solution of each stage is as follows: the solvent is dimethyl sulfoxide, the coagulant is water, and the regulator is ammonia water.

[0255] a) First stage solidification molding

[0256] Coagulation bath temperature: 52°C; coagulation draw ratio: 0.65 times; coagulation bath circulation flow rate: 13 mm / s; coagulation residence time: 0.25 min; solvent mass fraction in coagulation bath: 53%; mass concentration of regulator: 0.04 mol / L.

[0257] b) Second stage solidification molding

[0258] Coagulation bath temperature: 55°C; coagulation draw ratio: 1.05 times; coagulation bath circulation flow rate: 4 mm / s; coagulation residence time: 0.3 min; solvent mass fraction in coagulation bath: 40%; regulator mass concentration: 0.03 mol / L.

[0259] c) Three-stage solidification molding

[0260] The temperature of the coagulation bath is 70°C; the coagulation drawing ratio is 1.4 times; the circulation flow rate of the coagulation bath is 5 mm / s; the coagulation residence time is 1.0 min; the mass fraction of the solvent in the coagulation bath is 10%; the mass concentration of the regulator is 0.02 mol / L.

[0261] 3) Water washing process

[0262] Washing temperature: 66℃; Washing time: 2.5min.

[0263] 4) Hot water drawing process

[0264] The filament bundle is drawn in hot water at 80-95°C, with a drawing ratio of 2.5 times.

[0265] 5) Oiling process

[0266] Use self-prepared silicone oil, the oil concentration is 0.85%.

[0267] 6) Drying and densification process

[0268] Temperature gradient levels: 6 levels, the temperature gradients are: 98℃, 103℃, 109℃, 110℃, 116℃, 122℃; drying time for each level: 5s.

[0269] 7) Steam drawing and shrinkage heat setting process

[0270] Steam drawing process: using saturated water vapor as the medium, high-multiple drawing is applied to the tow, saturated water vapor pressure: 0.35MPa; drawing multiple: 3.5 times; steam drawing residence time: 3s.

[0271] Shrinkage heat setting process: superheated steam as the medium, superheated steam temperature: 150°C; draft ratio: 0.95 times; shrinkage heat setting residence time: 3s.

[0272] 8) Pre-oxidation process

[0273] Under hot air medium, a 5-stage pre-oxidation process is adopted. The pre-oxidation temperatures are: 198℃, 213℃, 228℃, 240℃, 265℃; the total pre-oxidation drafting multiple is 1.25 times; the total pre-oxidation residence time is 62min.

[0274] 9) Low temperature carbonization process

[0275] Under high-purity nitrogen protection conditions, low-temperature carbonization temperature: 300-900℃; low-temperature carbonization time: 4min.

[0276] 10) High temperature carbonization process

[0277] Under high-purity nitrogen protection conditions, 5-stage high-temperature carbonization treatment. The temperatures of high-temperature carbonization treatment are: 950℃, 1150℃, 1650℃, 1950℃, 1380℃; high-temperature carbonization time: 3.6min.

[0278] 11) Graphitization process

[0279] Under the protection of high-purity argon, 4-stage high-temperature graphitization treatment was carried out. The high-temperature graphitization treatment temperatures were 2050℃, 2450℃, 3000℃, and 2750℃, respectively; the graphitization draft ratio was 1.1 times; and the graphitization treatment time was 4.0 min.

[0280] Comparative Example 1

[0281] Comparative Example 1 proposes a method for developing a preparation process for high-strength and high-modulus polyacrylonitrile-based carbon fiber, wherein the specific steps are as follows:

[0282] (I) Constructing prediction models (overall model and sub-models)

[0283] Compared with Example 1, the second overall model and the second type of sub-model are removed, and the rest are the same as Example 1.

[0284] (II) Output criteria of the overall model and sub-models

[0285] Compared with Example 1, the second overall model and the second type of sub-model are removed, and the rest are the same as Example 1.

[0286] (III) Setting the parameter range of each process point in the full-line process parameters

[0287] Same as Example 1.

[0288] (IV) Fiber test performance and verified optimal full-line process parameters

[0289] With carbon fiber strength and modulus as the target (target range of fiber performance in comparative example 1: tensile strength 3.2-3.8 GPa, tensile modulus 665-730 GPa), process search is performed based on the predicted outputs of the first overall model and the first sub-model.

[0290] After 28 rounds of prediction-test-correction model, 20 sets of process parameters were finally output. After experimental verification, the fiber performance of one set of full-line process parameters reached the target range.

[0291] The test performance of the two groups of fibers that achieved the target is shown in Table 4:

[0292] Table 4

[0293]

[0294] The full-line process parameters corresponding to No. 1 are basically the same as the process parameters corresponding to No. 3 in Example 1. Compared with Example 1, Comparative Example 1 lacks the second general model and the second type of sub-model. The second general model and the second type of sub-model output the full-line and local process parameters based on the target performance. After the lack of them, the process search space range is increased, and the trial and error probability of process verification is greatly improved, resulting in a significant increase in the target process development cycle and a substantial increase in investment consumption (Comparative Example 1 undergoes 28 rounds while Example 1 only undergoes 4 rounds).

[0295] Comparative Example 2

[0296] Comparative Example 2 proposes a method for developing a preparation process for high-strength and high-modulus polyacrylonitrile-based carbon fiber, wherein the specific steps are as follows:

[0297] (I) Constructing prediction models (overall model and sub-models)

[0298] Compared with Example 1, the first overall model and the first type of sub-model are removed, and the rest are the same as Example 1.

[0299] (II) Output criteria of the overall model and sub-models

[0300] Compared with Example 1, the first overall model and the first type of sub-model are removed, and the rest are the same as Example 1.

[0301] (III) Setting the parameter range of each process point in the full-line process parameters

[0302] Same as Example 1.

[0303] (IV) Fiber test performance and verified optimal full-line process parameters

[0304] With carbon fiber strength and modulus as the target (target range of fiber performance in comparative example 2: tensile strength 3.2-3.8 Gpa, tensile modulus 665-730 Gpa), process search is performed based on the predicted output of the second overall model and the second sub-model.

[0305] After 36 rounds of prediction-test-correction model, 20 sets of process parameters were finally output. After experimental verification, the fiber performance of all process parameter groups was still far from the target performance.

[0306] Compared with Example 1, Comparative Example 2 removes the first general model and the first type of sub-model. Because the global process flow is too long, there are too many process parameters, and the search volume of the cyclic iterative process search process is too huge, which consumes computing resources; and because the prediction accuracy of the second general model is low, the probability of trial and error in experimental verification is greatly increased. In short, while the computing resources and experimental input are greatly increased compared with Example 1, Comparative Example 2 still does not have a good process development result.

[0307] Comparative Example 3

[0308] Comparative Example 3 proposes a method for developing a preparation process for high-strength and high-modulus polyacrylonitrile-based carbon fiber, wherein the details are as follows:

[0309] (I) Constructing prediction models (overall model and sub-models)

[0310] Compared with Example 1, the third sub-model is removed.

[0311] The overall model and sub-model construction process are the same as in Example 1.

[0312] (II) Output criteria of the overall model and sub-models

[0313] Compared with Example 1, the third sub-model is removed, and the rest is the same as Example 1.

[0314] (III) Setting the parameter range of each process point in the full-line process parameters

[0315] Same as Example 1.

[0316] (IV) Fiber test performance and verified optimal full-line process parameters

[0317] With carbon fiber strength and modulus as the target (target range of fiber performance in comparative example 3: tensile strength 3.2-3.8 GPa, tensile modulus 665-730 GPa), process search is performed based on the predicted output of the overall model and sub-model.

[0318] After 30 rounds of prediction-experiment-correction modeling, 20 sets of process parameters were finally outputted, and experimental verification showed that the fiber performance of one set of full-line process parameters was close to the target range (but not to the current range).

[0319] Compared with Example 1, after the number of the first type of sub-models is reduced (the third sub-model is removed), the restrictions on the core performance indicators of the fiber are reduced. Therefore, the space for process screening becomes larger, the demand for computing resources becomes larger, and the probability of process verification trial and error is significantly increased, and the development cycle and investment are significantly increased.

[0320] Comparative Example 4

[0321] Comparative Example 3 proposes a method for developing a preparation process for high-strength and high-modulus polyacrylonitrile-based carbon fiber, wherein the details are as follows:

[0322] (I) Constructing prediction models (overall model and sub-models)

[0323] Compared with Example 1, the sixth sub-model is removed.

[0324] The overall model and sub-model construction process are the same as in Example 1.

[0325] (II) Output criteria of the overall model and sub-models

[0326] Compared with Example 1, the sixth sub-model is removed, and the rest is the same as Example 1.

[0327] (III) Setting the parameter range of each process point in the full-line process parameters

[0328] Same as Example 1.

[0329] (IV) Fiber test performance and verified optimal full-line process parameters

[0330] With carbon fiber strength and modulus as the target (target range of fiber performance in comparative example 4: tensile strength 3.2-3.8 GPa, tensile modulus 665-730 GPa), process search is performed based on the prediction output of the overall model and sub-model.

[0331] After 40 rounds of prediction-test-correction model, 20 sets of process parameters were finally output. After experimental verification, the fiber performance of all process parameter groups was still far from the target performance.

[0332] Compared with Example 1, after the number of the second type of sub-models is reduced (the sixth sub-model is removed), the output restrictions of the key process points in the process group parameters are reduced, resulting in a larger space for process screening, a larger demand for computing resources, and a significantly increased probability of process verification trial and error, and a significantly increased development cycle and investment. Compared with Example 1, while the computing resources and experimental investment are greatly increased, Comparative Example 4 still does not have a good process development result.

[0333] Comparative Example 5

[0334] Comparative Example 5 proposes a method for developing a preparation process for high-strength and high-modulus polyacrylonitrile-based carbon fiber, wherein the specific steps are as follows:

[0335] (I) Constructing prediction models (overall model and sub-models)

[0336] Compared with Example 1, the third sub-model is different, and the rest are the same as Example 1.

[0337] The third sub-model: construct a BP neural network model (the third sub-model) of spinning process parameters → raw yarn orientation according to the data set and train it.

[0338] (II) Output criteria of the overall model and sub-models

[0339] Compared with Example 1, the output criterion of the third sub-model is: 75-90%;

[0340] The rest is the same as in Example 1.

[0341] (III) Setting the parameter range of each process point in the full-line process parameters

[0342] Same as Example 1.

[0343] (IV) Fiber test performance and verified optimal full-line process parameters

[0344] With carbon fiber strength and modulus as the target (target range of fiber performance in comparative example 5: tensile strength 3.2-3.8 GPa, tensile modulus 665-730 GPa), process search is performed based on the predicted output of the overall model and sub-model.

[0345] After 30 rounds of prediction-test-correction model, 20 sets of process parameters were finally output. After experimental verification, 2 sets of full-line process parameters had fiber strength reaching the target, 3 sets of full-line process parameters had modulus reaching the target, and neither strength nor modulus reached the target at the same time.

[0346] Compared with Example 1, the content of the third sub-model in Comparative Example 5 is different. Regarding the high-modulus performance characteristics, the parameters output by the sub-model and their correlation have priority. Selecting a sub-model with a high priority can speed up process development. After the content of the third sub-model in Comparative Example 5 is changed, its priority is obviously reduced from the development process, resulting in an increase in the development cycle for achieving both high-strength and high-modulus performance.

[0347] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber, It is characterized in that The intelligent development method comprises the following steps: Constructing a prediction model step: based on the process data set of polyacrylonitrile-based carbon fiber, constructing a first general model, a first type of sub-model, a second general model and a second type of sub-model; wherein, the first general model is a machine learning algorithm model that takes the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and the performance data of polyacrylonitrile-based carbon fiber as output; the first type of sub-model is a machine learning algorithm model that takes the local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and the fiber performance data and / or structure data after local process treatment as output; the second general model is a machine learning algorithm model that takes the performance data of polyacrylonitrile-based carbon fiber as input and the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output; the second type of sub-model is a machine learning algorithm model that takes the performance data of polyacrylonitrile-based carbon fiber as input and the local process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output; The step of setting the criterion conditions: setting the target performance index of the required polyacrylonitrile-based carbon fiber; setting the parameter range of each process point in the full-line process parameters; setting the output criterion conditions of the first general model, the second general model, the first sub-model, and the second sub-model; Process search step: according to the set target performance index of polyacrylonitrile-based carbon fiber, first run the second general model to generate a set of full-line process parameters; if the generated full-line process parameters meet the output criterion conditions of the second general model, run the second type of sub-model, otherwise, re-run the second general model in a loop iteration to generate a set of full-line process parameters; If the outputs of the second-type sub-models all meet the output criterion conditions of the second-type sub-models, the outputs of the second-type sub-models will replace the corresponding contents of the output process parameters of the second overall model, and a set of full-line process parameters will be re-formed. Next, the first-type sub-models will be run. Otherwise, the second overall model will be re-run in a loop iteration to generate a set of full-line process parameters. If the outputs of the first type of sub-models all meet the output criterion conditions of the first type of sub-models, the first overall model is run, otherwise the second overall model is re-run in a loop iteration to generate a set of full-line process parameters; If the output of the first overall model meets the output criterion conditions of the first overall model, the corresponding full-line process parameters are recorded as the screened full-line process parameters; if the output of the first overall model does not meet the output criterion conditions of the first overall model, the second overall model is re-run in an iterative loop to generate a set of full-line process parameters.

2. The intelligent development method of the preparation process of polyacrylonitrile-based carbon fiber according to claim 1, It is characterized in that After the process search step, the method further includes: Test verification step: test and verify the selected process parameters of the entire line; Wherein, if the performance of the prepared polyacrylonitrile-based carbon fiber reaches the target performance index, the process development is completed; Among them, if the performance of the prepared polyacrylonitrile-based carbon fiber does not meet the target performance indicators, the test process data is added to the process data set of the polyacrylonitrile-based carbon fiber, the first general model, the first type of sub-model, the second general model and the second type of sub-model are retrained, and the process search step and the test verification step are re-performed.

3. The intelligent development method of the preparation process of polyacrylonitrile-based carbon fiber according to claim 1 or 2, It is characterized in that The process data set of the polyacrylonitrile-based carbon fiber includes: full-line process parameter data in the preparation process of the polyacrylonitrile-based carbon fiber, fiber structure data, fiber performance data and fiber single filament diameter; Preferably, the full-line process parameter data include: polymerization process parameters, spinning process parameters, pre-oxidation process parameters, low-temperature carbonization process parameters and high-temperature carbonization process parameters; preferably, the full-line process parameters also include graphitization process parameters; Further preferably, the polymerization reaction process parameters include: polymerization reaction formula, temperature and time; Further preferably, the spinning process parameters include: coagulation molding process parameters, water washing process parameters, hot water drawing process parameters, oiling concentration, drying densification process parameters, steam drawing process parameters, shrinkage heat setting process parameters; Further preferably, the pre-oxidation process parameters include pre-oxidation temperature parameters, pre-oxidation time parameters and draft ratio; Further preferably, the low temperature carbonization process parameters include a temperature parameter of low temperature carbonization and a time parameter of low temperature carbonization; Further preferably, the high temperature carbonization process parameters include a temperature parameter of the high temperature carbonization and a time parameter of the high temperature carbonization; Preferably, the fiber performance data includes performance data of polyacrylonitrile precursor and performance data of polyacrylonitrile-based carbon fiber; Preferably, the structural data of the fiber include: the cross-sectional roundness of the primary fiber and the skin thickness of the skin-core structure of the polyacrylonitrile precursor.

4. The intelligent development method of the preparation process of polyacrylonitrile-based carbon fiber according to any one of claims 1 to 3, It is characterized in that In the step of building a prediction model: The machine learning algorithm model is a BP neural network model; and / or According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as input and the performance data of polyacrylonitrile-based carbon fiber as output, and training is performed to obtain a first overall model; and / or According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with the performance data of polyacrylonitrile-based carbon fiber as input and the full-line process parameters in the preparation process of polyacrylonitrile-based carbon fiber as output, and training is performed to obtain a second overall model; and / or The performance data of the polyacrylonitrile-based carbon fiber include: tensile strength, tensile modulus; preferably, the performance data of the polyacrylonitrile-based carbon fiber also include: one or more of the CV value of tensile strength, CV value of tensile modulus, elongation at break, CV value of elongation at break, bulk density, single filament diameter, carbon content, and ash content.

5. An intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber according to any one of claims 1 to 4, It is characterized in that The first type of sub-models includes: a first sub-model, a second sub-model, and a third sub-model; wherein, According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with polymerization reaction process parameters and coagulation molding process parameters as input and cross-sectional roundness of primary fiber as output, and training is performed to obtain the first sub-model; According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with polymerization reaction process parameters and spinning process parameters as input and performance data of polyacrylonitrile precursor as output, and training is performed to obtain a second sub-model; preferably, the performance data of the polyacrylonitrile precursor includes tensile strength data and tensile modulus data of the polyacrylonitrile precursor; According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning algorithm model is constructed with the spinning process parameters as input and the skin thickness data of the skin-core structure of polyacrylonitrile precursor as output, and training is performed to obtain the third sub-model; Preferably, in the step of setting the criterion condition, the output criterion condition of the first sub-model is set as follows: the cross-sectional roundness of the primary fiber is in a set range, preferably 0.65-0.95; wherein the cross-sectional roundness of the primary fiber is the ratio of the minor axis to the major axis of the cross-sectional area of ​​the primary fiber; Preferably, in the step of setting the criterion condition, the output criterion condition of the second sub-model is set as follows: the tensile strength of the polyacrylonitrile precursor is within the set strength range, preferably 750-1350 MPa; the tensile modulus of the polyacrylonitrile precursor is within the set modulus range, preferably 7.5-16 Gpa; Preferably, in the step of setting the criterion condition, the output criterion condition of the third sub-model is set as follows: the cortical thickness of the polyacrylonitrile precursor is a set thickness, preferably 120-430 nm.

6. An intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber according to any one of claims 1 to 5, It is characterized in that The second type of sub-models includes: a fourth sub-model, a fifth sub-model, and a sixth sub-model; wherein, According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with performance parameters of polyacrylonitrile-based carbon fiber as input and high-temperature carbonization process parameters as output is constructed and trained to obtain a fourth sub-model; preferably, the performance parameters of the polyacrylonitrile-based carbon fiber include tensile strength, tensile modulus and carbon content; According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the pre-oxidation process parameters as output is constructed, and training is performed to obtain a fifth sub-model; preferably, the performance parameters of the polyacrylonitrile-based carbon fiber include tensile strength, tensile strength CV value, tensile modulus, tensile modulus CV value, and bulk density; According to the process data set of polyacrylonitrile-based carbon fiber, a machine learning model with the performance parameters of polyacrylonitrile-based carbon fiber as input and the solidification molding process parameters as output is constructed, and training is performed to obtain a sixth sub-model; preferably, the performance parameters of the polyacrylonitrile-based carbon fiber include tensile strength, tensile strength CV value, tensile modulus, tensile modulus CV value, and bulk density; Preferably, in the step of setting the criterion condition, the output criterion condition of the fourth sub-model is set as follows: the highest temperature of the high-temperature carbonization process is ≤ 2100°C; Preferably, in the step of setting the criterion conditions, the output criterion conditions of the fifth sub-model are set as follows: the maximum temperature of the pre-oxidation process is ≤275°C; the total draft ratio of the pre-oxidation process is ≤1.3; the time of the pre-oxidation process is ≤80min; Preferably, in the step of setting the criterion conditions, the output criterion conditions of the sixth sub-model are set as follows: the temperature of the first coagulation bath ≤ 62°C; the concentration of the first coagulation bath ≤ 68%; and the draft ratio in the first coagulation bath ≤ 0.

95.

7. An intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber according to any one of claims 1 to 6, It is characterized in that In the step of setting the criterion conditions: The output criterion condition of the second overall model is: making the parameter range of each process point in the whole process parameters within the set effective range; and / or The criterion condition of the first general model is: the deviation between the performance data of the polyacrylonitrile-based carbon fiber output by the first general model and the target performance index is ≤5%.

8. An intelligent development method for the preparation process of polyacrylonitrile-based carbon fiber according to any one of claims 1 to 7, It is characterized in that In the process search step: when the number of groups of screened process parameters for the entire line is less than the set number of groups, the second overall model is re-run in an iterative loop to generate a set of process parameters for the entire line, and the process search step is performed again until the number of groups of screened process parameters for the entire line reaches the set number of groups; preferably, the set number of groups is 20 groups.

9. The intelligent development method of the preparation process of polyacrylonitrile-based carbon fiber according to any one of claims 2 to 8, It is characterized in that During the test verification step: Conducting test verification on one or more selected groups of full-line process parameters. If the performance of polyacrylonitrile-based carbon fibers prepared by one or more selected groups of full-line process parameters reaches the target value, the process development is completed. and / or Conducting test verification on one or more groups of full-line process parameters that have been screened out. If the performance of the prepared polyacrylonitrile-based carbon fibers does not reach the target value, the test verification data needs to be added to the process data set of the polyacrylonitrile-based carbon fibers, and the first general model, the second general model, the first sub-model, and the second sub-model need to be retrained and corrected. Then, the second overall model is re-run in an iterative cycle to generate a set of full-line process parameters, and the process search step and the test verification step are repeated.

10. A polyacrylonitrile-based carbon fiber, It is characterized in that The polyacrylonitrile-based carbon fiber is prepared by a process developed by the intelligent development method for the preparation process of the polyacrylonitrile-based carbon fiber according to any one of claims 1 to 9; Preferably, the tensile strength of the polyacrylonitrile-based carbon fiber is 2.6-5.1 GPa, the tensile modulus is 560-810 GPa, the diameter is 4.2-5.2 μm, and the bulk density is 1.88-1.95 g / cm 3 ; Preferably, the polyacrylonitrile-based carbon fiber is characterized by small-angle X-ray scattering, when the scattering vector When the scattering intensity I(q)∝q -α , α takes a value of 3.25-3.7; when the scattering vector When the scattering intensity I(q)∝q -α , α takes a value of 1.5-2.0.

Images