Intelligent analysis and self-adaptive regulation and control system for glass fiber wiredrawing forming process

Abstract

The invention discloses an intelligent analysis and self-adaptive regulation and control system for a glass fiber wiredrawing forming process, which relates to the technical field of high silica glass fiber product processing and comprises a multi-modal data acquisition unit, a process prediction model unit, a reference process model unit, a self-adaptive regulation and control decision unit and a control execution unit. By introducing the prediction model and the self-adaptive decision-making unit, two strong coupling parameters of speed and temperature can be decoupled, collaborative optimization regulation and control are carried out, precise control over the wire drawing process is achieved, the uniformity of the fiber diameter is remarkably improved, the future process trend can be predicted on the basis of real-time data, and the method has good application prospects. Traditional feedback control is improved into predictive feedforward control, fine adjustment is carried out before process parameters are obviously deviated, the rate of finished products is greatly improved, meanwhile, the optimal production process is solidified into a reference model, complex regulation and control logic is precipitated into a decision algorithm, and dependence on artificial experience is eliminated.

Description

Technical Field

[0001] This invention relates to the field of high-silica glass fiber product processing technology, and in particular to an intelligent analysis and adaptive control system for the glass fiber drawing and forming process. Background Technology

[0002] High-silica glass fiber plays an irreplaceable role in aerospace, defense, and high-end environmental protection fields due to its excellent high-temperature resistance, dielectric properties, and ablation resistance. The fiber drawing process is a key step in its production, determining the uniformity of fiber diameter, mechanical properties, and yield. This process involves drawing glass droplets through a baffle plate at high temperatures (usually above 1700°C) to form micron-sized continuous fibers.

[0003] Currently, the main technical challenges in controlling the high-silica glass fiber drawing process are as follows: I. Strong Coupling and Time Lag of Parameters: Drawing temperature and drawing speed are two core process parameters. There is a strong nonlinear coupling relationship between them. At the same time, temperature regulation has significant inertia and time lag. When a deviation in fiber diameter is detected, the root cause of the fault may have occurred several minutes ago, resulting in a lag in control response and making it difficult to achieve real-time correction.

[0004] Second, reliance on manual experience: Existing control methods heavily depend on the experience of senior operators. Operators indirectly judge process stability by visually observing the shape, brightness, and fiber state of the glass droplets and manually adjust temperature and speed setpoints. This method is highly subjective, with differences between operators, resulting in poor product quality consistency and making it difficult to pass on knowledge and scale up.

[0005] Third, lack of predictive maintenance: Traditional control systems are mostly based on PID (proportional-integral-derivative) feedback control, which is a reactive adjustment. The response is only triggered when the fiber breaks or the diameter is seriously out of tolerance. It cannot predict and avoid potential process fluctuations, resulting in low yield and high energy consumption.

[0006] IV. Characteristics of high-silica materials: High-silica glass is extremely sensitive to viscosity changes at high temperatures. Its forming process window is narrower than that of ordinary glass fiber, and it requires extremely high precision in temperature and speed control. Traditional control methods are difficult to meet its stringent production requirements.

[0007] Therefore, there is an urgent need to develop a system that can overcome the above-mentioned defects and realize intelligent analysis, advanced prediction and adaptive control of the high silica glass fiber drawing process, so as to solve the problem of precise and coordinated control of drawing speed and drawing temperature. Summary of the Invention

[0008] The purpose of this invention is to decouple the two strongly coupled parameters of speed and temperature by introducing a predictive model and an adaptive decision-making unit, and to achieve precise control of the fiber drawing process through coordinated optimization and regulation. This significantly improves the uniformity of fiber diameter and enables the prediction of future process trends based on real-time data. It upgrades traditional feedback control to predictive feedforward control, allowing for fine-tuning before significant deviations in process parameters occur. This effectively avoids fiber breakage and waste, greatly improving the yield. At the same time, it solidifies the optimal production process into a benchmark model and condenses complex control logic into a decision-making algorithm, eliminating reliance on human experience and ensuring a high degree of consistency in product quality across different shifts and production lines.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent analysis and adaptive control system for glass fiber drawing and forming process, comprising a multimodal data acquisition unit, a process prediction model unit, a baseline process model unit, an adaptive control decision unit, and a control execution unit; The multimodal data acquisition unit is used to collect multi-dimensional data in real time during the wire drawing process. The multi-dimensional data includes wire drawing temperature, wire drawing speed, fiber diameter, wire drawing tension, and glass droplet morphology characteristics. The glass droplet morphology characteristics include droplet height, maximum diameter, and cone angle. The multi-dimensional data is timestamped and synchronized in time, and then integrated into a high-dimensional real-time data stream and sent to the process prediction model unit. The process prediction model unit is used to acquire and process multi-dimensional data. Based on the pre-trained process prediction model, it predicts key process indicators within a preset period and generates a prediction state vector to output to the adaptive control decision unit. The prediction state vector includes the predicted fiber diameter, the predicted drawing tension, and the predicted fiber breakage probability. The benchmark process model unit is used to obtain the ideal mapping relationship between process parameters and state parameters under optimal production conditions and to form a benchmark state model. When the target fiber diameter setting value is received, the unit outputs the ideal state vector corresponding to the setting value to the adaptive control decision unit. The ideal state vector includes the ideal drawing tension and the ideal glass droplet morphology characteristics. The adaptive control decision unit is used to simultaneously receive the predicted state vector and the ideal state vector, calculate the deviation vector between the two, and based on the deviation vector, obtain the optimal control strategy vector required to minimize the deviation vector through a preset strategy optimization model, and output it to the control execution unit. The control strategy vector includes the target adjustment amount of drawing speed and the target adjustment amount of drawing temperature. The control execution unit is used to receive the regulation strategy vector and convert it into real-time control commands, which are then sent to the frequency converter of the wire drawing machine and the power controller of the heating furnace to precisely adjust the actual wire drawing speed and wire drawing temperature.

[0010] Furthermore, the following sensors are deployed within the distribution area of ​​the wire drawing and forming equipment: laser diameter gauge, industrial camera, infrared thermometer, tension sensor, and encoder, among which: Laser diameter gauges are used to measure the diameter of newly formed fibers in real time without damage. Industrial cameras are used to focus on the glass droplets below the platinum stencil and extract the droplet's height, maximum diameter, and cone angle in real time. Infrared thermometers are used for non-contact measurement of the surface temperature of glass droplets, serving as direct feedback on the drawing temperature. Tension sensors are installed on the fiber path to monitor tension changes in real time during the fiber drawing process; The encoder is used to integrate with the winding head motor of the wire drawing machine to accurately provide feedback on the real-time wire drawing speed.

[0011] Furthermore, the specific process for generating the predicted state vector is as follows: S101. Obtain multi-dimensional data and integrate it into a key data vector. Where T(t) is the drawing temperature, V(t) is the drawing speed, D(t) is the real-time fiber diameter, F(t) is the drawing tension, H(t) is the droplet height, Dmax(t) is the maximum diameter, and θ(t) is the cone angle; S102. Obtain historical processing data of multiple sets of glass fiber drawing and forming process, construct time series samples, and use the normalized data as training samples. Construct a process prediction model based on long short-term memory network, download the weight file and load it onto the corresponding network to initialize the transfer network parameters. S103. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the predicted state vector, initialize the weights of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters. Retrain the entire network to obtain the process prediction model. S104. During the training process, small batches of samples are randomly and non-repeatedly drawn from the training set for training. One training cycle is completed after all training samples are drawn. The training is completed after a certain number of cycles to obtain the basic process prediction model. S105. Input the key data vector into the basic process prediction model to obtain the basic process prediction model.

[0012] Furthermore, the specific process for generating the ideal state vector is as follows: S201. Obtain the process parameters and state parameters under optimal production conditions as target data. The selection criteria for the target data are as follows: Under the condition of highest product quality: the process parameters corresponding to the minimum standard deviation of fiber diameter and the optimal mechanical properties; Under the most stable production conditions: the process parameters corresponding to the lowest wire breakage rate and the smallest fluctuation in wire drawing tension; Under the condition of highest production efficiency: the process parameters corresponding to the fastest wire drawing speed while ensuring quality. Under the condition of minimum energy consumption: the process parameters corresponding to the minimum energy consumption per unit output; S202. The selected target data is time-aligned and cleaned to remove outliers, resulting in optimized target data used as training samples. The training samples are divided into training and validation sets. Based on a nonlinear regression model, the target fiber diameter and ideal state parameters are combined, and their functional relationship is shown below: ; S203. Verify the baseline state model using the validation set until the validation result shows that the difference between the output ideal state and the actual optimal production state reaches the preset difference value. S204. Input the target fiber diameter setting value into the reference state model to obtain the ideal state vector.

[0013] Furthermore, the specific process for generating the optimal control strategy vector is as follows: S301. Construct a simulation environment for the fiber drawing process based on digital twins, which is used to input control actions, specifically the fiber drawing speed adjustment amount ΔV and the fiber drawing temperature adjustment amount ΔT. Based on the mechanism model calibrated according to fluid mechanics, heat transfer and historical data, accurately simulate and output the fiber drawing process state data at the next moment. The fiber drawing process state data includes fiber diameter, fiber drawing tension and droplet morphology. S302. Construct a Markov decision process model to formalize the adaptive control problem of the wire drawing process into a Markov decision process; S303. A policy optimization model is constructed based on the dual-delay deep deterministic policy gradient algorithm. The training process is as follows: In the simulation environment, the agent outputs action A(t) through its actor network based on the current state S(t). After the simulation environment executes this action, it feeds back the state S(t+1) and reward R(t) for the next moment. The agent stores experience tuples (S(t), A(t), R(t), S(t+1)) into the experience replay pool and samples them to update its actor network and critic network. This process is iterated repeatedly until the agent's policy converges and can stably obtain high cumulative rewards, thus obtaining a policy optimization model. S304. Taking the real-time state vector S(t) as input, the actor network performs a forward propagation calculation, and its output is the optimal action A(t) that maximizes the long-term cumulative reward under the current state S(t). This action vector is the required optimal control strategy vector, as specifically described below: .

[0014] Furthermore, the core elements for constructing a Markov decision process model are defined as follows: The intelligent agent is the deep neural network model embedded in the adaptive control decision-making unit. State S(t) is defined as the set of information that comprehensively describes the current working condition and future trends, which the agent relies on when making decisions. Its specific composition is as follows: Deviation vector component: The deviation calculated from the predicted state vector and the ideal state vector; Current state component: Real-time data acquired by the multimodal data acquisition unit at the current moment; Predicted state components: The predicted state vector output by the process prediction model unit, i.e., [predicted fiber diameter, predicted drawing tension, predicted fiber breakage probability].

[0015] By combining the above components, state S(t) provides the agent with a complete basis for decision-making regarding the past, present, and future as follows: Where ΔV is the target adjustment amount of the drawing speed and ΔT is the target adjustment amount of the drawing temperature; The reward function R(t) maximizes the long-term cumulative reward by enabling the agent to perform actions, and is composed of the following weighted components: Bias penalty term: A penalty is applied based on the L2 norm of the bias vector. The larger the bias, the larger the negative reward value. This is used to drive the agent to learn how to minimize the difference between the predicted state and the ideal state, and is described as follows: ; Smoothness control penalty: This penalty applies to the magnitude of action A(t) to encourage the agent to produce smoother and more stable control commands. It is expressed as follows: ; Risk aversion penalty term: A penalty is imposed on the predicted wire breakage probability in the predicted state vector. When the predicted risk of wire breakage increases, a significant negative reward is given, as described below: ; Stable operation reward: If the system does not enter a high-risk state in each control cycle, a small positive reward value is given to encourage the agent to maintain the long-term stable operation of the system.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This intelligent analysis and adaptive control system for the glass fiber drawing process, by introducing a predictive model and an adaptive decision-making unit, decouples the two strongly coupled parameters of speed and temperature and performs collaborative optimization control. This achieves precise control of the drawing process, significantly improves the uniformity of fiber diameter, and can predict future process trends based on real-time data. It upgrades traditional feedback control to predictive feedforward control, making fine adjustments before significant deviations in process parameters occur, effectively avoiding fiber breakage and waste, and greatly improving the yield. At the same time, it solidifies the optimal production process into a benchmark model and condenses complex control logic into a decision-making algorithm, eliminating reliance on human experience and ensuring a high degree of consistency in product quality across different shifts and production lines. Attached Figure Description

[0017] Figure 1 A schematic diagram of the overall structure of the present invention is shown. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0019] like Figure 1 As shown, an intelligent analysis and adaptive control system for glass fiber drawing and forming process includes a multimodal data acquisition unit, a process prediction model unit, a baseline process model unit, an adaptive control decision unit, and a control execution unit. The multimodal data acquisition unit is used to collect multi-dimensional data in real time during the wire drawing process. The multi-dimensional data includes wire drawing temperature, wire drawing speed, fiber diameter, wire drawing tension, and glass droplet morphology characteristics. The glass droplet morphology characteristics include droplet height, maximum diameter, and cone angle. The multi-dimensional data is timestamped and synchronized in time, and then integrated into a high-dimensional real-time data stream and sent to the process prediction model unit. The following sensors are deployed within the distribution area of ​​the wire drawing and forming equipment, including laser diameter gauges, industrial cameras, infrared thermometers, tension sensors, and encoders, among which: Laser diameter gauges are used to measure the diameter of newly formed fibers in real time without damage. Industrial cameras are used to focus on the glass droplets below the platinum stencil and extract the droplet's height, maximum diameter, and cone angle in real time. Infrared thermometers are used for non-contact measurement of the surface temperature of glass droplets, serving as direct feedback on the drawing temperature. Tension sensors are installed on the fiber path to monitor tension changes in real time during the fiber drawing process; The encoder is used to integrate with the winding head motor of the wire drawing machine to accurately provide feedback on the real-time wire drawing speed.

[0020] The process prediction model unit is used to acquire and process multi-dimensional data. Based on the pre-trained process prediction model, it predicts key process indicators within a preset period and generates a prediction state vector to output to the adaptive control decision unit. The prediction state vector includes the predicted fiber diameter, the predicted drawing tension, and the predicted fiber breakage probability. The specific process of generating the predicted state vector is as follows: S101. Obtain multi-dimensional data and integrate it into a key data vector. Where T(t) is the drawing temperature, V(t) is the drawing speed, D(t) is the real-time fiber diameter, F(t) is the drawing tension, H(t) is the droplet height, Dmax(t) is the maximum diameter, and θ(t) is the cone angle; S102. Obtain historical processing data of multiple sets of glass fiber drawing and forming process, construct time series samples, and use the normalized data as training samples. Construct a process prediction model based on long short-term memory network, download the weight file and load it onto the corresponding network to initialize the transfer network parameters. S103. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the predicted state vector, initialize the weights of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters. Retrain the entire network to obtain the process prediction model. S104. During the training process, small batches of samples are randomly and non-repeatedly drawn from the training set for training. One training cycle is completed after all training samples are drawn. The training is completed after a certain number of cycles to obtain the basic process prediction model. S105. Input the key data vector into the basic process prediction model to obtain the basic process prediction model.

[0021] The benchmark process model unit is used to obtain the ideal mapping relationship between process parameters and state parameters under optimal production conditions and to form a benchmark state model. When the target fiber diameter setting value is received, the unit outputs the ideal state vector corresponding to the setting value to the adaptive control decision unit. The ideal state vector includes the ideal drawing tension and the ideal glass droplet morphology characteristics. The specific process of generating the ideal state vector is as follows: S201. Obtain the process parameters and state parameters under optimal production conditions as target data. The selection criteria for the target data are as follows: Under the condition of highest product quality: the process parameters corresponding to the minimum standard deviation of fiber diameter and the optimal mechanical properties; Under the most stable production conditions: the process parameters corresponding to the lowest wire breakage rate and the smallest fluctuation in wire drawing tension; Under the condition of highest production efficiency: the process parameters corresponding to the fastest wire drawing speed while ensuring quality. Under the condition of minimum energy consumption: the process parameters corresponding to the minimum energy consumption per unit output; S202. The selected target data is time-aligned and cleaned to remove outliers, resulting in optimized target data used as training samples. The training samples are divided into training and validation sets. Based on a nonlinear regression model, the target fiber diameter and ideal state parameters are combined, and their functional relationship is shown below: ; S203. Verify the baseline state model using the validation set until the validation result shows that the difference between the output ideal state and the actual optimal production state reaches the preset difference value. S204. Input the target fiber diameter setting into the reference state model to obtain the ideal state vector. The adaptive control decision unit is used to simultaneously receive the predicted state vector and the ideal state vector, calculate the deviation vector between the two, and based on the deviation vector, obtain the optimal control strategy vector required to minimize the deviation vector through a preset strategy optimization model, and output it to the control execution unit. The control strategy vector includes the target adjustment amount of drawing speed and the target adjustment amount of drawing temperature. The specific process of generating the optimal control strategy vector is as follows: S301. Construct a simulation environment for the fiber drawing process based on digital twins, which is used to input control actions, specifically the fiber drawing speed adjustment amount ΔV and the fiber drawing temperature adjustment amount ΔT. Based on the mechanism model calibrated according to fluid mechanics, heat transfer and historical data, accurately simulate and output the fiber drawing process state data at the next moment. The fiber drawing process state data includes fiber diameter, fiber drawing tension and droplet morphology. S302. Construct a Markov decision process model to formalize the adaptive control problem of the wire drawing process into a Markov decision process; The core elements for constructing a Markov decision process model are defined as follows: The intelligent agent is the deep neural network model embedded in the adaptive control decision-making unit. State S(t) is defined as the set of information that comprehensively describes the current working condition and future trends, which the agent relies on when making decisions. Its specific composition is as follows: Deviation vector component: The deviation calculated from the predicted state vector and the ideal state vector, for example, [predicted fiber diameter - target fiber diameter, predicted drawing tension - ideal drawing tension]; Current state component: Real-time data acquired by the multimodal data acquisition unit at the current moment, such as [current drawing temperature, current drawing speed, current fiber diameter]; Predicted state components: The predicted state vector output by the process prediction model unit, i.e., [predicted fiber diameter, predicted drawing tension, predicted fiber breakage probability].

[0022] By combining the above components, state S(t) provides the agent with a complete basis for decision-making regarding the past, present, and future as follows: Where ΔV is the target adjustment amount of the drawing speed and ΔT is the target adjustment amount of the drawing temperature; The reward function R(t) maximizes the long-term cumulative reward by enabling the agent to perform actions, and is composed of the following weighted components: Bias penalty term: A penalty is applied based on the L2 norm of the bias vector. The larger the bias, the larger the negative reward value. This is used to drive the agent to learn how to minimize the difference between the predicted state and the ideal state, and is described as follows: ; Smoothness control penalty: This penalty applies to the magnitude of action A(t) to encourage the agent to produce smoother and more stable control commands. It is expressed as follows: ; Risk aversion penalty term: A penalty is imposed on the predicted wire breakage probability in the predicted state vector. When the predicted risk of wire breakage increases, a significant negative reward is given, as described below: ; Stable operation reward: If the system does not enter a high-risk state in each control cycle, a small positive reward value is given to encourage the agent to maintain the long-term stable operation of the system; S303. A policy optimization model is constructed based on the dual-delay deep deterministic policy gradient algorithm. The training process is as follows: In the simulation environment, the agent outputs action A(t) through its actor network based on the current state S(t). After the simulation environment executes this action, it feeds back the state S(t+1) and reward R(t) for the next moment. The agent stores experience tuples (S(t), A(t), R(t), S(t+1)) into the experience replay pool and samples them to update its actor network and critic network. This process is iterated repeatedly until the agent's policy converges and can stably obtain high cumulative rewards, thus obtaining a policy optimization model. S304. Taking the real-time state vector S(t) as input, the actor network performs a forward propagation calculation, and its output is the optimal action A(t) that maximizes the long-term cumulative reward under the current state S(t). This action vector is the required optimal control strategy vector, as specifically described below: .

[0023] The control execution unit is used to receive the regulation strategy vector and convert it into real-time control commands, which are then sent to the frequency converter of the wire drawing machine and the power controller of the heating furnace to precisely adjust the actual wire drawing speed and wire drawing temperature.

[0024] This invention, by introducing a predictive model and an adaptive decision-making unit, decouples the two strongly coupled parameters of speed and temperature and performs collaborative optimization and control, achieving precise control of the fiber drawing process. This significantly improves the uniformity of fiber diameter and can predict future process trends based on real-time data, upgrading traditional feedback control to predictive feedforward control. Fine-tuning is performed before process parameters deviate significantly, effectively avoiding fiber breakage and waste, and greatly improving the yield. At the same time, the optimal production process is solidified into a benchmark model, and the complex control logic is condensed into a decision-making algorithm, eliminating reliance on human experience and ensuring a high degree of consistency in product quality across different shifts and production lines.

[0025] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0026] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart analysis and adaptive control system for glass fiber drawing and forming process, characterized in that, It includes a multimodal data acquisition unit, a process prediction model unit, a baseline process model unit, an adaptive control decision unit, and a control execution unit; The multimodal data acquisition unit is used to collect multi-dimensional data in real time during the wire drawing process. The multi-dimensional data includes wire drawing temperature, wire drawing speed, fiber diameter, wire drawing tension, and glass droplet morphology characteristics. The glass droplet morphology characteristics include droplet height, maximum diameter, and cone angle. The multi-dimensional data is timestamped and synchronized in time, and then integrated into a high-dimensional real-time data stream and sent to the process prediction model unit. The process prediction model unit is used to acquire and process multi-dimensional data. Based on the pre-trained process prediction model, it predicts key process indicators within a preset period and generates a prediction state vector to output to the adaptive control decision unit. The prediction state vector includes the predicted fiber diameter, the predicted drawing tension, and the predicted fiber breakage probability. The benchmark process model unit is used to obtain the ideal mapping relationship between process parameters and state parameters under optimal production conditions and to form a benchmark state model. When the target fiber diameter setting value is received, the unit outputs the ideal state vector corresponding to the setting value to the adaptive control decision unit. The ideal state vector includes the ideal drawing tension and the ideal glass droplet morphology characteristics. The adaptive control decision unit is used to simultaneously receive the predicted state vector and the ideal state vector, calculate the deviation vector between the two, and based on the deviation vector, obtain the optimal control strategy vector required to minimize the deviation vector through a preset strategy optimization model, and output it to the control execution unit. The control strategy vector includes the target adjustment amount of drawing speed and the target adjustment amount of drawing temperature. The control execution unit is used to receive the regulation strategy vector and convert it into real-time control commands, which are then sent to the frequency converter of the wire drawing machine and the power controller of the heating furnace to precisely adjust the actual wire drawing speed and wire drawing temperature.

2. The intelligent analysis and adaptive control system for glass fiber drawing and forming process according to claim 1, characterized in that, The following sensors are deployed within the distribution area of ​​the wire drawing and forming equipment, including laser diameter gauges, industrial cameras, infrared thermometers, tension sensors, and encoders, among which: Laser diameter gauges are used to measure the diameter of newly formed fibers in real time without damage. Industrial cameras are used to focus on the glass droplets below the platinum stencil and extract the droplet's height, maximum diameter, and cone angle in real time. Infrared thermometers are used for non-contact measurement of the surface temperature of glass droplets, serving as direct feedback on the drawing temperature. Tension sensors are installed on the fiber path to monitor tension changes in real time during the fiber drawing process; The encoder is used to integrate with the winding head motor of the wire drawing machine to accurately provide feedback on the real-time wire drawing speed.

3. The intelligent analysis and adaptive control system for glass fiber drawing and forming process according to claim 1, characterized in that, The specific process of generating the predicted state vector is as follows: S101. Obtain multi-dimensional data and integrate it into a key data vector. Where T(t) is the drawing temperature, V(t) is the drawing speed, D(t) is the real-time fiber diameter, F(t) is the drawing tension, H(t) is the droplet height, Dmax(t) is the maximum diameter, and θ(t) is the cone angle; S102. Obtain historical processing data of multiple sets of glass fiber drawing and forming process, construct time series samples, and use the normalized data as training samples. Construct a process prediction model based on long short-term memory network, download the weight file and load it onto the corresponding network to initialize the transfer network parameters. S103. Modify the last fully connected layer of the network, keep the input unchanged, set the output to the predicted state vector, initialize the weights of the last layer, use the gradient descent algorithm for learning, and use fixed step size decay to optimize the training parameters. Retrain the entire network to obtain the process prediction model. S104. During the training process, small batches of samples are randomly and non-repeatedly drawn from the training set for training. One training cycle is completed after all training samples are drawn. The training is completed after a certain number of cycles to obtain the basic process prediction model. S105. Input the key data vector into the basic process prediction model to obtain the basic process prediction model.

4. The intelligent analysis and adaptive control system for glass fiber drawing and forming process according to claim 1, characterized in that, The specific process of generating the ideal state vector is as follows: S201. Obtain the process parameters and state parameters under optimal production conditions as target data. The selection criteria for the target data are as follows: Under the condition of highest product quality: the process parameters corresponding to the minimum standard deviation of fiber diameter and the optimal mechanical properties; Under the most stable production conditions: the process parameters corresponding to the lowest wire breakage rate and the smallest fluctuation in wire drawing tension; Under the condition of highest production efficiency: the process parameters corresponding to the fastest wire drawing speed while ensuring quality. Under the condition of minimum energy consumption: the process parameters corresponding to the minimum energy consumption per unit output; S202. The selected target data is time-aligned and cleaned to remove outliers, resulting in optimized target data used as training samples. The training samples are divided into training and validation sets. Based on a nonlinear regression model, the target fiber diameter and ideal state parameters are related functionally as follows: ； S203. Verify the baseline state model using the validation set until the validation result shows that the difference between the output ideal state and the actual optimal production state reaches the preset difference value. S204. Input the target fiber diameter setting value into the reference state model to obtain the ideal state vector.

5. The intelligent analysis and adaptive control system for glass fiber drawing and forming process according to claim 1, characterized in that, The specific process of generating the optimal control strategy vector is as follows: S301. Construct a simulation environment for the fiber drawing process based on digital twins, which is used to input control actions, specifically the fiber drawing speed adjustment amount ΔV and the fiber drawing temperature adjustment amount ΔT. Based on the mechanism model calibrated according to fluid mechanics, heat transfer and historical data, accurately simulate and output the fiber drawing process state data at the next moment. The fiber drawing process state data includes fiber diameter, fiber drawing tension and droplet morphology. S302. Construct a Markov decision process model to formalize the adaptive control problem of the wire drawing process into a Markov decision process; S303. A policy optimization model is constructed based on the dual-delay deep deterministic policy gradient algorithm. The training process is as follows: In the simulation environment, the agent outputs action A(t) through its actor network based on the current state S(t). After the simulation environment executes this action, it feeds back the state S(t+1) and reward R(t) for the next moment. The agent stores experience tuples (S(t), A(t), R(t), S(t+1)) into the experience replay pool and samples them to update its actor network and critic network. This process is iterated repeatedly until the agent's policy converges and can stably obtain high cumulative rewards, thus obtaining a policy optimization model. S304. Taking the real-time state vector S(t) as input, the actor network performs a forward propagation calculation, and its output is the optimal action A(t) that maximizes the long-term cumulative reward under the current state S(t). This action vector is the required optimal control strategy vector, as specifically described below: 。 6. The intelligent analysis and adaptive control system for glass fiber drawing and forming process according to claim 1, characterized in that, The core elements for constructing a Markov decision process model are defined as follows: The intelligent agent is the deep neural network model embedded in the adaptive control decision-making unit. State S(t) is defined as the set of information that comprehensively describes the current working condition and future trends, which the agent relies on when making decisions. Its specific composition is as follows: Deviation vector component: The deviation calculated from the predicted state vector and the ideal state vector; Current state component: Real-time data acquired by the multimodal data acquisition unit at the current moment; Predicted state components: The predicted state vector output by the process prediction model unit, i.e., [predicted fiber diameter, predicted drawing tension, predicted fiber breakage probability].

7. By combining the above components, the state S(t) provides the agent with a complete basis for decision-making regarding the past, present, and future as follows: ,in, ΔV is the target adjustment value for drawing speed, and ΔT is the target adjustment value for drawing temperature; The reward function R(t) maximizes the long-term cumulative reward by enabling the agent to perform actions, and is composed of the following weighted components: Bias penalty term: A penalty is applied based on the L2 norm of the bias vector. The larger the bias, the larger the negative reward value. This is used to drive the agent to learn how to minimize the difference between the predicted state and the ideal state, and is described as follows: ； Smoothness control penalty: This penalty applies to the magnitude of action A(t) to encourage the agent to produce smoother and more stable control commands. It is expressed as follows: ； Risk aversion penalty term: A penalty is imposed on the predicted wire breakage probability in the predicted state vector. When the predicted risk of wire breakage increases, a significant negative reward is given, as described below: ； Stable operation reward: If the system does not enter a high-risk state in each control cycle, a small positive reward value is given to encourage the agent to maintain the long-term stable operation of the system.

Images