Flexible glass thickness difference regulation method and system based on the langevin algorithm

Abstract

The application discloses a flexible glass thickness difference regulation method and system based on the Langevin algorithm, and relates to the field of advanced manufacturing and intelligent control technology, wherein thickness data of a flexible glass substrate is collected in real time in a monitoring period, a thickness difference evaluation value is calculated through sliding window segmentation, key process parameters are extracted synchronously, and an optimization model with the minimum thickness difference as the target is constructed; the motion trajectory of particles in a deterministic driving force and a random fluctuation force potential field is simulated by using a Langevin equation optimization algorithm, global optimization is performed within a process parameter safety boundary, optimization adjustment instructions are generated and are sent to corresponding process execution mechanisms in real time, the system continuously monitors thickness feedback after adjustment, automatically triggers secondary optimization according to a judgment result, forms closed-loop regulation, and realizes online intelligent optimization of the thickness difference of the flexible glass through a thickness detection unit, a process parameter acquisition unit, a control unit integrated with a Langevin algorithm module and a process execution mechanism.

Description

Technical Field

[0001] This invention belongs to the field of advanced manufacturing and intelligent control technology, specifically relating to a method and system for controlling the thickness difference of flexible glass based on the Langevin algorithm. Background Technology

[0002] The in-depth application of intelligent manufacturing technology is driving the transformation of glass production towards higher precision and efficiency. Combining advanced optimization algorithms with process control enables real-time optimization and dynamic adjustment of key process parameters, becoming an important technical path to improve the quality and stability of flexible glass production.

[0003] Existing technologies for dealing with thickness variations in flexible glass substrates often employ traditional statistical process control or PID-based closed-loop control methods. These methods heavily rely on pre-set fixed thresholds and linearized models, failing to effectively address the nonlinear, time-varying, and strongly coupled characteristics of high-temperature melt flow during flexible glass forming. Furthermore, after collecting thickness data, only simple mean or range values ​​are often calculated as evaluation indicators, leading to an oversimplification of the mapping relationship between key process parameters and thickness variations, resulting in the loss of much relevant information hidden in the dynamic process. Secondly, existing optimization methods mostly rely on empirical rules for adjustment. These methods are prone to getting trapped in local optima, especially in the high-dimensional, multi-peak objective function space of flexible glass production. Often, only a single process parameter can be corrected successively. This makes the existing system slow to react when process drift or external disturbances occur, and the adjustment direction is prone to deviating from the global optimum. In addition, existing technologies mostly use offline modeling and periodic manual intervention. The model update cycle is long, and there is a lack of adaptive closed-loop logic based on real-time adjustment feedback to determine whether to initiate secondary optimization. This means that thickness difference exceeding the standard on the production line is often only discovered and corrected after a large number of scraps have been produced.

[0004] To address the aforementioned issues, this invention proposes a method and system for controlling the thickness difference of flexible glass based on the Langevin algorithm. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for controlling the thickness difference of flexible glass based on the Langevin algorithm, which solves the problem that existing technologies are difficult to optimize the thickness difference of flexible glass in real time and globally in nonlinear strongly coupled systems.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for controlling the thickness difference of flexible glass based on the Langevin algorithm, the method comprising: Step 1: Within the predetermined monitoring period, the thickness data of the flexible glass substrate on the target production line is locked in real time, a thickness data sequence is generated, the thickness difference assessment value of the target production line is calculated and the key process parameter sequence of the target production line within the monitoring period is correlated with the feature vector to generate a set of data sequences to be optimized. Step 2: Analyze the set of data sequences to be optimized, map the key process parameters into optimization vectors, map the thickness difference evaluation value into the target output, construct an optimization model with the objective of minimizing the thickness difference evaluation value, and use the optimization model as the energy function of the Langevin algorithm; Step 3: Solve the optimization model online based on the Langevin algorithm to generate optimization adjustment instructions for the corresponding key process parameters; The Langevin equation optimization algorithm simulates the trajectory of a particle in a potential field containing deterministic driving forces and random fluctuation forces, and performs global optimization within the allowable space of key process parameters. Step four: Issue the determined optimization and adjustment instructions to the corresponding process execution mechanism on the target production line to make real-time adjustments to the target production line. Based on the feedback of the real-time adjustment results, determine whether to perform secondary optimization.

[0007] As a further aspect of the present invention, in step one, the specific method for real-time locking of the thickness data of the flexible glass substrate on the target production line within a predetermined monitoring period and generating the thickness data sequence is as follows: Identify the target production line; Taking the current time t1 as the start time of a monitoring cycle, the thickness data of the flexible glass substrate generated on the target production line is extracted in real time within this monitoring cycle and arranged in the order of generation time, denoted as the thickness data sequence H1, H2, ..., Hn, where n is the total number of times within the monitoring cycle and the duration of the monitoring cycle is a preset value.

[0008] As a further aspect of the present invention, the specific method for generating the set of data sequences to be optimized in step one is as follows: Extract the thickness data sequence H1, H2, ..., Hn, and perform sliding window segmentation on the monitoring period with a preset window length L and sliding step size S to obtain K sliding windows. Extract the thickness data of each of the K sliding windows in the thickness data sequence H1, H2, ..., Hn, and arrange them in the original order to obtain K thickness subsequences, where S < L < n. Extract the maximum and minimum values ​​of each thickness subsequence and calculate the difference as the thickness difference evaluation value of the corresponding sliding window. Summarize K thickness difference evaluation values ​​and arrange them in chronological order as: D1, D2, ..., DK; Based on various pre-deployed sensors, key process parameters of all process actuators on the target production line are collected at a preset synchronous sampling frequency within the monitoring period. The key process parameters include at least one of the following: temperature field parameter HC of molten glass, flow velocity field parameter VC of the forming zone, tension parameter FC of the traction device, and temperature gradient parameter TC of the cooling zone. Arrange the key process parameters in chronological order to obtain the temperature field parameter sequence T-HC, the flow velocity field parameter sequence T-VC, the tension parameter sequence T-FC, and the temperature gradient parameter sequence T-TC. For K sliding windows within the monitoring period, all key process parameters within the corresponding sliding window time period are extracted based on T-HC, T-VC, T-FC, and T-TC, and their standard deviations are calculated and combined as the key process parameter feature vectors for the corresponding sliding window. A total of K key process parameter feature vectors are generated, represented as: TZ1, TZ2, ..., TZK. The thickness difference evaluation values ​​D1, D2, ..., DK in each sliding window are combined with the key process parameter feature vectors TZ1, TZ2, ..., TZK through time-series correlation to generate a set of data sequences to be optimized, DSO={(D1,TZ1),(D2,TZ2),...,(DK,TZK)}.

[0009] As a further aspect of the present invention, the specific method for parsing the set of data sequences to be optimized and constructing an optimization model with the objective of minimizing the thickness difference evaluation value in step two is as follows: Extract the key process parameter feature vectors TZ1, TZ2, ..., TZK and the thickness difference evaluation values ​​D1, D2, ..., DK associated with the key process parameters from the set of data sequences to be optimized, DSO; The key process parameter feature vectors are mapped to optimization vectors, and the thickness difference evaluation value is mapped to the target output. An optimization model is constructed with the objective of minimizing the thickness difference evaluation value, specifically as follows: minJ(X)=f(ΔD(X)) Where X is the optimization vector, including at least one set of key process parameter feature vectors, J(X) is the thickness difference comprehensive cost function, ΔD(X) represents the thickness difference evaluation value based on the corresponding key process parameter feature vector, and min represents minimization; J(X) is the energy function of Langevin's algorithm.

[0010] As a further aspect of the present invention, in step three, the specific method for generating optimization adjustment instructions for corresponding key process parameters by solving the optimization model online based on the Langevin algorithm is as follows: Extract the optimization vector X, and the operator defines the population size N, the current iteration number it, and the maximum iteration number Itmax of the Langevin algorithm in combination with the types of key process parameters; Based on deterministic driving force The initial position is constructed by the trajectory of motion in the potential field of random fluctuation forces, as shown in the following formula: X k,j =LB j +rand×(UB j -LB j ) Among them, X k,j This represents the initial position of the k-th search volume in the j-th dimension, i.e., the initial value of the key process parameter feature vector corresponding to the optimization vector, corresponding to one of HC, VC, FC, TC, and LB. j With UB j These represent the preset lower and upper safety limits of the corresponding key process parameter feature vectors, respectively. rand is a random number randomly distributed within the range [0,1], where 1≤j≤N; Based on the constructed initialization position, further processing is performed using the Langevin equation to generate optimization and adjustment instructions.

[0011] As a further aspect of the present invention, in step three, based on the constructed initialization position and combined with the Langevin equation, the specific method for generating optimization adjustment instructions includes: S61, the terms mapping Langevin's equation are: Where, m(d) 2 x / dt 2 ) represents the acceleration term, m(d) 2 x / dt 2 ) = ma, where m is the mass of the particle, a is the acceleration, -λ(dx / dt) is the damping term, -λ(dx / dt) = -λV, where V is the velocity of the particle, and λ is the damping coefficient. Random items, To satisfy the random force of fluctuation-dissipation relationship, T env This represents the actual ambient temperature of the simulated target production line; S62, define an equivalent algorithm to simplify the terms of the mapping Langevin equation to: a=-λV+randomForce The acceleration of the candidate solution corresponding to the k-th search volume at the current iteration number it is defined as: d k it =-λ×V k it +randomForce; The random force model is: randomForce = (√2λT)randn × (X) k it -X avg it), where randn represents a random number from the standard normal distribution (0,1), and X k it Let X be the position vector of the k-th search body at the it-th iteration, which is the value of a set of key process parameter feature vectors. k it As the current solution, X is used to calculate the directional difference of the random force. avg it The average of the two randomly selected search body positions in the current iteration is used as a dynamic reference center, causing the random force to point towards or away from this dynamic reference center, thereby promoting information interaction among the populations. X avg it =(X avg1 it +X avg2 it ) / 2, X avg1 it With X avg2 it The position vectors of two different search bodies randomly selected in the current iteration are drawn from the population without replacement. Where ℮ is the base of the natural logarithm, and T = T initial ℮ -(it / Itmax) T initial This is the initial simulation temperature; Based on the constructed equivalent algorithm, an update rate formula is generated, and the position of a single candidate solution is calculated.

[0012] As a further aspect of the present invention, in step three, the specific method for generating the update rate formula based on the constructed equivalent algorithm and calculating the position of a single candidate solution is as follows: The formula for generating the update rate is as follows: V k it+1 =V k it +a k it V k it+1 =b×a k it Where b = randn × β, β = sin[(π / 2) × it + π × (it / Itmax)] × [1 - (it / Itmax)] × ℮ -[-2×(it / Itmax)] b is the inertia weight, used to control the degree of influence of the previous velocity on the current velocity, and β is the standard deviation, which decreases dynamically with iteration; After the speed update, based on X k it+1 =Xk it +V k it+1 Calculate the new position X for each candidate solution k it+1 .

[0013] As a further aspect of the present invention, step three, in which the optimization adjustment instructions for the corresponding key process parameters are generated, further includes: The formula for building global search capabilities is as follows: X k it+1 =[X avg it +F k ×(X pb it -X k it )]+V k it+1 , where X pb it To find the optimal solution, F is randomly selected from the first four candidate solutions in each iteration. k F is the weighting coefficient. k =0.5+0.1×sin(randn); The determined X k it+1 As an optimization and adjustment instruction for the corresponding key process parameters, the output operation is executed.

[0014] As a further aspect of the present invention, the specific method for determining whether to perform secondary optimization based on the feedback of real-time adjustment results in step four is as follows: The generated optimization and adjustment instructions are sent to the corresponding process execution mechanism on the target production line. The corresponding process execution mechanism on the target production line is adjusted in real time and continuously monitored for a monitoring cycle. The thickness data sequence within this monitoring cycle is extracted, and the thickness difference evaluation value of the target production line is calculated and marked as HD2. Extract the thickness difference assessment value of the target production line within the monitoring period where the generated optimization adjustment instruction is located, and mark it as HD1; If HD2 < HD1 and the thickness difference assessment value HD2 < HD_th, then the real-time adjustment is considered qualified and no secondary optimization is required. The thickness difference assessment value is continuously monitored, where HD_th is the preset thickness difference qualification threshold. Conversely, if the real-time adjustment is not satisfactory, the optimization and adjustment instructions associated with various key process parameters within the current monitoring cycle will be regenerated, and secondary optimization will be performed.

[0015] A flexible glass thickness difference control system based on the Langevin algorithm, used to implement the method described in any one of the above, characterized in that the system comprises: Thickness detection unit, which measures the thickness data of flexible glass substrate in real time; The process parameter acquisition unit is used to acquire key process parameters of the process actuators on the production line in real time. The control unit is equipped with the Langevin equation optimization algorithm module, which receives thickness data and key process parameters and generates optimization adjustment instructions. The process actuator receives and executes optimization and adjustment instructions issued by the control unit to adjust key process parameters.

[0016] The beneficial effects of this invention are: This invention introduces the Langevin algorithm into the control of thickness difference in flexible glass, achieving dynamic and intelligent process optimization. Its core advantage lies in the combination of deterministic driving forces and random fluctuation forces in the Langevin equation, enabling the algorithm to possess both precise local search and global escape capabilities within the allowable space of key process parameters. This effectively avoids getting trapped in local optima, thereby improving the global optimization accuracy of thickness difference control. Based on real-time monitoring data, the online optimization model can quickly respond to production line fluctuations and generate optimization adjustment instructions. Combined with a feedback mechanism, it determines whether secondary optimization is needed, forming a closed-loop adaptive control. This shortens the control lag time, reduces the need for manual intervention, and achieves a stable improvement in the uniformity of flexible glass thickness and continuous optimization of production yield. This invention utilizes a sliding window segmentation technique to time-align and correlate continuously acquired flexible glass substrate thickness data with multi-source key process parameters, constructing a set of data sequences to be optimized, including thickness difference assessment values ​​and corresponding process feature vectors. This achieves precise quantification of the dynamic coupling relationship between process fluctuations and product quality indicators during production. By mapping the sequence set to an optimization model that aims to minimize the thickness difference assessment value, and introducing the energy function of the Langevin algorithm into the cost function, this invention effectively utilizes the algorithm's global optimization capability in non-convex, high-dimensional parameter spaces, avoiding the shortcomings of traditional gradient methods that are prone to getting trapped in local extrema. This allows for rapid convergence to the optimal combination of process parameters, reducing the thickness difference of the flexible glass substrate, improving product thickness consistency, and enhancing the adaptability and interpretability of process control. This invention introduces the Langevin equation to simulate the motion of physical particles in a potential field, combining deterministic driving forces with random fluctuation forces to achieve an adaptive balance between global search and local exploitation. This improves the convergence accuracy of the optimization process. Based on dynamically adjusted damping coefficients, simulated temperature, and inertia weights, the search step size and random perturbation intensity can be automatically adjusted according to the iteration process, avoiding premature convergence problems. At the same time, the dynamic reference center introduced in the random forces promotes information interaction among the population, while the target position enhancement strategy further helps the algorithm escape local optima, thereby generating better key process parameter adjustment instructions, enhancing the real-time performance of online solutions, and steadily improving product quality and production efficiency in complex industrial environments. This invention introduces a secondary optimization judgment logic. Based on the dynamic comparison between the real-time monitored thickness difference assessment value and the thickness difference assessment value of the previous monitoring cycle, it automatically determines whether the adjustment effect is qualified and triggers secondary optimization to avoid the accumulation of deviations. This adaptive feedback mechanism reduces the cost of manual inspection and repeated trial and error, effectively suppresses the thickness unevenness problem caused by process fluctuations, and, combined with the collaboration between the control unit driven by Langzhiwan's algorithm and the thickness detection and process parameter acquisition units, realizes a closed-loop response from data perception to execution adjustment. This improves the stable compliance rate of the thickness difference of flexible glass, while also possessing self-learning and continuous optimization capabilities. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the method described in this invention; Figure 2 This is a schematic diagram of the system described in this invention. Detailed Implementation

[0019] 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.

[0020] like Figure 1 As shown, this application provides a method and system for controlling the thickness difference of flexible glass based on the Langevin algorithm; As an embodiment 1 of this application, it specifically includes: Step 1: Within the predetermined monitoring period, the thickness data of the flexible glass substrate on the target production line is locked in real time, a thickness data sequence is generated, the thickness difference assessment value of the target production line is calculated and the key process parameter sequence of the target production line within the monitoring period is correlated with the feature vector to generate a set of data sequences to be optimized. Step 2: Analyze the set of data sequences to be optimized, map the key process parameters into optimization vectors, map the thickness difference evaluation value into the target output, construct an optimization model with the objective of minimizing the thickness difference evaluation value, and use the optimization model as the energy function of the Langevin algorithm; Step 3: Solve the optimization model online based on the Langevin algorithm to generate optimization adjustment instructions for the corresponding key process parameters; The Langevin equation optimization algorithm simulates the trajectory of a particle in a potential field containing deterministic driving forces and random fluctuation forces, and performs global optimization within the allowable space of key process parameters. Step four: Issue the determined optimization and adjustment instructions to the corresponding process execution mechanism on the target production line to make real-time adjustments to the target production line. Based on the feedback of the real-time adjustment results, determine whether to perform secondary optimization.

[0021] Example 2

[0022] This embodiment further discloses a flexible glass thickness difference control system based on the Langevin algorithm, building upon Embodiment 1. Figure 2 As shown, it specifically includes the following: The thickness detection unit includes one or more confocal displacement sensors arranged in a linear array along the width direction of the glass substrate. The sensors measure the cross-sectional thickness profile of the moving flexible glass substrate in real time at a preset sampling frequency and send the thickness data of the flexible glass substrate to the control unit. The process parameter acquisition unit consists of various intelligent sensors distributed throughout the target production line, specifically including: a distributed fiber optic temperature measurement system or a high-precision thermocouple matrix for acquiring temperature field parameters of molten glass; an ultrasonic Doppler velocimeter for acquiring flow field parameters in the forming zone; an integrated strain gauge tension sensor for acquiring tension parameters of the traction equipment, installed on the bearing seat of the traction roller; and a multi-point thermocouple tree for acquiring temperature gradient parameters in the cooling zone, vertically inserted into the cooling zone, to acquire key process parameters of the process actuators on the production line in real time. The control unit is equipped with the Langevin equation optimization algorithm module, which receives thickness data and key process parameters and generates optimization adjustment instructions. The process actuator receives and executes optimization and adjustment instructions issued by the control unit to adjust key process parameters.

[0023] Example 3

[0024] This embodiment, based on embodiment 2, further discloses a basic implementation of a method for controlling the thickness difference of flexible glass: In this embodiment, the target production line is a flexible glass substrate production line, specifically a production line that uses the overflow pull-down method to produce flexible glass substrates with a thickness of less than 0.1 mm. Its thickness difference assessment value is a key indicator affecting product yield. This method aims to dynamically adjust process parameters in real time to minimize the thickness difference assessment value. First, in step one, based on the operator's pre-set duration, such as 60 seconds, as a predetermined monitoring cycle, the system takes the current time t1 as the starting point of the monitoring cycle. In the following 60 seconds, through online laser displacement sensors and spectral confocal sensors, it locks and extracts the thickness data of the continuously generated flexible glass substrates on the target production line in real time. The collected thickness data is then arranged in chronological order to generate a thickness data sequence, denoted as H1, H2, ..., Hn, where n is the total number of sampling times within the monitoring cycle. The sampling frequency is set to 100Hz, so n = 6000. To further analyze the local fluctuations in the thickness difference of the flexible glass substrate, a sliding window technique was adopted. The window length L was set to 1000, and the total number of data n corresponding to the window length L was 10 seconds. The sliding step size S was set to 200, and the total number of data n corresponding to the sliding step size S was 2 seconds. Next, using a sliding window with a step size of S, the thickness data sequence of length n is divided into K sliding windows. The thickness data within each sliding window is extracted to form a thickness subsequence. The difference between the maximum and minimum values ​​in each thickness subsequence is calculated as the thickness difference evaluation value Di corresponding to that window, where i=1,...,K. The summations result in the thickness difference evaluation value sequence D1,D2,...,DK. It should be noted that the above steps are mainly based on a system, which includes a variety of sensors pre-deployed on the target production line to collect key process parameters at a sampling frequency of 100Hz, synchronized with the thickness data collection. The key process parameters include at least one of the following: temperature field parameter HC obtained by a thermocouple array arranged in the molten glass channel, flow velocity field parameter VC obtained by a laser Doppler velocimeter arranged in the forming zone, tension parameter FC obtained by a tension sensor on the traction roller, and temperature gradient parameter TC obtained by multiple thermocouples in the cooling zone. All the key process parameters obtained are arranged in chronological order to form temperature field parameter sequence T-HC, flow velocity field parameter sequence T-VC, tension parameter sequence T-FC, and temperature gradient parameter sequence T-TC, respectively. For each time period associated with a sliding window, data on all key process parameters within that time period are extracted from the aforementioned process parameter sequences, and their standard deviations are calculated. These parameters are then combined into a multidimensional vector, which is the feature vector of the key process parameters corresponding to that sliding window. , where HCI represents the standard deviation of the temperature field parameters within the i-th window, and so on; It is important to note that In This represents the transpose of a vector. In this way, we can obtain a total of K eigenvectors, which are represented as: TZ1, TZ2, ..., TZK.

[0025] Finally, the thickness difference evaluation value of each sliding window is combined with the corresponding key process parameter feature vector in a time series correlation to generate a set of data sequences to be optimized, DSO={(D1,TZ1),(D2,TZ2),...,(DK,TZK)}. The set of data sequences to be optimized represents the mapping relationship between the combination of process parameters at different times and the final thickness difference evaluation value within the monitoring period.

[0026] As described above, this embodiment uses sliding window technology to transform continuous time-series data into discrete sample pairs, aligning the thickness difference assessment value with process parameters in time, in order to capture the instantaneous impact of fluctuations in key process parameters on the actual quality of flexible glass substrates. Compared with the prior art, which only takes the extreme difference of the entire board thickness, this embodiment can effectively address the problem of non-uniform distribution of thickness difference in the length direction of the substrate.

[0027] Example 4

[0028] This embodiment, based on embodiment 3, further discloses a method for constructing an optimization model based on a set of data sequences to be optimized, specifically including the following: Based on the content described in Example 3, after obtaining the set of data sequences to be optimized, DSO, step two is executed. It should be noted that the control unit of the system in this solution is specifically an industrial programmable logic controller (PLC) with an embedded high-performance computing module. The control unit parses all key process parameter feature vectors TZ1,...,TZK and their corresponding thickness difference evaluation values ​​D1,...,DK from DSO. To minimize the thickness difference evaluation value, this embodiment transforms the control problem into a mathematical optimization problem. Specifically, the key process parameter feature vector TZ is mapped to an optimization vector. Where x1 corresponds to the temperature field parameter HC, x2 corresponds to the flow velocity field parameter VC, x3 corresponds to the tension parameter FC, and x4 corresponds to the temperature gradient parameter TC.

[0029] Simultaneously, the thickness difference evaluation value D is mapped to the objective output of the optimization problem, based on deterministic driving forces. The optimization model for the trajectory of motion in the potential field of random fluctuation force is: minJ(X)=f(ΔD(X)); Where J(X) is the comprehensive cost function of thickness difference, and ΔD(X) represents the thickness difference evaluation value under the current combination of key process parameters X; The function f() is the cost mapping function. In this embodiment, to simplify the calculation, ΔD(X) itself can be directly taken, that is, J(X) = ΔD(X); It should be noted that the deterministic driving force Fdet(X) can be determined by the perturbation observation method, and the perturbation stochastic approximation method can be used as an alternative, which is suitable for multi-parameter online optimization scenarios. The optimization objective of the optimization model is to find a set of X that minimizes J(X). In the subsequent Langevin algorithm, the J(X) function is directly defined as the energy function of the system. The lower the energy function value, the better the combination of key process parameters and the smaller the thickness difference.

[0030] Subsequently, step three is executed, and online solution is performed based on the Langevin algorithm. The system first extracts the domain of the optimization vector X, and the operator presets the hyperparameters of the Langevin algorithm through the human-computer interface. Specifically, the population size N=30, the current iteration number it, and the maximum iteration number Itmax=100, where the initial value of the current iteration number it is 0.

[0031] The core of the Langevin algorithm described in this scheme is to simulate the motion of particles in a potential field. First, within the allowable space of key process parameters, each search body in the population is initialized, i.e., the initial position of the corresponding particle. The initialization formula is: X k,j =LB j +rand×(UB j -LB j ); Among them, X k,j Let LB be the initial position of the k-th search volume in the j-th dimension (j=1,2,3,4), which corresponds to the feature vector of the key process parameters corresponding to the optimization vector and the initial candidate values ​​of the key process parameters HC, VC, FC, and TC. j With UB j These represent the preset lower and upper safety limits of the corresponding key process parameter feature vectors, respectively. For example, for HC, LB = 1100 degrees Celsius, UB = 1300 degrees Celsius, and rand is a random number randomly distributed in the range [0,1], where 1 ≤ j ≤ N. This embodiment transforms the abstract glass forming process control problem into a potential field optimization problem with clear physical meaning by constructing an energy function based on the thickness difference evaluation value. This lays the theoretical foundation for the introduction of the Langevin equation, making the optimization objective no longer a black box, but a calculable cost function directly related to product quality.

[0032] Example 5

[0033] This embodiment, based on embodiment 4, further discloses a specific implementation logic for online solving of the optimization model using the Langevin algorithm, as detailed below: After the population initialization is completed, the control unit executes the iterative solution process of the Langevin algorithm. In this embodiment, the classical Langevin equation is discretized and the algorithm is equivalent to adapt to the iterative calculation of digital computers.

[0034] First, let's map the classic Langevin equation, which becomes: Where, m(d) 2 x / dt 2 ) represents the acceleration term, m(d) 2 x / dt 2 ) = ma, where m is the mass of the particle, a is the acceleration, -λ(dx / dt) is the damping term proportional to the particle's velocity, -λ(dx / dt) = -λV, where V is the particle's velocity and λ is the damping coefficient. Random items, To satisfy the random force of fluctuation-dissipation relationship, T env This represents the actual ambient temperature of the simulated target production line; Next, taking the acceleration *a*, velocity *V*, damping coefficient *λ*, and random force *randomForce* as the direct operational objects in the algorithm, an equivalent algorithm is defined, resulting in the simplified equation: *a* = -λV + randomForce, where the acceleration of the candidate solution corresponding to the k-th search volume at the current iteration number *it* is discretized as: d k it =-λ×V k it +randomForce; In the equation, the model for the random force randomForce is: randomForce = (√2λT)randn × (X k it -X avg it ), where randn represents a random number from the standard normal distribution (0,1), and X k itLet X be the position vector of the k-th search body at the it-th iteration, which is the value of a set of key process parameter feature vectors. k it As the current solution, X is used to calculate the directional difference of the random force. avg it The average of the two randomly selected search body positions in the current iteration is used as a dynamic reference center, causing the random force to point towards or away from this dynamic reference center, thereby promoting information interaction among the populations. X avg it =(X avg1 it +X avg2 it ) / 2, X avg1 it With X avg2 it Let these be the position vectors of two different search bodies randomly selected without replacement from the population in the current iteration; Where ℮ is the base of the natural logarithm, and T = T initial ℮ -(it / Itmax) T initial This is the initial simulation temperature; Based on the above, the execution speed and position update are related, where the formula for the update speed is: V k it+1 =V k it +a k it V k it+1 =b×a k it Where b is the inertia weight, b = randn × β, β = sin[(π / 2) × it + π × (it / Itmax)] × [1 - (it / Itmax)] × ℮ -[-2×(it / Itmax)] The inertial weight b combines a sine function and exponential decay to control the degree of influence of the previous velocity on the current velocity, making the velocity update have oscillating and decaying characteristics, which helps to escape local optima. β is the standard deviation, which decreases dynamically with iteration. After the speed update, based on X k it+1 =X k it +V k it+1 Calculate the new position X for each candidate solution k it+1 ; To further enhance the global search capability, this embodiment discloses a target bit enhancement strategy to update the final candidate solution, specifically: Xk it+1 =[X avg it +F k ×(X pb it -X k it )]+V k it+1 , where X pb it F is the optimal solution in the current iteration, and in each iteration, one of the four candidate solutions is randomly selected. k F is the weighting coefficient. k =0.5 + 0.1 × sin(randn), whose value fluctuates between 0.4 and 0.6, used to control the step size for learning towards the optimal solution. After iteration, the final X is... k it+1 The position vector that minimizes the energy function J(X) is the optimal combination of key process parameters.

[0035] Finally, the system converts the optimal combination of key process parameters into specific optimization and adjustment instructions. For example, for the optimized value of the temperature field parameter HC, the instruction is encoded into a standardized analog signal message and sent to the process actuator.

[0036] It should be noted that the Langevin equation, in practical applications, also includes a parameter update step, specifically based on X. k+1 =X k -η×g k +σ×ξ k ; Where η is the learning rate, set to 0.01, σ is the diffusion coefficient, initially set to 0.05, used to control the amplitude of random perturbations, and η×g k For the deterministic descent term, g k The cost function J(X) is the cost function J(X) at the current X. k The gradient at a certain point, i.e., the gradient of the energy function, reflects the sensitive direction of the thickness difference as the parameter changes. This term drives the parameter to shift in the direction of decreasing thickness difference, σ×ξ k For the random exploration term, ξ k It is random noise with a standard normal distribution, a mean of 0, and a variance of 1.

[0037] Example 6

[0038] This embodiment further discloses a feedback mechanism and secondary optimization judgment logic in the control closed loop, based on embodiment 5, as follows: In step four, the control unit sends the optimization adjustment instructions generated in Example 5 to the corresponding process actuators on the target production line. For example, the temperature field adjustment instruction is sent to the thyristor power regulator of the electric heater in the molten glass channel; the flow rate field adjustment instruction is sent to the regulating valve actuator of the overflow trough; the tension adjustment instruction is sent to the servo driver of the traction roller; and the temperature gradient adjustment instruction is sent to the partitioned air valve actuator of the cooling zone. Each actuator adjusts its working state in real time according to the instructions.

[0039] Once the system issues the instruction, it immediately starts the next predetermined monitoring cycle. During this period, the system continuously monitors the thickness data of the newly produced flexible glass substrate and extracts the thickness data sequence within this new cycle according to the method described in Example 3. The thickness difference evaluation value of this cycle is calculated and denoted as HD2. At the same time, the thickness difference evaluation value of the monitoring cycle in which this optimization and adjustment instruction was generated and issued is retrieved and denoted as HD1.

[0040] Obtain the pre-set thickness difference acceptable threshold HD_th=15μm set by the operator, and judge HD1 and HD2 accordingly: If HD2 < HD1, and the thickness difference assessment value HD2 < HD_th, then the real-time adjustment is considered qualified and no secondary optimization is required. The thickness difference assessment value should be continuously monitored. Conversely, if the real-time adjustment is not satisfactory, the optimization and adjustment instructions associated with various key process parameters within the current monitoring cycle will be regenerated, and secondary optimization will be performed.

[0041] During the above process, secondary optimization may be triggered multiple times until the thickness difference evaluation value meets the dual requirements of improvement and compliance. If the standard cannot be met after more than 3 secondary optimizations, the system will issue an alarm to prompt the operator to intervene and check the target production line.

[0042] By implementing the above methods, a flexible glass substrate production line was controlled and compared with existing technologies such as empirical rule adjustment, traditional PID controller, genetic algorithm, and particle swarm optimization. Under the same production conditions, the line was run continuously for 8 hours, and the actual data was recorded for comparison. The comparison results are shown in Table 1. Main issues explained: Empirical rule-based regulation relies on manual adjustment and fixed experience windows, resulting in slow response to furnace temperature fluctuations, traction speed changes, and cooling condition drift, and poor regulation consistency.

[0043] Traditional PID controllers have fast calculation speeds, but they are difficult to tune when faced with the coupling of temperature field, flow velocity field, traction tension and cooling gradient, and are prone to hysteresis, overshoot or oscillation.

[0044] Genetic algorithms have strong global search capabilities, but require a large number of online iterations. The crossover, mutation, and selection processes are computationally expensive, and the convergence speed is affected by the population size and the initial population.

[0045] Particle swarm optimization (PSO) is faster than genetic algorithm, but it is prone to premature convergence under multi-peak, non-convex objective functions, and its convergence stability is affected by the initial particle distribution and parameter settings.

[0046] The Langevin optimization algorithm seeks optimization through the combined use of deterministic driving forces and stochastic fluctuation forces. It enhances the global search in the early stage, improves convergence stability in the later stage through damping and equivalent temperature decay, and ensures safe online adjustment through constraint projection.

[0047] All data in the formulas described above have been calculated with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0048] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0049] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. A method for controlling the thickness difference of flexible glass based on the Langevin algorithm, characterized in that, The method includes: Step 1: Within the predetermined monitoring period, the thickness data of the flexible glass substrate on the target production line is locked in real time, a thickness data sequence is generated, the thickness difference assessment value of the target production line is calculated and the key process parameter sequence of the target production line within the monitoring period is correlated with the feature vector to generate a set of data sequences to be optimized. Step 2: Analyze the set of data sequences to be optimized, map the key process parameters into optimization vectors, map the thickness difference evaluation value into the target output, construct an optimization model with the objective of minimizing the thickness difference evaluation value, and use the optimization model as the energy function of the Langevin algorithm; Step 3: Solve the optimization model online based on the Langevin algorithm to generate optimization adjustment instructions for the corresponding key process parameters; The Langevin equation optimization algorithm simulates the trajectory of a particle in a potential field containing deterministic driving forces and random fluctuation forces, and performs global optimization within the allowable space of key process parameters. Step four: Issue the determined optimization and adjustment instructions to the corresponding process execution mechanism on the target production line to make real-time adjustments to the target production line. Based on the feedback of the real-time adjustment results, determine whether to perform secondary optimization.

2. The method according to claim 1, characterized in that, In step one, the specific method for real-time locking of the thickness data of the flexible glass substrate on the target production line within a predetermined monitoring period and generating the thickness data sequence is as follows: Identify the target production line; Taking the current time t1 as the start time of a monitoring cycle, the thickness data of the flexible glass substrate generated on the target production line is extracted in real time within this monitoring cycle and arranged in the order of generation time, denoted as the thickness data sequence H1, H2, ..., Hn, where n is the total number of times within the monitoring cycle and the duration of the monitoring cycle is a preset value.

3. The method according to claim 2, characterized in that, In step one, the specific method for generating the set of data sequences to be optimized is as follows: Extract the thickness data sequence H1, H2, ..., Hn, and perform sliding window segmentation on the monitoring period with a preset window length L and sliding step size S to obtain K sliding windows. Extract the thickness data of each of the K sliding windows in the thickness data sequence H1, H2, ..., Hn, and arrange them in the original order to obtain K thickness subsequences, where S < L < n. Extract the maximum and minimum values ​​of each thickness subsequence and calculate the difference as the thickness difference evaluation value of the corresponding sliding window. Summarize K thickness difference evaluation values ​​and arrange them in chronological order as: D1, D2, ..., DK; Based on various pre-deployed sensors, key process parameters of all process actuators on the target production line are collected at a preset synchronous sampling frequency within the monitoring period. The key process parameters include at least one of the following: temperature field parameter HC, flow velocity field parameter VC, tension parameter FC, and temperature gradient parameter TC. Arrange the key process parameters in chronological order to obtain the temperature field parameter sequence T-HC, the flow velocity field parameter sequence T-VC, the tension parameter sequence T-FC, and the temperature gradient parameter sequence T-TC. For K sliding windows within the monitoring period, all key process parameters within the corresponding sliding window time period are extracted based on T-HC, T-VC, T-FC, and T-TC, and their standard deviations are calculated and combined as the key process parameter feature vectors for the corresponding sliding window. A total of K key process parameter feature vectors are generated, represented as: TZ1, TZ2, ..., TZK. The thickness difference evaluation values ​​D1, D2, ..., DK in each sliding window are combined with the key process parameter feature vectors TZ1, TZ2, ..., TZK through time-series correlation to generate a set of data sequences to be optimized, DSO={(D1,TZ1),(D2,TZ2),...,(DK,TZK)}.

4. The method according to claim 3, characterized in that, In step two, the specific method for parsing the set of data sequences to be optimized and constructing an optimization model with the objective of minimizing the thickness difference evaluation value is as follows: Extract the key process parameter feature vectors TZ1, TZ2, ..., TZK and the thickness difference evaluation values ​​D1, D2, ..., DK associated with the key process parameters from the set of data sequences to be optimized, DSO; The key process parameter feature vectors are mapped to optimization vectors, and the thickness difference evaluation value is mapped to the target output. An optimization model is constructed with the objective of minimizing the thickness difference evaluation value, specifically as follows: minJ(X)=f(ΔD(X)) Where X is the optimization vector, including at least one set of key process parameter feature vectors, J(X) is the thickness difference comprehensive cost function, ΔD(X) represents the thickness difference evaluation value based on the corresponding key process parameter feature vector, and min represents minimization; J(X) is the energy function of Langevin's algorithm.

5. The method according to claim 4, characterized in that, In step three, the specific method for generating optimization adjustment instructions for corresponding key process parameters by solving the optimization model online based on the Langevin algorithm is as follows: Extract the optimization vector X, and the operator defines the population size N, the current iteration number it, and the maximum iteration number Itmax of the Langevin algorithm in combination with the types of key process parameters; Based on deterministic driving force The initial position is constructed by the trajectory of motion in the potential field of random fluctuation forces, as shown in the following formula: X k,j =LB j +rand×(UB j -LB j ) Among them, X k,j This represents the initial position of the k-th search volume in the j-th dimension, i.e., the initial value of the key process parameter feature vector corresponding to the optimization vector, corresponding to one of HC, VC, FC, TC, and LB. j With UB j These represent the preset lower and upper safety limits of the corresponding key process parameter feature vectors, respectively. rand is a random number randomly distributed within the range [0,1], where 1≤j≤N; Based on the constructed initialization position, further processing is performed using the Langevin equation to generate optimization and adjustment instructions.

6. The method according to claim 5, characterized in that, In step three, based on the constructed initialization position and further processed using the Langevin equation, the specific methods for generating optimization and adjustment instructions include: S61, the terms mapping Langevin's equation are: Where, m(d) 2 x / dt 2 ) represents the acceleration term, m(d) 2 x / dt 2 ) = ma, where m is the mass of the particle, a is the acceleration, -λ(dx / dt) is the damping term, -λ(dx / dt) = -λV, where V is the velocity of the particle, and λ is the damping coefficient. Random items, To satisfy the random force of fluctuation-dissipation relationship, T env This represents the actual ambient temperature of the simulated target production line; S62, define an equivalent algorithm to simplify the terms of the mapping Langevin equation to: a=-λV+randomForce The acceleration of the candidate solution corresponding to the k-th search volume at the current iteration number it is defined as: d k it =-λ×V k it +randomForce; The random force model is: randomForce = (√2λT)randn × (X) k it -X avg it ), where randn represents a random number from the standard normal distribution (0,1), and X k it Let X be the position vector of the k-th search body at the it-th iteration, which is the value of a set of key process parameter feature vectors. k it As the current solution, X is used to calculate the directional difference of the random force. avg it The average of the two randomly selected search body positions in the current iteration is used as a dynamic reference center, causing the random force to point towards or away from this dynamic reference center, thereby promoting information interaction among the populations. X avg it =(X avg1 it +X avg2 it ) / 2, X avg1 it With X avg2 it The position vectors of two different search bodies randomly selected in the current iteration are drawn from the population without replacement. Where ℮ is the base of the natural logarithm, and T = T initial ℮ -(it / Itmax) T initial This is the initial simulation temperature; Based on the constructed equivalent algorithm, an update rate formula is generated, and the position of a single candidate solution is calculated.

7. The method according to claim 6, characterized in that, In step three, the specific method for generating the update rate formula based on the constructed equivalent algorithm and calculating the position of a single candidate solution is as follows: The formula for generating the update rate is as follows: V k it+1 =V k it +a k it ,V k it+1 =b×a k it Where b = randn × β, β = sin[(π / 2) × it + π × (it / Itmax)] × [1 - (it / Itmax)] × ℮ -[-2×(it / Itmax)] b is the inertia weight, used to control the degree of influence of the previous velocity on the current velocity, and β is the standard deviation, which decreases dynamically with iteration; After the speed update, based on X k it+1 =X k it +V k it+1 Calculate the new position X for each candidate solution k it+1 .

8. The method according to claim 7, characterized in that, In step three, generating the optimization and adjustment instructions for the corresponding key process parameters also includes: The formula for building global search capabilities is as follows: X k it+1 =[X avg it +F k ×(X pb it -X k it )]+V k it+1 , where X pb it To find the optimal solution, F is randomly selected from the first four candidate solutions in each iteration. k F is the weighting coefficient. k =0.5+0.1×sin(randn); The determined X k it+1 As an optimization and adjustment instruction for the corresponding key process parameters, the output operation is executed.

9. The method according to claim 8, characterized in that, In step four, the specific method for determining whether to perform secondary optimization based on the feedback from real-time adjustments is as follows: The generated optimization and adjustment instructions are sent to the corresponding process execution mechanism on the target production line. The corresponding process execution mechanism on the target production line is adjusted in real time and continuously monitored for a monitoring cycle. The thickness data sequence within this monitoring cycle is extracted, and the thickness difference evaluation value of the target production line is calculated and marked as HD2. Extract the thickness difference assessment value of the target production line within the monitoring period where the generated optimization adjustment instruction is located, and mark it as HD1; If HD2 < HD1 and the thickness difference assessment value HD2 < HD_th, then the real-time adjustment is considered qualified and no secondary optimization is required. The thickness difference assessment value is continuously monitored, where HD_th is the preset thickness difference qualification threshold. Conversely, if the real-time adjustment is not satisfactory, the optimization and adjustment instructions associated with various key process parameters within the current monitoring cycle will be regenerated, and secondary optimization will be performed.

10. A flexible glass thickness difference control system based on Langevin algorithm, used to implement the method according to any one of claims 1 to 9, characterized in that, The system includes: Thickness detection unit, which measures the thickness data of flexible glass substrate in real time; The process parameter acquisition unit is used to acquire key process parameters of the process actuators on the production line in real time. The control unit is equipped with the Langevin equation optimization algorithm module, which receives thickness data and key process parameters and generates optimization adjustment instructions. The process actuator receives and executes optimization and adjustment instructions issued by the control unit to adjust key process parameters.

Images