Coating viscosity control method and system
By constructing a time-weighted state vector and fusing it with a multi-path model, and combining the prediction confidence to adjust the control strategy, the problem of working condition adaptability and stability of viscosity control in the coating process was solved, and efficient coating process control was achieved.
Patent Information
- Application Number
- CN202511551497.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-28
- Publication Date
- 2026-01-09
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing viscosity control methods in coating processes are unable to cope with changes in operating conditions. Model predictive control methods become mismatched when switching operating conditions, leading to a decline in control performance. Furthermore, the reliability of prediction results cannot be evaluated online, which can easily cause system oscillations or missed response opportunities.
A time-weighted state vector containing historical information is constructed, and multiple benchmark path models are used for comprehensive prediction. The path deviation and stability are evaluated by Mahalanobis distance and fusion weights. The control cost function is adjusted by combining the prediction confidence to generate the optimal control sequence.
It improves the accuracy of viscosity prediction and adaptability to changes in operating conditions, ensures the stability of the control system, reduces system oscillation, improves response speed, and guarantees the consistency of product coating quality.
Smart Images
Figure CN121300084A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of viscosity control, and particularly relates to a coating viscosity control method and system. BACKGROUND
[0002] In the coating process, the viscosity of the coating liquid is a core process parameter for determining the uniformity, thickness and comprehensive performance of the product coating. Too high or too low viscosity will lead to coating defects, resulting in product quality decline and raw material waste. However, the coating process is easily disturbed by uncertain factors such as environmental temperature and raw material batch fluctuation, making the stable control of viscosity extremely challenging. The current industrial site still generally adopts an empirical manual adjustment control strategy, which relies on the experience of operators and is highly subjective.
[0003] An advanced control strategy based on a model, such as model predictive control (MPC), predicts the future state of the system by establishing a dynamic model of the process, and optimizes the control input within a limited time range. However, the existing model predictive control method is a static process model. The coating production process usually includes multiple different working condition zones such as start-up, stable operation and material change, and the characteristics of each working condition zone differ significantly. When the working condition switches, the model mismatch problem will cause the control performance to drop sharply. Moreover, the reliability of the prediction results cannot be evaluated online, and the aggressiveness of the control strategy cannot be adjusted accordingly. When the model prediction confidence is low, if the original control parameters are still used, the system may produce excessive control actions, which may even cause system oscillation and affect stability. On the contrary, when the prediction confidence is high, the system may miss the opportunity for fast response due to overly conservative control. SUMMARY
[0004] The application provides a coating viscosity control method and system to solve the technical problems of difficult control of coating liquid viscosity and defects in the existing MPC method.
[0005] In a first aspect, the application provides a coating viscosity control method, comprising the following steps: S1, obtaining real-time data of preset process parameters in the coating process, constructing a time sequence weighted state vector containing current and historical parameter information based on the real-time data and a preset exponential decay coefficient, and using a plurality of reference path models associated with different coating working condition zones to respectively perform forward evolution on the time sequence weighted state vector to obtain a plurality of candidate viscosity prediction paths; S2, calculating the Mahalanobis distance between the time sequence weighted state vector and the historical trajectory of each reference path model as a path deviation degree, evaluating the stability of each candidate viscosity prediction path, and determining a fusion weight for each candidate viscosity prediction path in combination with the path deviation degree and the stability of the candidate viscosity prediction path; S3, weighting and fusing the plurality of candidate viscosity prediction paths based on the fusion weights to generate a comprehensive viscosity prediction trajectory, and obtaining a confidence degree representing prediction reliability based on distribution characteristics of the fusion weights; S4, constructing a control sequence composed of control variable setting values of a plurality of future time steps, obtaining a joint cost function including a prediction error term between the comprehensive viscosity prediction trajectory and a target viscosity value, and a control cost term for penalizing variation of the control sequence, adjusting a penalty weight of the control cost term using the confidence degree, and solving an optimal control sequence minimizing the joint cost function to output the optimal control sequence as a control command to an actuator of the coating device.
[0006] Further, a time-series weighted state vector including current and historical parameter information is constructed based on real-time data and a preset exponential decay coefficient, including: At each sampling time t, real-time process parameters including coating speed, slurry supply flow rate, and drying temperature are collected to form an original state vector s(t) of the current sampling time; Exponential decay coefficient The value range is 0.1 to 0.3; The time-series weighted state vector X(t) of the current sampling time is generated by recursive calculation through the formula Wherein X(t-1) is the time-series weighted state vector of the previous sampling time.
[0007] Further, a fusion weight is determined for each candidate viscosity prediction path in combination with the path deviation degree and the stability of the candidate viscosity prediction path, including: For the i-th candidate viscosity prediction path, the path deviation degree corresponding to the candidate viscosity prediction path is calculated The variance of the viscosity prediction value of the candidate viscosity prediction path within the future prediction time window is calculated ; The path stability is calculated according to the formula Wherein is a preset positive number for ensuring numerical stability; The initial score of each candidate viscosity prediction path is calculated by the formula Wherein is a preset distance penalty factor; The initial scores of all candidate viscosity prediction paths are normalized to obtain the fusion weight of the i-th candidate viscosity prediction path .
[0008] Further, a confidence degree representing prediction reliability is obtained based on the distribution characteristics of the fusion weights, including: Fusion weight of all candidate viscosity prediction paths Fusion weight distribution vector Wherein N is the total number of candidate viscosity prediction paths; According to the Shannon entropy formula Calculate the information entropy H(W) of the fusion weight distribution vector; The information entropy is normalized and inverted to calculate the confidence The calculation formula is Wherein is the maximum possible entropy value.
[0009] Further, the confidence is used to adjust the penalty weight of the control cost term, including: Obtain the joint cost function Wherein the first term is the sum of squared prediction errors in the prediction time domain P, and the second term is the control increment penalty term in the control time domain M, is the control increment of the future k time step, and R is the penalty weight; Obtain a reference penalty weight And an adjustment sensitivity coefficient ; Based on the confidence C calculated at the last sampling time, the penalty weight R of the current control period is dynamically calculated by the formula .
[0010] Further, in S1, the coating process is divided into several typical coating condition areas by a clustering algorithm.
[0011] Further, in S1, an independent long short-term memory network LSTM model or state space model is trained for the data of each coating condition area as a reference path model.
[0012] Further, in S1, the process parameters related to viscosity in the PLC or DCS system of the coating machine are collected through the field bus.
[0013] Further, in S4, the minimum value of the joint cost function is solved by a numerical optimization algorithm such as sequential quadratic programming to obtain the optimal control sequence.
[0014] Secondly, the present application provides a coating viscosity control system, comprising a memory and a processor, and the memory stores computer program instructions, which realize the coating viscosity control method when executed by the processor.
[0015] The beneficial effects are as follows: By constructing a time-weighted state vector containing historical information and utilizing multiple benchmark path models associated with different coating conditions for comprehensive prediction, the accuracy of viscosity prediction and adaptability to changes in operating conditions are improved. This invention combines prediction confidence with the cost function of model predictive control. When the prediction confidence is high, larger adjustments are allowed to quickly respond to deviations; while when the confidence is low, excessive control actions are suppressed, prioritizing the stable operation of the system and reducing system oscillations caused by model uncertainties. While ensuring the stability of the control system, the control response speed is improved, enabling better handling of various disturbances during the coating process and ensuring the consistency of product coating quality. Attached Figure Description
[0016] Figure 1 This is a flowchart of a coating viscosity control method. Detailed Implementation
[0017] 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, not all, of the embodiments of the present invention. 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.
[0018] Examples of the coating viscosity control method provided by this invention: like Figure 1 As shown, the coating viscosity control method includes the following steps: S1. Obtain real-time data of preset process parameters during the coating process. Based on the real-time data and preset exponential decay coefficient, construct a time-weighted state vector containing current and historical parameter information. Using multiple benchmark path models associated with different coating conditions, perform forward evolution on the time-weighted state vector to obtain multiple candidate viscosity prediction paths.
[0019] Viscosity-related process parameters, such as coating speed, feed pump speed, oven temperature, and material tank level, are acquired via fieldbus from the PLC or DCS system of the coating machine. Let the vector formed by the parameters acquired at sampling time n be... Obtain an exponential decay coefficient between 0 and 1. Time-weighted state vector pass The calculation incorporates historical data into the current state in an exponentially decaying manner, resulting in an enhanced state representation of the trend.
[0020] In an optional embodiment, a time-weighted state vector containing current and historical parameter information is constructed based on real-time data and a preset exponential decay coefficient, including: At each sampling time t, real-time process parameters including coating speed, pulp supply flow rate, drying temperature are collected to form the original state vector s(t) of the current sampling time; Exponential decay coefficient The value range is 0.1 to 0.3; The recursive calculation is carried out through the formula to generate the time series weighted state vector X(t) of the current sampling time, where X(t-1) is the time series weighted state vector of the previous sampling time.
[0021] For example, at a certain sampling time t, the collected coating speed is 50 meters per minute, the pulp supply flow rate is 100 milliliters per minute, and the drying temperature is 80 degrees Celsius. At this time, the three components of the original state vector s(t) of the current sampling time are 50, 100 and 80. Assuming that the three components of the time series weighted state vector X(t-1) of the previous sampling time are 48, 102 and 81.
[0022] If the exponential decay coefficient a is 0.2, according to the formula, the calculation result of the time series weighted state vector X(t) of the current sampling time is the component of the vector, which is 48.4, 101.6 and 80.8 in turn. The new time series weighted state vector X(t) not only contains the latest process parameter information of the current time, but also retains the inertial influence of historical data through the time series weighted state vector X(t-1) of the previous sampling time, filters out the instantaneous noise, and makes the state evaluation more reliable.
[0023] According to historical production data, the coating process is divided into several typical coating condition areas, such as start-up acceleration area, stable production area and roll change transition area, by clustering algorithm. For the data of each coating condition area, an independent prediction model is trained, such as long short-term memory network LSTM model or state space model, and the model is the benchmark path model. In the forward evolution, the current time series weighted state vector is taken as the initial input of each benchmark path model, and the viscosity value of multiple time steps in the future is iteratively predicted in a preset future control time domain by using each model, and then multiple independent candidate viscosity prediction paths with the same number as the benchmark path model are obtained.
[0024] S2, calculate the Mahalanobis distance between the time series weighted state vector and the historical trajectory of each benchmark path model as the path deviation, evaluate the stability of each candidate viscosity prediction path, and determine a fusion weight for each candidate viscosity prediction path by combining the path deviation and the stability of the candidate viscosity prediction path.
[0025] Each reference path model corresponds to a historical trajectory distribution composed of model training data, which contains a mean vector and a covariance matrix. The Mahalanobis distance between the current time-weighted state vector and the mean vector of each model historical trajectory distribution is calculated, and the Mahalanobis distance value is the path deviation degree, which represents the matching degree of the current working condition and the working condition represented by each model. At the same time, the sum of the squares of the first-order differences between the internal prediction points of each candidate viscosity prediction path generated by each model is calculated as the stability index of the candidate viscosity prediction path, and the smaller the value, the smoother and more stable the prediction trajectory. The reciprocal of the path deviation degree and the reciprocal of the stability index are weighted and multiplied, and then normalized to obtain the fusion weight of each candidate viscosity prediction path.
[0026] In an optional embodiment, the path deviation degree and the stability of the candidate viscosity prediction path are combined to determine a fusion weight for each candidate viscosity prediction path, including: For the i-th candidate viscosity prediction path, the path deviation degree corresponding to the candidate viscosity prediction path is calculated , and the variance of the viscosity prediction value of the candidate viscosity prediction path in the future prediction time window is calculated ; The path stability is calculated according to the formula , wherein is a preset positive number for ensuring numerical stability; The initial score of each candidate viscosity prediction path is calculated by the formula , wherein is a preset distance penalty factor; The initial scores of all candidate viscosity prediction paths are normalized to obtain the fusion weight of the i-th candidate viscosity prediction path .
[0027] Suppose two candidate viscosity prediction paths, path one and path two, are generated. The path deviation degree of path one is 0.5, and the variance of the future viscosity prediction value is 0.04; the deviation degree of path two is larger, which is 1.2, but the prediction value is more stable, and the variance of the future viscosity prediction value is 0.01. A preset positive number ε for ensuring numerical stability is 0.001, and a distance penalty factor β is 0.5.
[0028] The stability of each path is calculated. The path stability of path one is approximately equal to 24.39. The path stability of path two approximately equal to 90.91. The score of path one is approximately 18.99. The score of path two is approximately 49.89 after similar calculation. The initial scores are normalized by dividing each path score by the sum of all path scores, which is 68.88. The final fusion weight of path one is approximately 0.276, and the fusion weight of path two is approximately 0.724.
[0029] S3, fusing the plurality of candidate viscosity prediction paths based on the fusion weights to generate a comprehensive viscosity prediction trajectory, and obtaining a confidence representing prediction reliability based on distribution characteristics of the fusion weights.
[0030] At each future prediction time step, the prediction values of all candidate viscosity prediction paths at the time step are weighted and averaged according to the respective fusion weights to obtain a comprehensive viscosity prediction value at the time step. The comprehensive viscosity prediction values of all future time steps are connected to obtain a comprehensive viscosity prediction trajectory. At the same time, the distribution of the fusion weight vector is analyzed, for example, the reciprocal of the Shannon entropy of the fusion weight vector is calculated or the maximum value of the Shannon entropy is directly taken as the confidence. If a fusion weight is much larger than other fusion weights, it indicates that the judgment of the current working condition is very accurate, and the confidence is high at this time; on the contrary, if multiple fusion weights are similar in size, it indicates that the working condition is ambiguous, and the confidence is low.
[0031] In an optional embodiment, obtaining a confidence representing prediction reliability based on distribution characteristics of the fusion weights comprises: using a fusion weight distribution vector composed of fusion weights of all candidate viscosity prediction paths where N is the total number of candidate viscosity prediction paths; calculating the information entropy H(W) of the fusion weight distribution vector according to the Shannon entropy formula normalizing and inverting the information entropy to calculate the confidence , and the calculation formula is where is the maximum possible entropy value.
[0032] The confidence of the overall prediction result is quantified by evaluating the concentration degree of the fusion weight distribution. It is assumed that there are three candidate viscosity prediction paths, that is, N is equal to 3. In a high certainty scenario, the calculated fusion weight distribution vector W may be a vector with components of 0.9, 0.05 and 0.05, indicating that most of the weights are concentrated in the first path.
[0033] The information entropy H(W) of the fusion weight distribution is calculated according to the Shannon entropy formula, and the result is about 0.598. Since there are three paths, the maximum possible entropy value is about 1.585. The confidence C is about 0.623. On the contrary, if the fusion weight distribution is very uniform, for example, the components of the fusion weight distribution vector W are 0.33, 0.34 and 0.33, the information entropy is close to the maximum value 1.585, and the calculated confidence C will be close to 0.
[0034] S4, constructing a control sequence composed of control variable setting values of future multiple time steps, obtaining a joint cost function, the joint cost function including a prediction error term between a comprehensive viscosity prediction trajectory and a target viscosity value, and a control cost term for penalizing a control sequence variation; adjusting a penalty weight of the control cost term by using the confidence; solving an optimal control sequence that minimizes the joint cost function, and outputting the optimal control sequence as a control instruction to an actuator of the coating device.
[0035] The control sequence is a sequence of feed pump speed setting values in future M control periods. The joint cost function is composed of two parts: the first part is the sum of squares of differences between the target viscosity value and the comprehensive viscosity prediction trajectory in future P prediction periods, representing the tracking performance; and the second part is the sum of squares of adjacent control action variations, i.e., feed pump speed increments in future M control periods, representing the control smoothness. In an embodiment, the penalty weight R of the control cost term is a reference penalty weight divided by the confidence C. When the confidence C is high, the penalty weight R is reduced, allowing more aggressive control adjustment; when the confidence C is low, the penalty weight R is increased, inhibiting large amplitude changes of the control action. The minimum value of the joint cost function is solved by a numerical optimization algorithm such as sequential quadratic programming, to obtain the optimal control sequence. Only the first control variable setting value of the optimal control sequence, such as the feed pump speed of the next second, is output to the PLC, which drives the frequency converter and then drives the feed pump, completing a control cycle.
[0036] In an optional embodiment, the penalty weight of the control cost term is adjusted by using the confidence, including: obtaining a joint cost function , wherein the first term is the sum of squares of prediction errors in the prediction time domain P, and the second term is a control increment penalty term in the control time domain M, is the control increment of the future kth time step, and R is the penalty weight; obtaining a reference penalty weight and an adjustment sensitivity coefficient ; Based on the confidence C calculated at the last sampling time, the penalty weight R of the current control period is dynamically calculated by the formula .
[0037] a reference penalty weight For example, when the confidence C is 0.9, the penalty weight R is about 3.57. The lower penalty weight encourages the controller to take larger adjustments to quickly and accurately track the target viscosity. On the other hand, when the prediction uncertainty is high, for example, the confidence C is 0.1, the penalty weight R is about 8.33. The higher penalty weight close to the reference penalty weight limits the controller's action magnitude, making the behavior more conservative, avoiding system oscillation due to overreaction when the prediction is unreliable, and ensuring the stability of the production process.
[0038] An embodiment of the coating viscosity control system provided by the present application comprises: A coating viscosity control system comprises a processor and a memory, the memory storing computer program instructions which, when executed by the processor, implement the coating viscosity control method described above.
[0039] The coating viscosity control system further comprises a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art and thus will not be described here.
[0040] In this disclosure, a "storage medium" or "computer readable medium" can be any available medium that can be accessed by a general purpose or special purpose computer system. By way of example, and not limitation, such computer readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code means in the form of computer readable instructions or data structures and that can be accessed by a general purpose or special purpose computer system, or a combination thereof. When information is transferred or provided over a network or another communications connection (either hardwired, wireless, or combination thereof) to a computer, the computer properly views the connection as a computer readable medium. Thus, any such connection is properly termed a computer readable medium. Combinations of the above should also be included within the scope of the computer readable media.
[0041] The above are preferred embodiments of the present application, not to limit the protection scope of the present application, therefore: any equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.
Claims
1. A method for controlling coating viscosity, characterized in that, Includes the following steps: S1. Obtain real-time data of preset process parameters during the coating process. Based on the real-time data and preset exponential decay coefficient, construct a time-weighted state vector containing current and historical parameter information. Using multiple benchmark path models associated with different coating conditions, perform forward evolution on the time-weighted state vector to obtain multiple candidate viscosity prediction paths. S2, calculate the Mahalanobis distance between the time-weighted state vector and the historical trajectory of each baseline path model as the path deviation, evaluate the stability of each candidate viscosity prediction path, and combine the path deviation and the stability of the candidate viscosity prediction path to determine a fusion weight for each candidate viscosity prediction path. S3, weighted fusion of multiple candidate viscosity prediction paths based on fusion weights to generate a comprehensive viscosity prediction trajectory, and confidence level representing prediction reliability is obtained based on the distribution characteristics of fusion weights; S4. Construct a control sequence consisting of control variable setpoints for multiple future time steps, obtain a joint cost function, which includes a prediction error term between the comprehensive viscosity prediction trajectory and the target viscosity value, and a control cost term that penalizes the change in the control sequence; adjust the penalty weight of the control cost term using confidence level; solve for the optimal control sequence that minimizes the joint cost function, and output the optimal control sequence as a control command to the actuator of the coating equipment.
2. The coating viscosity control method according to claim 1, characterized in that, Based on real-time data and a preset exponential decay coefficient, a time-weighted state vector containing current and historical parameter information is constructed, including: At each sampling time t, real-time process parameters including coating speed, slurry supply flow rate, and drying temperature are collected to form the original state vector s(t) at the current sampling time. Exponential decay coefficient The value ranges from 0.1 to 0.3; Through formula Perform recursive calculations to generate the time-weighted state vector X(t) at the current sampling time, where X(t-1) is the time-weighted state vector at the previous sampling time.
3. The coating viscosity control method according to claim 1, characterized in that, Combining path deviation and the stability of candidate viscosity prediction paths, a fusion weight is determined for each candidate viscosity prediction path, including: For the i-th candidate viscosity prediction path, calculate the path deviation corresponding to the candidate viscosity prediction path. Calculate the variance of the viscosity prediction values of candidate viscosity prediction paths within future prediction time windows. ; According to the formula Calculate path stability ,in This is a preset positive number used to ensure numerical stability; Through formula Calculate the initial score for each candidate viscosity prediction path. ,in The preset distance penalty factor; The initial scores of all candidate viscosity prediction paths are normalized to obtain the fusion weight of the i-th candidate viscosity prediction path. .
4. The coating viscosity control method according to claim 1, characterized in that, The confidence level representing the reliability of the prediction is obtained based on the distribution characteristics of the fusion weights, including: Utilizing the fusion weights of all candidate viscosity prediction paths The fusion weight distribution vector , where N is the total number of candidate viscosity prediction paths; According to Shannon's entropy formula Calculate the information entropy H(W) of the fusion weight distribution vector; Normalize and invert the information entropy, then calculate the confidence level. The calculation formula is: ,in This represents the maximum possible entropy value.
5. The coating viscosity control method according to claim 1, characterized in that, Adjusting the penalty weights of the control cost term using confidence levels includes: Obtain the joint cost function The first term is the sum of squared prediction errors in the prediction time domain P, and the second term is the control increment penalty term in the control time domain M. R represents the control increment at the k-th future time step, and R is the penalty weight. Obtain a baseline penalty weight And an adjustment sensitivity coefficient ; Based on the confidence level C calculated at the previous sampling time, using the formula... The penalty weight R for the current control cycle is dynamically calculated.
6. The coating viscosity control method according to claim 1, characterized in that, In S1, the coating process is divided into several typical coating conditions using a clustering algorithm.
7. The coating viscosity control method according to claim 6, characterized in that, In S1, for the data of each coating condition area, an independent Long Short-Term Memory (LSTM) network model or a state-space model is trained as the baseline path model.
8. The coating viscosity control method according to claim 1, characterized in that, In S1, viscosity-related process parameters are acquired from the PLC or DCS system of the coating machine via fieldbus.
9. The coating viscosity control method according to claim 1, characterized in that, In S4, the minimum value of the joint cost function is obtained by solving numerical optimization algorithms such as sequential quadratic programming, thus obtaining the optimal control sequence.
10. A coating viscosity control system, characterized in that, It includes a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the coating viscosity control method according to any one of claims 1-9 is implemented.