A purification process adaptive control method and system based on transfer learning
Patent Information
- Application Number
- CN202611022627.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-10-09
AI Technical Summary
[0004]本发明的目的就在于解决工艺参数的数据稀缺和工艺参数的前序历史参数相关联的问题，而提出一种基于迁移学习的纯化过程自适应控制方法及系统
通过构建了实时感知、偏差判定、参数寻优、反馈迭代的闭环控制架构。利用迁移成分分析和最大均值差异最小化策略，将源域知识有效迁移至目标域，解决了不同批次数据不匹配导致的负迁移问题，使预测模型在新工况下仍保持高精度和强泛化能力，为过程控制提供可靠的在线检测输入；利用LSTM机制构建虚拟软测量模型，成功建模了前序操作历史对当前纯度的影响，精准捕捉时序耦合与长延迟特性，有效克服纯化环节的大滞后反馈难题，实现超前调节；以纯度预测值与目标阈值的偏差量作为控制指令触发条件，驱动全局粗调与局部精调的分段式自适应优化，并通过高斯过程代理模型的在线更新形成自校正迭代机制，使控制策略能够持续适应填料批次更换、原料差异等慢时变扰动。实现了从离线经验向在线数据驱动的过程控制范式转变，大幅降低了因控制滞后和模型失配导致的生产损耗。
Smart Images
Figure CN122883818A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of process control technology, specifically relating to an adaptive control method and system for a purification process based on transfer learning. Background Technology
[0002] Purification process control is crucial throughout the entire downstream process of biopharmaceutical manufacturing. It serves as a key link between efficient upstream expression and the safety of the final formulation. Precise chromatography control can effectively remove impurities such as host proteins, nucleic acids, and viruses, ensuring high purity and bioactivity of the product, directly impacting the efficacy and safety of clinical applications. Simultaneously, process control strategies can significantly reduce batch-to-batch variability, meet GMP compliance requirements, improve yield and batch success rate, thereby reducing unit production costs. Current methods involve sampling from the production line at fixed time intervals for testing. After the test results are available, operators adjust the process parameters for subsequent batches based on the results. This control method, centered on offline detection and dominated by human experience, is currently the most common implementation technology in industrial purification workshops.
[0003] In actual purification, on the one hand, there are often differences in the distribution of process parameter data between different batches and equipment, as well as data gaps, which leads to negative migration caused by the mismatch between process parameters in the source domain and the target domain. On the other hand, the purification process has strong temporal coupling and long delay response characteristics. The purity at the current moment is not only related to the current parameters, but also closely related to the operation history of multiple previous time points. Due to the lack of previous historical parameters, the prediction and control accuracy decreases significantly as the process progresses. Summary of the Invention
[0004] The purpose of this invention is to solve the problems of scarce process parameter data and correlation between process parameters and their preceding historical parameters, and to propose an adaptive control method and system for purification processes based on transfer learning.
[0005] In a first aspect of this invention, an adaptive control method for a purification process based on transfer learning is first proposed, the method comprising: Obtain the process characteristic parameters of the source purification control model and the target purification control model, and use them as source domain data and target domain data, respectively; and obtain the process parameters under the current operating conditions. Feature parameter transfer is performed on source domain data and target domain data to obtain a transfer process feature parameter sequence; after training the purification control model with the transfer process feature parameter sequence, the process parameters under the current operating conditions are input into the purification control model for processing, and the purity prediction value is output. If the predicted purity value deviates from the preset purity threshold, the migration process characteristic parameter sequence is optimized to obtain an optimized process characteristic parameter sequence.
[0006] Optionally, the step of performing feature parameter migration on the source domain data and target domain data to obtain the migration process feature parameter sequence includes: Matrix transformation is performed on the source domain data and the target domain data to obtain the source domain matrix and the target domain matrix; The source domain matrix and the target domain matrix are merged to obtain a joint data matrix; Based on the number of samples of the source domain data and the target domain data, a maximum mean difference matrix is constructed. The maximum mean difference matrix is used to characterize the distribution difference distance between the source domain data and the target domain data in the kernel space. Based on the joint data matrix and the maximum mean difference matrix, a generalized eigenvalue decomposition equation is constructed to obtain the transfer transformation matrix; the generalized eigenvalue decomposition equation is specifically as follows:
[0007] in, It is a joint data matrix. It is the transpose of the joint data matrix. It is the maximum mean difference matrix. It is the identity matrix. It is the transformation matrix. It is a diagonal matrix of generalized eigenvalues; The target domain data is projected onto the common feature space through the migration transformation matrix to obtain a sequence of migration process feature parameters.
[0008] By constructing a maximum mean difference matrix, the distribution differences between the source and target domains in the kernel space are quantified. Then, a transfer transformation matrix is obtained using generalized eigenvalue decomposition, projecting the target domain data onto a common feature space. This effectively solves the problem of inconsistent distribution between historical source domain data and current target domain data during the purification process.
[0009] Optionally, the target domain data can be input into the purification control model for processing, and the output purity prediction values include: The migration process feature parameter sequence is transformed into a migration process feature sequence; The migration process feature sequence is continuously processed through a convolution module to obtain the migration convolution feature sequence; The migration convolution feature sequence is substituted into the purification control module for processing to obtain the purified feature sequence; The purified feature sequence is processed through a fully connected module, a convolution module, and an activation module to obtain a purity prediction value.
[0010] Using a Long Short-Term Memory (LSTM) network as its core, this system employs a three-gate structure (forget gate, input gate, and output gate) to finely manage process timing information. It precisely adapts to the dynamic characteristics of the purification process, which involves multiple variables, strong coupling, and long delays. The forget gate adaptively determines the proportion of historical information retained, effectively solving the gradient vanishing problem of traditional RNNs and enabling the model to capture the impact of operational parameters from tens of minutes or even hours ago on the current purity. The input gate intelligently filters valid information based on the current process state and writes it into the cell state, avoiding interference from redundant features. The output gate controls the filtering ratio of the final output information, ensuring that the prediction results at each time step incorporate bidirectional contextual information, significantly improving the accuracy and robustness of purity prediction.
[0011] Optionally, optimizing the migration process characteristic parameter sequence to obtain an optimized process characteristic parameter sequence includes: Using the migration process characteristic parameter sequence as the initial solution, and minimizing the deviation between the purity prediction value and the preset purity threshold as the fitness function, a fitness value is obtained; a first deviation threshold and a second deviation threshold are preset, wherein the first deviation threshold is greater than the second deviation threshold; Using the initial solution as the current optimal solution, the purity prediction value corresponding to the current optimal solution is obtained as the current optimal purity prediction value; Calculate the deviation between the current optimal purity prediction value and the preset purity threshold; When the deviation is greater than the first deviation threshold, a candidate parameter sequence is determined in the global search space; when the first deviation threshold is less than or equal to the deviation and the second deviation threshold, a candidate parameter sequence is determined in the neighborhood of the current optimal solution; when the deviation is less than the second deviation threshold, the iteration is terminated, and the current optimal solution is output as the optimized process feature parameter sequence. The candidate parameter sequence determined in each iteration is input into the purification control model to obtain the corresponding purity prediction value. If the corresponding purity prediction value is better than the current optimal purity prediction value, the current optimal solution is updated with the candidate parameter sequence, and the current optimal purity prediction value is updated with the corresponding purity prediction value. Repeat the above iterative process until the preset maximum number of iterations is exceeded, at which point the iteration terminates.
[0012] A dynamic balance between optimization accuracy and computational efficiency is achieved. When the predicted purity deviates significantly from the target threshold, a global search mode is activated to quickly locate the optimal region within the multi-parameter space, avoiding getting trapped in local optima. When the deviation narrows to between the first and second deviation thresholds, a local development mode is switched to perform a fine-grained search in the neighborhood of the current optimal solution, accelerating convergence to the global optimum. When the deviation is less than the second deviation threshold, the iteration is terminated early to save computational resources. This hierarchical strategy ensures both a wide-range optimization capability under extreme conditions and avoids excessive redundant computation when approaching the target, making it a highly efficient adaptive optimization mechanism.
[0013] Optionally, determining the candidate parameter sequence within the global search space and within the neighborhood of the current optimal solution includes: Using the migration process feature parameter sequence and historically determined candidate parameter sequences as training samples, and the fitness value corresponding to each candidate parameter sequence as the training target, a Gaussian process surrogate model is constructed. The Gaussian process surrogate model is used to characterize the mapping relationship between the process feature parameter sequence and the fitness value. In global search mode, the candidate parameter sequence is determined with the goal of minimizing the predicted mean output by the Gaussian process surrogate model; In the local development mode, a local search space is constructed with the current optimal solution as the center and a preset neighborhood radius as the boundary. Several parameter sequences are randomly sampled in the local search space. Each parameter sequence is input into the Gaussian process surrogate model to obtain the corresponding prediction mean. The parameter sequence with the smallest prediction mean is selected as the candidate parameter sequence. The candidate parameter sequence and its corresponding fitness value determined in each iteration are added to the training samples to update the Gaussian process surrogate model.
[0014] A Gaussian process surrogate model is introduced to fit the mapping relationship between the process feature parameter sequence and fitness values, guiding the generation and optimization of candidate solutions in both global search and local exploitation modes. Replacing the expensive real purity detection process with a data-driven surrogate model significantly reduces optimization costs. The Gaussian process not only outputs the predicted mean but also provides an uncertainty measure, offering a reliable confidence interval for the search strategy and effectively balancing exploration and exploitation. After each iteration, new samples are added to the training set to update the surrogate model, continuously improving model accuracy as the optimization progresses, forming a positive feedback loop. Ultimately, high-quality optimized process parameters are obtained with fewer iterations.
[0015] In a second aspect of this invention, an adaptive control device for a purification process based on transfer learning is proposed, comprising: Data acquisition module: Acquires process characteristic parameters of the source purification control model and the target purification control model, which are used as source domain data and target domain data, respectively; and acquires process parameters under the current operating conditions; Migration module: Performs feature parameter migration on source domain data and target domain data to obtain a sequence of migration process feature parameters; Purity prediction module: After training the purification control model by transferring the process feature parameter sequence, the process parameters under the current operating conditions are input into the purification control model for processing, and the purity prediction value is output. Optimization module: If the predicted purity value deviates from the preset purity threshold, the migration process characteristic parameter sequence is optimized to obtain an optimized process characteristic parameter sequence.
[0016] Optionally, the migration module includes: a matrix transformation module, a merging module, a maximum mean difference matrix module, a migration transformation matrix module, and a matrix projection module. The matrix transformation module is used to perform matrix transformation on source domain data and target domain data to obtain source domain matrix and target domain matrix; The merging module is used to merge the source domain matrix and the target domain matrix to obtain a joint data matrix; The maximum mean difference matrix module is used to construct a maximum mean difference matrix based on the number of samples of the source domain data and the target domain data. The maximum mean difference matrix is used to characterize the distribution difference distance between the source domain data and the target domain data in the kernel space. The migration transformation matrix module is used to construct a generalized eigenvalue decomposition equation based on the joint data matrix and the maximum mean difference matrix to obtain the migration transformation matrix. The matrix projection module is used to project the target domain data onto a common feature space through the migration transformation matrix to obtain a migration process feature parameter sequence.
[0017] Optionally, the purity prediction module is further used for: The migration process feature parameter sequence is transformed into a migration process feature sequence; The migration process feature sequence is continuously processed through a convolution module to obtain the migration convolution feature sequence; The migration convolution feature sequence is substituted into the purification control module for processing to obtain the purified feature sequence; The purified feature sequence is processed through a fully connected module, a convolution module, and an activation module to obtain a purity prediction value.
[0018] Optionally, the optimization module is further configured to: Using the migration process characteristic parameter sequence as the initial solution, and minimizing the deviation between the purity prediction value and the preset purity threshold as the fitness function, a fitness value is obtained; a first deviation threshold and a second deviation threshold are preset, wherein the first deviation threshold is greater than the second deviation threshold; Using the initial solution as the current optimal solution, the purity prediction value corresponding to the current optimal solution is obtained as the current optimal purity prediction value; Calculate the deviation between the current optimal purity prediction value and the preset purity threshold; When the deviation is greater than the first deviation threshold, a candidate parameter sequence is determined in the global search space; when the first deviation threshold is less than or equal to the second deviation threshold, a candidate parameter sequence is determined in the neighborhood of the current optimal solution; when the deviation is less than the second deviation threshold, the iteration is terminated, and the current optimal solution is output as the optimized process feature parameter sequence. The candidate parameter sequence determined in each iteration is input into the purification control model to obtain the corresponding purity prediction value. If the corresponding purity prediction value is better than the current optimal purity prediction value, the current optimal solution is updated with the candidate parameter sequence, and the current optimal purity prediction value is updated with the corresponding purity prediction value. Repeat the above iterative process until the preset maximum number of iterations is exceeded, at which point the iteration terminates.
[0019] Optionally, determining the candidate parameter sequence within the global search space and within the neighborhood of the current optimal solution includes: Using the migration process feature parameter sequence and historically determined candidate parameter sequences as training samples, and the fitness value corresponding to each candidate parameter sequence as the training target, a Gaussian process surrogate model is constructed. The Gaussian process surrogate model is used to characterize the mapping relationship between the process feature parameter sequence and the fitness value. In global search mode, the candidate parameter sequence is determined with the goal of minimizing the predicted mean output by the Gaussian process surrogate model; In the local development mode, a local search space is constructed with the current optimal solution as the center and a preset neighborhood radius as the boundary. Several parameter sequences are randomly sampled in the local search space. Each parameter sequence is input into the Gaussian process surrogate model to obtain the corresponding prediction mean. The parameter sequence with the smallest prediction mean is selected as the candidate parameter sequence. The candidate parameter sequence and its corresponding fitness value determined in each iteration are added to the training samples to update the Gaussian process surrogate model.
[0020] The beneficial effects of this invention are: A closed-loop control architecture was constructed, encompassing real-time sensing, deviation determination, parameter optimization, and feedback iteration. Utilizing transfer component analysis and a maximum mean difference minimization strategy, source domain knowledge was effectively transferred to the target domain, resolving the negative transfer problem caused by data mismatch between different batches. This ensured the prediction model maintained high accuracy and strong generalization ability under new operating conditions, providing reliable online detection input for process control. A virtual soft sensor model was constructed using the LSTM mechanism, successfully modeling the impact of previous operation history on current purity, accurately capturing temporal coupling and long-delay characteristics, effectively overcoming the large lag feedback problem in the purification process, and achieving proactive adjustment. The deviation between the purity prediction value and the target threshold was used as the trigger condition for control commands, driving piecewise adaptive optimization of global coarse and local fine adjustments. A self-correcting iterative mechanism was formed through online updates of the Gaussian process surrogate model, enabling the control strategy to continuously adapt to slow time-varying disturbances such as packing batch changes and raw material differences. This achieved a paradigm shift from offline experience to online data-driven process control, significantly reducing production losses caused by control lag and model mismatch. Attached Figure Description
[0021] The invention will now be further described with reference to the accompanying drawings.
[0022] Figure 1 A flowchart illustrating an adaptive control method for a purification process based on transfer learning, provided as an embodiment of the present invention; Figure 2 This is a framework diagram of an adaptive control system for a purification process based on transfer learning, provided as an embodiment of the present invention. Detailed Implementation
[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0024] It should be noted that all formula calculations in the scheme are purely numerical calculations.
[0025] 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.
[0026] This invention provides an adaptive control method for a purification process based on transfer learning. See also... Figure 1 The method includes the following steps: S101: Obtain the process characteristic parameters of the source purification control model and the target purification control model, and use them as source domain data and target domain data, respectively; and obtain the process parameters under the current operating conditions; S102: Perform feature parameter migration on the source domain data and target domain data to obtain a migration process feature parameter sequence; S103: After training the purification control model by transferring the process characteristic parameter sequence, the process parameters under the current operating conditions are input into the purification control model for processing, and the purity prediction value is output. S104: If the purity prediction value deviates from the preset purity threshold, the migration process characteristic parameter sequence is optimized to obtain the optimized process characteristic parameter sequence.
[0027] In one implementation, the process parameters that need to be controlled during the bioprocessing (protein) include flow rate, eluent concentration, pH setpoint, temperature, stirring speed, etc.; the source purification control model and the target purification control model are based on different data sources (spatiotemporal context), and they respectively represent existing historical / reference knowledge and the current working conditions to be adapted.
[0028] In one embodiment, performing feature parameter migration on source domain data and target domain data to obtain a migration process feature parameter sequence includes: Matrix transformation is performed on the source domain data and the target domain data to obtain the source domain matrix and the target domain matrix; The source domain matrix and the target domain matrix are merged to obtain a joint data matrix; Based on the sample size of the source domain data and the target domain data, a maximum mean difference matrix is constructed. The maximum mean difference matrix is used to characterize the distribution difference distance between the source domain data and the target domain data in the kernel space. Based on the joint data matrix and the maximum mean difference matrix, a generalized eigenvalue decomposition equation is constructed to obtain the transfer transformation matrix; the generalized eigenvalue decomposition equation is as follows:
[0029] in, It is a joint data matrix. It is the transpose of the joint data matrix. It is the maximum mean difference matrix. It is the identity matrix. It is the transformation matrix. It is a diagonal matrix of generalized eigenvalues; The target domain data is projected onto the common feature space through a migration transformation matrix to obtain a sequence of migration process feature parameters.
[0030] In a specific embodiment, taking the affinity chromatography purification process of a protein A as an example, the source domain consists of historical normal batches, containing time-series data (1000 samples) across five dimensions: flow rate, eluent concentration, pH, temperature, and stirring speed. The target domain consists of the current batch, where the data distribution has shifted due to a change in the chromatography column packing material (only 100 real-time samples are obtained). During implementation, firstly, the original five-dimensional data from both domains are normalized to zero mean and merged into a joint data matrix X. Secondly, based on the ratio of the 1000 samples in the source domain to the 100 samples in the target domain, 1 / 1000 of each sample is filled. 2 1 / 100 2 A MMD matrix L is constructed using -1 / (1000×100); then, it is substituted into the generalized eigenvalue equation for solution. The eigenvectors corresponding to the first 5 smallest generalized eigenvalues are selected to form the migration transformation matrix W. Finally, the original data of the target domain is projected onto the common feature space through W to obtain the migration process feature parameter sequence. This sequence is precisely aligned with the source domain at the distribution level, effectively overcoming the negative migration problem caused by packing material replacement.
[0031] In one implementation, the dimension of the common feature space is smaller than the dimension of the original process feature parameters. The maximum mean difference distance between the source domain data and the target domain data in the common feature space is minimized to achieve data alignment from the source domain to the target domain. Through migration component analysis, the original multidimensional process feature parameters of the source and target domains are synchronously projected to the low-dimensional common feature space, and the maximum mean difference is minimized as a distribution distance metric. This enables the data of the two domains to achieve precise alignment at the distribution level in the common feature space, thereby effectively overcoming the technical defect of decreased model prediction accuracy caused by data distribution shift under different operating conditions, and avoiding the negative migration problem caused by directly reusing the source domain model in traditional methods. In addition, the above optimization problem is equivalently transformed into generalized eigenvalue decomposition for solution. The optimal transformation matrix can be obtained in one step without iterative optimization, resulting in high computational efficiency and meeting the application requirements of real-time control of industrial purification processes.
[0032] In one embodiment, target domain data is input into a purification control model for processing, and the output purity prediction value includes: The migration process feature parameter sequence is transformed into a migration process feature sequence; The migration process feature sequence is continuously processed through a convolution module to obtain the migration convolution feature sequence; The transfer convolution feature sequence is substituted into the purification control module for processing to obtain the purified feature sequence; The purified feature sequence is processed through a fully connected module, a convolutional module, and an activation module to obtain the purity prediction value.
[0033] In one implementation, the purification control module processes the following steps: The process parameter sequence is arranged in chronological order, with a length of T (T consecutive sampling times, such as sampling once every 30 seconds, for a total of 100 time points). LSTM processes the parameters sequentially according to the time step. The system receives the transfer convolution feature sequence and inputs it into the first LSTM unit. It processes the sequence step by step and outputs the hidden state sequence. The hidden state sequence output from the first layer is then input into the second LSTM unit. The sequence is processed step by step and outputs the hidden state sequence. This process continues until the output of the last LSTM layer is used as the final purified feature sequence.
[0034] The purification control module consists of multiple LSTM units connected sequentially over time. Each LSTM unit corresponds to the input of a time step, and the information of the entire time series is linked together as a whole through the transfer of hidden states and cell states between units. The LSTM typically uses a two-layer stacked structure. The first layer has 64–128 neurons to capture short-term process features, and the second layer has 32–64 neurons to extract long-term time dependencies. Dropout layers (with a ratio of 0.1–0.3) are added between layers to prevent overfitting. The input layer dimension is equal to the number of process feature parameters (such as flow rate, pH, temperature, stirring speed, etc., 5 dimensions). The time step size depends on the length of the historical sampling window. The output layer has one neuron to output the purity prediction value. Finally, the prediction is mapped through 1–2 layers of fully connected networks. Results: During training, the Adam optimizer is typically used with an initial learning rate of 0.001 (which can be dynamically adjusted using learner rate decay strategies such as ReduceLROnPlateau). The batch size is set to 32–64, and the number of iterations is generally 50–200 rounds, with early stopping to prevent overfitting. The loss function is the mean squared error. For data preprocessing, the original process parameter time series data needs to be cleaned (missing and outliers are handled), then Min-Max normalization or Z-Score standardization is performed to eliminate the influence between different units. The time series is then converted into a supervised learning format (process parameters from the previous N time steps are used as input, and purity values from the current or future time steps are used as labels). Finally, the data is divided into training, validation, and test sets in a ratio of approximately 7:1.5:1.5.
[0035] In one implementation, the purification control module is based on a long short-term memory network. Through a three-gate structure of forget gate, input gate, and output gate, it controls the forgetting ratio of historical information, the writing ratio of current information, and the filtering ratio of output information, respectively. This enables the storage and transmission of key information of any time length, making it suitable for time-series prediction scenarios with multiple variables, strong coupling, and long delays during the purification process. Furthermore, the output at each time step incorporates the contextual information before and after that time step, making the expression of process characteristics more comprehensive.
[0036] In one embodiment, optimizing the migration process characteristic parameter sequence to obtain an optimized process characteristic parameter sequence includes: Using the sequence of migration process characteristic parameters as the initial solution, and minimizing the deviation between the predicted purity value and the preset purity threshold as the fitness function, the fitness value is obtained; a first deviation threshold and a second deviation threshold are preset, with the first deviation threshold being greater than the second deviation threshold; Using the initial solution as the current optimal solution, the purity prediction value corresponding to the current optimal solution is obtained as the current optimal purity prediction value; Calculate the deviation between the current optimal purity prediction value and the preset purity threshold; When the deviation is greater than the first deviation threshold, a candidate parameter sequence is determined in the global search space; when the first deviation threshold is less than or equal to the second deviation threshold, a candidate parameter sequence is determined in the neighborhood of the current optimal solution; when the deviation is less than the second deviation threshold, the iteration is terminated, and the current optimal solution is output as the optimized process characteristic parameter sequence. The candidate parameter sequence determined in each iteration is input into the purification control model to obtain the corresponding purity prediction value. If the corresponding purity prediction value is better than the current best purity prediction value, the current best solution is updated with the candidate parameter sequence and the current best purity prediction value is updated with the corresponding purity prediction value. Repeat the above iterative process until the preset maximum number of iterations is exceeded, at which point the iteration terminates.
[0037] In one implementation, the fitness function includes: Where X represents the sequence of candidate process parameters, specifically including flow rate, eluent concentration, pH setpoint, temperature, stirring speed, etc. It is the purity prediction value corresponding to the candidate parameter sequence output by the purification control model, and its value range is [0,1]. It is a preset purity threshold.
[0038] In one implementation, a segmented deviation threshold judgment mechanism divides the optimization control of the purification process into three stages: global exploration, local development, and termination output. This enables adaptive adjustment of the optimization strategy: when the predicted purity deviates significantly from the target threshold, a sequence of candidate parameters is determined in the global search space, allowing the search process to quickly approach the potential region within the feasible domain and avoid getting trapped in local extrema; when the deviation narrows to the threshold range, a fine search is conducted in the neighborhood of the current optimal solution to ensure optimization accuracy; and the process terminates promptly when the deviation reaches the target, avoiding unnecessary energy consumption and media loss due to over-purification. Simultaneously, the optimal solution is updated by feeding back the evaluation results of the actual model in each iteration, ensuring that the search direction is always based on the predictions of the actual purification control model, approaching the global optimum within a limited number of iterations. The deviation-driven segmented search effectively balances global convergence speed and local search accuracy, significantly reducing the number of invalid model calls and taking into account both the timeliness and solution quality requirements of the purification process.
[0039] In one embodiment, determining the candidate parameter sequence within the global search space and determining the candidate parameter sequence within the neighborhood of the current optimal solution includes: Using the migration process feature parameter sequence and historically determined candidate parameter sequences as training samples, and the fitness value corresponding to each candidate parameter sequence as the training objective, a Gaussian process surrogate model is constructed. The Gaussian process surrogate model is used to characterize the mapping relationship between the process feature parameter sequence and the fitness value. In global search mode, the candidate parameter sequence is determined with the goal of minimizing the predicted mean output by the Gaussian process surrogate model; In the local development mode, a local search space is constructed with the current optimal solution as the center and a preset neighborhood radius as the boundary. Several parameter sequences are randomly sampled in the local search space, and each parameter sequence is input into the Gaussian process surrogate model to obtain the corresponding prediction mean. The parameter sequence with the smallest prediction mean is selected as the candidate parameter sequence. The candidate parameter sequence and its corresponding fitness value determined in each iteration are added to the training samples to update the Gaussian process surrogate model.
[0040] In one implementation, a Gaussian process surrogate model (Gaussian process regression) is constructed to guide the search process of candidate parameter sequences. In the global exploration phase, the predicted mean of the surrogate model is used to quickly locate potential regions over a large area. In the local development phase, neighborhood sampling is used to achieve a fine search, balancing the convergence speed and solution accuracy of the optimization process. At the same time, by feeding back the real evaluation results to the training samples and updating the surrogate model after each iteration, the surrogate model continuously approaches the real objective function shape during the iteration process, achieving efficient use of data. This is especially suitable for industrial scenarios where the evaluation cost of purification control models is high, effectively reducing the number of iterations and actual model calls.
[0041] This invention also provides an adaptive control system for a purification process based on transfer learning. See also... Figure 2 The system includes the following modules: Data acquisition module (201): acquires the process characteristic parameters of the source purification control model and the target purification control model, respectively as source domain data and target domain data; and acquires the process parameters under the current operating conditions; Migration module (202): Performs feature parameter migration on source domain data and target domain data to obtain a sequence of migration process feature parameters; Purity prediction module (203): After training the purification control model by transferring the process characteristic parameter sequence, the process parameters under the current operating conditions are input into the purification control model for processing, and the purity prediction value is output. Optimization module (204): If the purity prediction value deviates from the preset purity threshold, the migration process feature parameter sequence is optimized and controlled to obtain the optimized process feature parameter sequence.
[0042] The foregoing has described one embodiment of the present invention in detail, but this is merely a preferred embodiment and should not be considered as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims.
Claims
1. An adaptive control method for a purification process based on transfer learning, characterized in that, The method includes: Obtain the process characteristic parameters of the source purification control model and the target purification control model, and use them as source domain data and target domain data, respectively; and obtain the process parameters under the current operating conditions. A sequence of migration process feature parameters is obtained by performing feature parameter transfer on source domain data and target domain data; After training the purification control model by transferring the process feature parameter sequence, the process parameters under the current operating conditions are input into the purification control model for processing, and the purity prediction value is output. If the predicted purity value deviates from the preset purity threshold, the migration process characteristic parameter sequence is optimized to obtain an optimized process characteristic parameter sequence.
2. The adaptive control method for purification process based on transfer learning according to claim 1, characterized in that, The feature parameter sequence obtained by performing feature parameter migration on source domain data and target domain data includes: Matrix transformation is performed on the source domain data and the target domain data to obtain the source domain matrix and the target domain matrix; The source domain matrix and the target domain matrix are merged to obtain a joint data matrix; Based on the number of samples of the source domain data and the target domain data, a maximum mean difference matrix is constructed. The maximum mean difference matrix is used to characterize the distribution difference distance between the source domain data and the target domain data in the kernel space. Based on the joint data matrix and the maximum mean difference matrix, a generalized eigenvalue decomposition equation is constructed to obtain the migration transformation matrix; The target domain data is projected onto the common feature space through the migration transformation matrix to obtain a sequence of migration process feature parameters.
3. The adaptive control method for purification process based on transfer learning according to claim 1, characterized in that, The target domain data is input into the purification control model for processing, and the output purity prediction values include: The migration process feature parameter sequence is transformed into a migration process feature sequence; The migration process feature sequence is continuously processed through a convolution module to obtain the migration convolution feature sequence; The migration convolution feature sequence is substituted into the purification control module for processing to obtain the purified feature sequence; The purified feature sequence is processed through a fully connected module, a convolution module, and an activation module to obtain a purity prediction value.
4. The adaptive control method for purification process based on transfer learning according to claim 1, characterized in that, The optimized process characteristic parameter sequence obtained by optimizing and controlling the migration process characteristic parameter sequence includes: Using the migration process characteristic parameter sequence as the initial solution, and minimizing the deviation between the purity prediction value and the preset purity threshold as the fitness function, a fitness value is obtained; a first deviation threshold and a second deviation threshold are preset, wherein the first deviation threshold is greater than the second deviation threshold; Using the initial solution as the current optimal solution, the purity prediction value corresponding to the current optimal solution is obtained as the current optimal purity prediction value; Calculate the deviation between the current optimal purity prediction value and the preset purity threshold; When the deviation is greater than the first deviation threshold, a candidate parameter sequence is determined in the global search space; when the first deviation threshold is less than or equal to the second deviation threshold, a candidate parameter sequence is determined in the neighborhood of the current optimal solution; when the deviation is less than the second deviation threshold, the iteration is terminated, and the current optimal solution is output as the optimized process feature parameter sequence. The candidate parameter sequence determined in each iteration is input into the purification control model to obtain the corresponding purity prediction value. If the corresponding purity prediction value is better than the current optimal purity prediction value, the current optimal solution is updated with the candidate parameter sequence, and the current optimal purity prediction value is updated with the corresponding purity prediction value. Repeat the above iterative process until the preset maximum number of iterations is exceeded, at which point the iteration terminates.
5. The adaptive control method for purification process based on transfer learning according to claim 4, characterized in that, The steps of determining the candidate parameter sequence in the global search space and determining the candidate parameter sequence in the neighborhood of the current optimal solution include: Using the migration process feature parameter sequence and historically determined candidate parameter sequences as training samples, and the fitness value corresponding to each candidate parameter sequence as the training target, a Gaussian process surrogate model is constructed. The Gaussian process surrogate model is used to characterize the mapping relationship between the process feature parameter sequence and the fitness value. In global search mode, the candidate parameter sequence is determined with the goal of minimizing the predicted mean output by the Gaussian process surrogate model; In the local development mode, a local search space is constructed with the current optimal solution as the center and a preset neighborhood radius as the boundary. Several parameter sequences are randomly sampled in the local search space. Each parameter sequence is input into the Gaussian process surrogate model to obtain the corresponding prediction mean. The parameter sequence with the smallest prediction mean is selected as the candidate parameter sequence. The candidate parameter sequence and its corresponding fitness value determined in each iteration are added to the training samples to update the Gaussian process surrogate model.
6. An adaptive control device for a purification process based on transfer learning, characterized in that, The device includes: Data acquisition module: Acquires process characteristic parameters of the source purification control model and the target purification control model, which are used as source domain data and target domain data, respectively; and acquires process parameters under the current operating conditions; Migration module: Performs feature parameter migration on source domain data and target domain data to obtain a sequence of migration process feature parameters; Purity prediction module: After training the purification control model by transferring the process feature parameter sequence, the process parameters under the current operating conditions are input into the purification control model for processing, and the purity prediction value is output. Optimization module: If the predicted purity value deviates from the preset purity threshold, the migration process characteristic parameter sequence is optimized to obtain an optimized process characteristic parameter sequence.
7. The adaptive control device for a purification process based on transfer learning according to claim 6, characterized in that, The migration module includes: a matrix transformation module, a merging module, a maximum mean difference matrix module, a migration transformation matrix module, and a matrix projection module. The matrix transformation module is used to perform matrix transformation on source domain data and target domain data to obtain source domain matrix and target domain matrix; The merging module is used to merge the source domain matrix and the target domain matrix to obtain a joint data matrix; The maximum mean difference matrix module is used to construct a maximum mean difference matrix based on the number of samples of the source domain data and the target domain data. The maximum mean difference matrix is used to characterize the distribution difference distance between the source domain data and the target domain data in the kernel space. The migration transformation matrix module is used to construct a generalized eigenvalue decomposition equation based on the joint data matrix and the maximum mean difference matrix to obtain the migration transformation matrix. The matrix projection module is used to project the target domain data onto a common feature space through the migration transformation matrix to obtain a migration process feature parameter sequence.
8. The adaptive control device for a purification process based on transfer learning according to claim 6, characterized in that, The purity prediction module is also used for: The migration process feature parameter sequence is transformed into a migration process feature sequence; The migration process feature sequence is continuously processed through a convolution module to obtain the migration convolution feature sequence; The migration convolution feature sequence is substituted into the purification control module for processing to obtain the purified feature sequence; The purified feature sequence is processed through a fully connected module, a convolution module, and an activation module to obtain a purity prediction value.
9. The adaptive control device for a purification process based on transfer learning according to claim 6, characterized in that, The optimization module is also used for: Using the migration process characteristic parameter sequence as the initial solution, and minimizing the deviation between the purity prediction value and the preset purity threshold as the fitness function, the fitness value is obtained. A first deviation threshold and a second deviation threshold are preset, wherein the first deviation threshold is greater than the second deviation threshold; Using the initial solution as the current optimal solution, the purity prediction value corresponding to the current optimal solution is obtained as the current optimal purity prediction value; Calculate the deviation between the current optimal purity prediction value and the preset purity threshold; When the deviation is greater than the first deviation threshold, a candidate parameter sequence is determined in the global search space; when the first deviation threshold is less than or equal to the second deviation threshold, a candidate parameter sequence is determined in the neighborhood of the current optimal solution; when the deviation is less than the second deviation threshold, the iteration is terminated, and the current optimal solution is output as the optimized process feature parameter sequence. The candidate parameter sequence determined in each iteration is input into the purification control model to obtain the corresponding purity prediction value. If the corresponding purity prediction value is better than the current optimal purity prediction value, the current optimal solution is updated with the candidate parameter sequence, and the current optimal purity prediction value is updated with the corresponding purity prediction value. Repeat the above iterative process until the preset maximum number of iterations is exceeded, at which point the iteration terminates.
10. The adaptive control device for a purification process based on transfer learning according to claim 9, characterized in that, The steps of determining the candidate parameter sequence in the global search space and determining the candidate parameter sequence in the neighborhood of the current optimal solution include: Using the migration process feature parameter sequence and historically determined candidate parameter sequences as training samples, and the fitness value corresponding to each candidate parameter sequence as the training target, a Gaussian process surrogate model is constructed. The Gaussian process surrogate model is used to characterize the mapping relationship between the process feature parameter sequence and the fitness value. In global search mode, the candidate parameter sequence is determined with the goal of minimizing the predicted mean output by the Gaussian process surrogate model; In the local development mode, a local search space is constructed with the current optimal solution as the center and a preset neighborhood radius as the boundary. Several parameter sequences are randomly sampled in the local search space. Each parameter sequence is input into the Gaussian process surrogate model to obtain the corresponding prediction mean. The parameter sequence with the smallest prediction mean is selected as the candidate parameter sequence. The candidate parameter sequence and its corresponding fitness value determined in each iteration are added to the training samples to update the Gaussian process surrogate model.