Multi-parameter-based intelligent control method for chromatographic separation of xylose mother liquor

Abstract

The application belongs to the technical field of intelligent control, and particularly relates to a kind of intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters. A chromatographic separation multi-parameter multi-objective optimization control model is established, with feed flow rate, elution water flow rate, column temperature, feed concentration and switching time as key operating parameters, and with xylose purity and yield maximization as optimization objectives. A layered Sobol quasi-random sequence is combined with a constraint perception repair strategy and dynamic reverse learning to initialize a population. A leader-driven particle swarm optimization algorithm is improved by a role self-adaptive multi-leader mechanism, a trigonometric function enhanced position update, a sensitivity guided dimensional selective update and a hyper-volume external archive maintenance to obtain a Pareto optimal control parameter solution set and issue an execution. The application can realize online intelligent regulation and control of the chromatographic separation process of xylose mother liquor, effectively improve product purity and yield, reduce energy consumption and adapt to fluctuations in raw material components.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and particularly to an intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters. Background Technology

[0002] Xylose mother liquor obtained from the hydrolysis of agricultural and forestry byproducts such as corn cobs and straw contains various monosaccharides with similar properties, including xylose, arabinose, and glucose. Chromatographic separation is the mainstream industrial method for purifying high-purity xylose, often using simulated moving bed continuous chromatography equipment. The core control parameters include five types of variables: feed flow rate, elution water flow rate, column temperature, feed concentration, and valve switching time. Process optimization needs to simultaneously maximize xylose purity and product yield, representing a typical multi-constraint, multi-objective coupled optimization scenario. The current mainstream control mode in the industry relies on manual trial and error to calibrate parameters. Operators adjust the flow rate, temperature, and other operating conditions one by one based on offline detection results, which has drawbacks such as strong lag, long debugging cycles, and high elution water and energy consumption. When the composition and concentration of the mother liquor raw materials fluctuate continuously, fixed parameters are prone to problems such as substandard purity or significant xylose loss. Existing traditional offline simulation optimization can only provide static optimal operating conditions and cannot be connected to the on-site chromatographic actuator to achieve online automatic control. Moreover, the multi-component separation variables of xylose mother liquor are highly coupled and different parameters have significantly different sensitivities to separation effects. Conventional uniform optimization strategies have low iteration efficiency and cannot take into account both global exploration and local fine optimization, which has long restricted the yield and production cost of xylose production. Summary of the Invention

[0003] To address the technical problems existing in the background art described above, this invention proposes a multi-parameter-based intelligent control method for xylose mother liquor chromatographic separation.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] S1. Based on the key operating parameters and separation quality objectives in the chromatographic separation process of xylose mother liquor, establish a multi-parameter, multi-objective optimization control model for chromatographic separation;

[0006] S2. Initialize the particle population using a hierarchical Sobol quasi-random sequence combined with a constraint-aware repair strategy, and generate an enhanced population through dynamic reverse learning. Select the best from the original population and the reverse population to form the initial iterative population.

[0007] S3. Improvements to the leader-driven particle swarm optimization algorithm include: introducing a role-adaptive multi-leader mechanism, dynamically dividing leader particles into three roles—reconnaissance leader, anchoring leader, and bridging leader—based on the fitness-distance balance index, with the number of each role adaptively adjusted according to the population convergence state; employing a trigonometric function-enhanced position update strategy for leaders of different roles, with different rotation angle update rules for different roles; introducing a sensitivity-guided dimension-selective update mechanism, using differentiated update step sizes for high-sensitivity and low-sensitivity dimensions based on the sensitivity differences of each decision variable dimension to the objective function; and using an external archive maintenance strategy based on the hypervolume index to store and manage non-dominated solutions, while ensuring the uniformity of solution set distribution in the target space through adaptive grid partitioning.

[0008] S4. The improved leader-driven particle swarm optimization algorithm is used to solve the multi-parameter, multi-objective optimization control model of chromatographic separation, and the Pareto optimal control parameter solution set is obtained. The optimal solution that meets the actual production preference is selected from the solution set, and the corresponding control parameter adjustment command is sent to the actuator of the chromatographic separation system to realize intelligent control of the xylose mother liquor chromatographic separation process.

[0009] Preferably, the key operating parameters in step S1 include feed flow rate, washing water flow rate, column temperature, feed concentration, and switching time. The optimization objectives include maximizing the purity of xylose products and maximizing the xylose yield. The constraints include the physical feasible range of the key operating parameters and the lower limit of the xylose product quality indicators.

[0010] As a preferred option, the specific steps for initializing the particle population in step S2 are as follows:

[0011] Based on the key operational parameter variable dimension m and the population size N, an m-dimensional Sobol quasi-random sequence is generated, producing N initial particle positions;

[0012] For particles that do not meet the constraints, a repair strategy based on the constraint violation gradient direction is adopted, which projects the infeasible particles to the nearest feasible domain boundary along the direction of the fastest decrease in constraint violation.

[0013] Dynamic back-learning is applied to the repaired population, and the offset weights of the back-learning points are determined based on the variance of the current population fitness distribution; the back-learning is... ,in, Let be the position of the reverse direction of the i-th particle. For weights in dynamic back-learning, These are the lower and upper bounds of the search space, respectively. Let be the current position of the i-th particle in the search space;

[0014] The original population and the reverse population are merged, and the top N particles are selected according to fitness to form the initial iterative population.

[0015] As a preferred option, the specific steps for introducing the role-adaptive multi-leader mechanism in step S3 are as follows:

[0016] Calculate the fitness-distance balance index for each particle in the population. The calculation method is as follows: ,in, Let i be the normalized rank of particle i in the fitness ranking. Let be the normalized ranking of particle i from its nearest neighbor particle in the decision space using Euclidean distance. This is the balance coefficient;

[0017] In each iteration, the total number of leaders is set to... All particles in the population are sorted according to Sort in ascending order and select the first one. The particles in position constitute the leader set;

[0018] The leader particle is divided into three roles: reconnaissance leader, anchoring leader, and bridging leader.

[0019] Anchored Leader: Sort the particles in the leader set from best to worst fitness value, and select the top-ranked ones. The particles in the position guide the population to carry out localized and refined development in high-quality areas;

[0020] Bridging Leader: After sorting the non-dominated solutions on the current Pareto front in ascending order along the first objective function value in the objective space, divide them equally into... Each segment contains the same number of solutions. The particle closest to the geometric center of the segment and belonging to the leader set is selected from each segment as the bridging leader of that segment, thus maintaining the diversity of the solution set distribution in the target space.

[0021] Scout Leader: The remaining particles in the leader set that are not included in the anchored leader and bridging leader are regarded as scout leaders, guiding the population to conduct global exploration;

[0022] The number of leaders in each role is based on population diversity indicators. Adjustment: When When the diversity threshold is exceeded, the particle with the worst fitness among the anchored leaders is reclassified as a scout leader; when When the diversity threshold is lower than the threshold, the particle with the best FDB value among the reconnaissance leaders is reclassified as an anchored leader; when the maximum normalized Euclidean distance between adjacent non-dominated solutions on the Pareto front is greater than the preset gap threshold, the number of bridging leaders is increased by 1, while the number of reconnaissance leaders is decreased by 1, ensuring that the total number of leaders remains unchanged.

[0023] Preferably, in step S3, a trigonometric function-enhanced position update strategy is adopted for leaders with different roles. Different roles correspond to different rotation angle update rules, specifically including:

[0024] For reconnaissance-type leaders, rotation angle ,in, The preset maximum rotation angle has a value range of [value range missing]. t is the current iteration number, and T is the total number of iterations;

[0025] For anchored leaders, rotation angle ,in, The preset minimum rotation angle has a range of values. ;

[0026] For bridging leaders, rotation angle ;

[0027] Then, the particle position is updated using trigonometric functions: ,in, Let i be the updated position of the i-th particle. Let i be the position of the i-th particle before the update. Let r be the position vector of the particle guiding the direction of reference, and r be a random number uniformly distributed in [0,1]. Let be the position vectors of two different particles randomly selected from the population. The following rules apply: for a reconnaissance leader, the position of the particle with the best fitness in the current iteration is selected; for an anchoring leader, the position of the particle with the best fitness in the neighborhood of the decision space with the particle as the center and a preset neighborhood radius is selected; for a bridging leader, the position vector corresponding to the non-dominated solution in the external archive that has the closest Euclidean distance to the particle in the target space is selected.

[0028] As a preferred option, the specific steps for introducing the sensitivity-guided dimension-selective update mechanism in step S3 are as follows:

[0029] In the early iteration stage of the algorithm, small positive perturbations are applied to each decision variable dimension in turn, the change magnitude of all objective functions before and after the perturbation is statistically analyzed, and the sensitivity index corresponding to each dimension is calculated.

[0030] Calculate the median of all sensitivity indices and use this median as the dividing line to separate the high-sensitivity dimension set from the low-sensitivity dimension set.

[0031] During particle position update, the high-sensitivity dimension uses the baseline update step size, while the low-sensitivity dimension uses an amplified update step size; the corresponding dimension update magnitude is performed in the position update operation.

[0032] Every fixed number of iterations, the sensitivity of each dimension is recalculated, and the results of the high and low sensitivity dimension division are updated.

[0033] As a preferred option, the specific steps of adopting the external archive maintenance strategy based on the super-volume index in step S3 are as follows:

[0034] Store all non-dominated solutions obtained during the optimization process into an external archive.

[0035] If the number of solutions in the file exceeds the preset maximum capacity, calculate the decrease in the overall over-volume index after removing a single solution as the contribution of that solution to the removal.

[0036] The solution with the smallest contribution is removed in a loop until the number of solutions in the file does not exceed the maximum capacity.

[0037] Adaptive grid partitioning is implemented for external archives, and the number of grid partitions is determined based on the archive capacity. When there are multiple non-dominated solutions in the same grid area, only the solution with the highest hypervolume contribution is retained, and the other solutions are removed to ensure that the Pareto solution set is uniformly distributed in the target space.

[0038] A multi-parameter-based intelligent control system for xylose mother liquor chromatographic separation, characterized in that it includes:

[0039] The model building module establishes a multi-parameter, multi-objective optimization control model for chromatographic separation based on the key operating parameters and separation quality objectives of the xylose mother liquor chromatographic separation process.

[0040] The algorithm initialization module uses a hierarchical Sobol quasi-random sequence combined with a constraint-aware repair strategy to initialize the particle population, and generates an enhanced population through dynamic back learning.

[0041] The algorithm improvement module improves the leader-driven particle swarm optimization algorithm by: introducing a role-adaptive multi-leader mechanism, dynamically dividing leader particles into three roles: scout leader, anchor leader, and bridging leader; adopting a position update strategy enhanced by trigonometric functions for leaders of different roles; introducing a sensitivity-guided dimension-selective update mechanism; and adopting an external archive maintenance strategy based on hypervolume indices.

[0042] The control execution module uses an improved leader-driven particle swarm optimization algorithm to solve the multi-parameter, multi-objective optimization control model for chromatographic separation, and performs online parameter adjustment and intelligent control of the chromatographic separation system based on the solution results.

[0043] Compared with existing technologies, the advantages and positive effects of this invention are as follows: It employs Sobol sequences combined with constraint repair and dynamic back-learning to complete population initialization, resulting in a more uniform initial solution distribution and a higher proportion of feasible solutions; it designs three types of adaptive leader roles paired with differentiated trigonometric function rotation updates to dynamically balance global exploration and local fine-grained optimization; it utilizes parameter sensitivity to achieve step-by-step dimension updates, reducing computational waste; and it maintains external archives based on hypervolume indices, resulting in better Pareto solution set convergence and distribution uniformity. It can solve for optimal chromatographic conditions online and distribute the results to the actuators, responding in real-time to fluctuations in mother liquor raw materials, significantly shortening the debugging cycle, improving xylose purity and yield, and reducing elution water and production energy consumption. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram illustrating the implementation process of an intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters. Detailed Implementation

[0046] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0047] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0048] Example 1: Considering that in existing technologies for chromatographic separation of xylose mother liquor, the mainstream control mode in the industry has long relied on manual trial-and-error calibration of parameters. Operators adjust process conditions such as flow rate, temperature, and switching time based on offline detection results, which presents the following prominent problems: The raw material components and concentrations of xylose mother liquor fluctuate continuously due to the influence of upstream hydrolysis processes and batch differences in raw materials. The offline detection results relied upon for manual control are usually several hours or even a shift behind, resulting in parameter adjustments failing to respond in real time to changes in raw materials. During periods of fluctuation, economic losses are easily incurred due to substandard product purity (below the 98% industry access standard) or significant xylose loss into waste liquid. The debugging cycle is long, and there are strong coupling relationships among the five types of control parameters. For example, the feed flow rate and switching time jointly determine the retention time distribution within the chromatographic column, the column temperature affects the distribution coefficient of each component between the stationary and mobile phases, and the feed concentration affects the column load and resolution. The manual trial-and-error method is difficult to effectively decouple the nonlinear interaction effects between multidimensional parameters, and the parameter optimization process usually requires repeated experiments for several days to several weeks, seriously affecting production efficiency.

[0049] To address the above problems, this invention proposes a specific implementation method for intelligent control of xylose mother liquor chromatographic separation based on multiple parameters, comprising the following steps:

[0050] First, based on the key operating parameters and separation quality objectives in the chromatographic separation process of xylose mother liquor, a multi-parameter, multi-objective optimization control model for chromatographic separation is established. Specifically, the key operating parameters include feed flow rate, elution water flow rate, column temperature, feed concentration, and switching time. The optimization objectives include maximizing the purity of xylose products and maximizing the xylose yield. The constraints include the physical feasible range of the key operating parameters and the lower limit of the quality indicators of xylose products.

[0051] The feed flow rate is defined as the volumetric flow rate of the xylose mother liquor feed solution delivered to the chromatographic separation system via the feed pump, and its physical feasible range is limited. Too low a feed flow rate will lead to insufficient equipment capacity and increased fixed costs per ton of product; too high a feed flow rate will cause column overload, decreased resolution, and increased overlap of chromatographic peaks between xylose and impurity sugars. The eluent flow rate is defined as the volumetric flow rate of the eluent pure water delivered to the chromatographic separation system via the eluent pump, and its physical feasible range is set. The eluent flow rate directly affects the linear velocity of the mobile phase within the column and the elution order and resolution of each component. A low eluent flow rate will lead to incomplete elution and residual impurities in the xylose product; a high eluent flow rate will lead to excessive dilution of xylose. The column temperature is defined as the column bed temperature maintained by the column jacket circulating heating system, and its physical feasible range is set. The column temperature affects the partition coefficient (Henry constant) of sugar molecules between the stationary phase (ion exchange resin) and the mobile phase, thus affecting the retention time differences of each component. At excessively low temperatures, the distribution coefficients of each component are small, resulting in insufficient separation; at excessively high temperatures, the resin may undergo thermal degradation, shortening its service life. The feed concentration is defined as the mass fraction of total dissolved solids in the xylose mother liquor feed solution, and a physically feasible range is set. The feed concentration is determined by the upstream hydrolysis process but can be monitored in real time using an online refractive index detector. Excessively high feed concentrations increase column load and decrease separation efficiency; excessively low feed concentrations reduce the amount of xylose entering the system per unit time. The switching time is defined as the time interval (i.e., valve switching cycle) during which the inlet and outlet of the SMB device move one column position along the flow direction, and a physically feasible range is set. The switching time determines the effective utilization period of the column and the matching relationship between the migration rate of each component and the port migration rate. If the switching time is too short, xylose is switched to the next region before being fully eluted, resulting in high impurity levels in the product; if the switching time is too long, xylose reaches the waste outlet prematurely, leading to yield loss. The aforementioned key operating parameters need to be normalized before being input into the model to address dimensional inconsistencies.

[0052] This optimization model comprises two mutually constraining objectives: Objective 1 is to maximize xylose product purity, defined as the percentage of xylose mass in the chromatographically separated product stream relative to the total dissolved solids; Objective 2 is to maximize xylose yield, defined as the percentage of recovered xylose mass in the chromatographically separated product stream relative to the total xylose mass in the feed. In actual industrial operation, there is a natural constraint between purity and yield. Improving purity usually requires stricter separation conditions (such as higher elution water volume or shorter switching time), which can lead to some xylose loss into the waste liquid, thus reducing the yield. Therefore, this optimization model adopts a multi-objective optimization framework, outputting a Pareto optimal solution set, providing production decision-makers with a global perspective on the purity-yield tradeoff. Constraints include the physical feasible range of key operating parameters and the lower limit of xylose product quality indicators. This indicator is the operational baseline set to ensure production economics. The two lower limits of quality indicators jointly define the boundary of the feasible solution domain, and the algorithm searches for Pareto optimal solutions only within the region that satisfies all constraints.

[0053] Then, a hierarchical Sobol quasi-random sequence combined with a constraint-aware repair strategy is used to initialize the particle population, and an enhanced population is generated through dynamic back-learning. The initial iterative population is then selected from the original population and the back-learning population. Specifically, considering that the initial particles generated by conventional random initialization methods (such as uniform random sampling) have poor uniformity in the search space, it is easy for particles to cluster in local areas while other areas are sparse or even empty, causing the algorithm to miss potential high-quality search areas in the early stages of iteration and reducing the reliability of global optimization. This invention uses a hierarchical Sobol quasi-random sequence to generate the initial particle positions. The Sobol sequence is a low-discrepancy quasi-random sequence, and the uniformity of the distribution of its generated sampling points in multidimensional space is significantly better than that of pseudo-random numbers. The core principle of the Sobol sequence is based on the bit-by-bit construction of binary digit sets (Generating Matrices), and the coordinate values ​​of each point in the sequence are recursively calculated using Gray Code, so that the marginal distribution and joint distribution of each dimension can approximately achieve theoretical uniform coverage under a limited sample size. The specific implementation process is as follows: first, the population size is set to N, and the dimension of the decision variable is m=5. The algorithm initialization module generates a Sobol quasi-random sequence of length N within an m-dimensional hypercube. When generating the Sobol sequence, the Bratley-Fox algorithm is used to construct the direction numbers, which are obtained recursively from pre-computed primitive polynomials. The components of each dimension at each sampling point in the Sobol sequence are linearly mapped from the [0,1] interval to the physically feasible range of the corresponding decision variable to obtain the initial particle positions. Each particle corresponds to a complete set of chromatographic separation operation parameters.

[0054] Considering that the initial particle positions generated by the Sobol sequence only guarantee a uniform distribution within the physically feasible range of the decision variables, but do not consider the satisfaction of the constraints: purity lower bound and yield lower bound, some particle process parameter combinations may not enable the separation process to meet performance requirements, belonging to constrained infeasible solutions. If infeasible particles are directly included in the initial population, the algorithm will waste a lot of resources processing infeasible regions in the early stages of iteration, reducing optimization efficiency. Therefore, this invention adopts a repair strategy based on the constraint violation gradient direction, projecting infeasible particles to the nearest feasible region boundary along the direction of the fastest decrease in constraint violation degree. The specific implementation process is as follows: For each initial particle, the model is used to evaluate its corresponding purity value and yield value, and the satisfaction of all constraints is checked one by one. The constraint violation degree function is defined as the sum of the violation amounts of all violated constraints, that is, for each constraint condition, if its left-hand expression is greater than zero, the value is taken as the violation amount, otherwise it is taken as zero, and finally all violation amounts are summed. When the constraint violation degree of a particle is greater than zero, the particle is confirmed as an infeasible particle and a repair operation needs to be performed. The repair direction is determined by numerical gradient calculation. For each violated constraint, its partial derivative with respect to each decision variable dimension is calculated. The partial derivative is approximated using the forward difference method: a small perturbation step size is applied to a dimension at the current particle position, which is 0.5% of the width of the physically feasible range in that dimension. The difference between the constraint function values ​​before and after the perturbation is then calculated and divided by the perturbation step size to obtain an approximate value of the partial derivative in that dimension. The repair direction is the opposite direction of the weighted sum of the gradients of all violated constraints, where the weight of each constraint gradient is the constraint violation amount divided by the square of the gradient magnitude plus a small constant to prevent division by zero. This weighting strategy ensures that constraints with greater violation contribute more to the repair direction, while avoiding numerical instability caused by excessively large or small gradient magnitudes. After determining the repair direction, the particle position is moved along that direction with a certain step size. The step size is determined through line search: starting from the current particle position, the particle moves gradually along the repair direction. After each move, all constraints are checked, and the smallest non-negative step size that satisfies all constraints is taken as the final movement distance. The repaired particle position is equal to the original particle position plus the step size multiplied by the repair direction vector. If a component of a repaired particle exceeds the physically feasible range in any dimension, it is clamped to the nearest boundary value. The repair process is executed iteratively until the constraint violation drops below a minimum threshold or the preset maximum number of repairs is reached. If no feasible solution can be found after the maximum number of repairs, the particle is repositioned to the vicinity of any feasible particle position found in the search space, with the offset controlled within five percent of the width of each dimension, to ensure that the initial population does not contain infeasible solutions.

[0055] After constraint repair, all particles in the population represent feasible solutions, but their positional distribution remains constrained by the deterministic patterns of the Sobol sequence, potentially overlooking some high-quality regions in the search space. To expand the initial search coverage, enhance population diversity, and increase the probability of discovering high-quality solutions in the early stages of iteration, this step applies dynamic back-learning to the repaired population, generating a back-enhanced population. The core idea of ​​dynamic back-learning is based on the observation that for a known candidate solution in the search space, its back-point, which is symmetric about the center of the search space, has a higher probability of being closer to the global optimum than the original candidate solution. By simultaneously evaluating the quality of the original and back-point solutions and selectively retaining the best, the initial search range of the algorithm can be effectively expanded. Unlike standard back-learning, which uses a fixed symmetry center point, dynamic back-learning introduces an adjustable offset weight, adaptively controlling the generation strategy of the back-point based on the fitness distribution characteristics of the current population. Specifically, for each repaired particle, the variance of the objective function fitness values ​​of all particles in the current population is first calculated. Fitness values ​​are calculated using a weighted comprehensive score method, which is the purity value multiplied by its purity weight plus the yield value multiplied by its yield weight. The purity weight is set to 0.6, and the yield weight to 0.4. The variance of the comprehensive fitness values ​​of all particles in the population is calculated and normalized by dividing it by the square of the difference between the maximum and minimum fitness values, yielding the normalized fitness variance. The weights for dynamic back-learning are then derived from this normalized variance value. The value is clamped to between 0.1 and 0.8. When the population fitness variance is small, it indicates that the fitness of particles in the population tends to be consistent, and the search diversity is insufficient. At this time, the weight approaches 0.8, and the reverse point is pushed to the boundary region far away from the center of the search space, tending to generate reverse points far away from the current population gathering area to enhance the global exploration capability. When the population fitness variance is large, it indicates that the population fitness is dispersed and the diversity is good. At this time, the weight approaches 0.5, and the reverse point degenerates into a standard reverse point, that is, a symmetrical point about the center of the search space, maintaining appropriate exploration coverage without introducing excessive offset perturbation. Using the calculated weight, a reverse point is generated for each particle. The calculation formula for the reverse point is: The reverse learning is ,in, Let be the position of the reverse direction of the i-th particle. For weights in dynamic back-learning, These are the lower and upper bounds of the search space, respectively. Let be the current position of the i-th particle in the search space. The lower and upper bounds of the search space are taken as zero (lower bound) and one (upper bound) of the physically feasible range of each decision variable, respectively. This calculation is performed in the normalized space. For each generated reverse point, check if its dimensional components are within the range of zero to one in the normalized space. If they are outside this range, clamp them to the nearest boundary value. Subsequently, perform the same constraint-aware repair strategy as the previous step on all reverse points to ensure that all particles in the reverse population satisfy all constraints.

[0056] After generating a reverse population through dynamic back learning, the system simultaneously possesses two particle populations, each of size N: the original population and the reverse population. The original population is the set of feasible solutions generated by Sobol sequence generation and constraint repair, with its particle positions uniformly distributed in the search space. The reverse population is the set of feasible solutions transformed by dynamic back learning and constraint repair, with its particle positions biased towards symmetrical positions far from the original population's aggregation region in the search space. The two populations are merged into a candidate population of size 2N. From the candidate population, N particles are selected based on their fitness to form the final initial iteration population, ensuring that the quality of the initial population entering the iterative optimization stage is as high as possible. The specific implementation process is as follows: The N particles from the original population and the N particles from the reverse population are merged into a candidate set containing 2N particles. For all particles in the candidate set, the xylose product purity and xylose yield are evaluated using a chromatographic separation multi-parameter multi-objective optimization control model, and the satisfaction of all constraints is checked. Since both the original and reverse populations have undergone constraint repair in the preceding steps, all particles in the candidate set theoretically satisfy the constraints. The secondary check here is used to eliminate minute violations that may remain due to numerical precision issues during the repair process. If any individual particles still exhibit minor violations, the repair is performed again. A comprehensive fitness score is calculated for the candidate particles that satisfy the constraints. A higher comprehensive fitness score indicates better overall performance in terms of purity and yield for the corresponding process parameters. All 2N particles in the candidate set are sorted in descending order of their comprehensive fitness scores. The top N particles after sorting are selected to form the final initial iterative population. Since the original population generated by the Sobol sequence and the reverse population generated by dynamic reverse learning occupy complementary spatial regions in the search space, the initial iterative population selected after merging has two advantages: on the one hand, it retains particles with excellent fitness and uniform spatial distribution in the original population, ensuring that the initial population basically covers the search space; on the other hand, it incorporates particles with better fitness in the reverse population, replacing the positions with poor fitness but valuable spatial distribution in the original population with higher quality reverse solutions, thereby improving the overall quality of the initial population.

[0057] Next, an improved leader-driven particle swarm optimization (PSO) algorithm is used to solve the problem. However, existing standard PSO and multi-objective PSO algorithms typically rely on a single globally optimal individual or a randomly selected few leaders to guide the population search. When dealing with complex multi-objective optimization problems such as xylose mother liquor chromatography separation with five decision variables, two coupled objectives, and multiple constraints, the single-leader structure suffers from the following problems: all particles converge in the same direction, population diversity is rapidly lost, and it is prone to getting trapped in local Pareto fronts; it cannot simultaneously address the fine-grained search of different regions of the Pareto front, resulting in uneven front distribution; and the algorithm lacks a regulatory mechanism for exploration and development capabilities, leading to insufficient exploration in the early stages of iteration (premature convergence) or insufficient development in the later stages of iteration (oscillatory non-convergence). This merge-selection strategy is essentially a population enhancement method based on an elite retention mechanism. To address these problems, this invention introduces a role-adaptive multi-leader mechanism, dividing leader particles into three roles based on their positional characteristics and functional roles within the population. These roles cooperate and complement each other, and the number of roles is adaptively adjusted according to the population convergence status. Specifically, the fitness-distance balance index is calculated for each particle in the population. The calculation method is as follows: ,in, Let i be the normalized rank of particle i in the fitness ranking. Let be the normalized ranking of particle i from its nearest neighbor particle in the decision space using Euclidean distance. This is a balance coefficient, ranging from 0.4 to 0.7. In each iteration, the total number of leaders is set to... All particles in the population are sorted according to Sort in ascending order and select the first one. The leaders are comprised of particles of a certain rank. These leaders are categorized into three roles: scout leaders, anchoring leaders, and bridging leaders. Anchoring leaders are defined as those who perform a fine-grained local search within a region of discovered high-quality solutions, guiding the population to precise convergence toward the Pareto front. The selection rule involves ranking the particles in the leader set from best to worst fitness value, and selecting the top-ranked particles. The particles at position guide the population to conduct localized, refined development in high-quality regions. If the value is not an integer, it needs to be rounded down. All subsequent operations requiring integer rounding in this scheme will use this rounding down method. A bridging leader is defined as maintaining the uniformity of the Pareto front distribution in the target space and connecting regions with large gaps on the Pareto front. The selection rule is to sort the non-dominated solutions on the current Pareto front in ascending order along the first objective function value (xylose purity) in the target space, and then divide them into equal parts. Each segment contains the same number of solutions. From each segment, the particle closest to the geometric center and belonging to the leader set is selected as the bridging leader for that segment, maintaining the diversity of the solution distribution in the target space. If a segment has no particles belonging to the leader set, the non-dominated solution closest to the geometric center of that segment (even if it is not in the leader set) is selected as a bridging leader candidate and added to the leader set, replacing the worst-fit particle in the current set. A scout leader is defined as someone who guides the population to conduct a global exploration, searching for new regions not covered by the current population, preventing the algorithm from getting trapped in local optima. The selection rule is to consider the remaining particles in the leader set that are not included in the anchored or bridging leaders as scout leaders, guiding the population to conduct a global exploration.

[0058] The number of leaders in the three roles is not fixed throughout the optimization process, but rather depends on the diversity index of the population convergence state. Adaptive adjustments are made to the diversity index. The calculation method involves, for each particle in the current iteration population, first calculating the Euclidean distance from that particle to its nearest neighbor particle in the normalized decision space. Since the physical dimensions and numerical magnitudes of the five decision variables differ, to eliminate the influence of dimensional differences on the distance metric, the distance calculation is performed in the normalized space. For a given particle, the Euclidean distances between it and all other particles in the population are calculated one by one, and the minimum value is taken as the nearest neighbor distance. The Euclidean distance is calculated as follows: the normalized position components of the two particles in each dimension are subtracted, the squares of the differences are summed, and the square root of the sum is taken. This nearest neighbor distance reflects the spatial distance between the particle and its nearest neighbor in the decision space; a larger distance indicates a more isolated particle and a greater contribution to population diversity. Subsequently, the nearest neighbor distance of each particle is standardized by dividing it by the diagonal length of the normalized search space. The normalized search space is a five-dimensional unit hypercube mapped to the interval between zero and one for each dimension, with its diagonal length equal to the square root of the dimension number five. The standardized nearest neighbor distance ranges from zero to one, eliminating the influence of the number of dimensions on the absolute value of the distance metric and making the index comparable across problems of different dimensions. Finally, the arithmetic mean of the standardized nearest neighbor distances for all N particles in the population is used to obtain the population diversity index. When the diversity threshold is exceeded, the particle with the worst fitness among the anchored leaders is reclassified as a scout leader; when When the diversity threshold is lower than the threshold, the particle with the best FDB value among the reconnaissance leaders is reclassified as an anchored leader; when the maximum normalized Euclidean distance between adjacent non-dominated solutions on the Pareto front is greater than the preset gap threshold, the number of bridging leaders is increased by 1, while the number of reconnaissance leaders is decreased by 1, ensuring that the total number of leaders remains unchanged.

[0059] Next, considering that the position update formula of the particle swarm optimization algorithm in the prior art usually uses linear inertia weights or linear compression factors to control the search step size, the following problems exist: Linearly decaying inertia weights cause the search step size to decrease uniformly with the number of iterations, failing to reflect the stage characteristics of the search process. In the middle stage of the algorithm (when the population has initially located near the Pareto front but has not yet precisely converged), it is necessary to strengthen local development while maintaining a certain level of exploration capability, but linear decay cannot provide this non-linear balance adjustment; furthermore, all particles (regardless of their role) use the same position update rule, failing to reflect the different needs of different leaders in search behavior. To address these problems, this invention adopts a position update strategy enhanced with trigonometric functions. Different leaders correspond to different rotation angle update rules, ensuring that each role maintains a search behavior pattern consistent with its functional role throughout the entire iteration process. Specifically, for scout leaders, the rotation angle... ,in, The preset maximum rotation angle has a value range of [value range missing]. t is the current iteration number, and T is the total number of iterations. For anchored leaders, the rotation angle... ,in, The preset minimum rotation angle has a range of values. For bridging leaders, the rotation angle Then, the particle positions are updated using trigonometric functions: ,in, , Let i be the updated position of the i-th particle. Let i be the position of the i-th particle before the update. Let r be the position vector of the particle guiding the direction of reference, and r be a random number uniformly distributed in [0,1]. Let be the position vectors of two different particles randomly selected from the population. The following rules apply: for a reconnaissance leader, the position of the particle with the best fitness in the current iteration is selected; for an anchoring leader, the position of the particle with the best fitness in the neighborhood of the decision space with the particle as the center and a preset neighborhood radius is selected; for a bridging leader, the position vector corresponding to the non-dominated solution in the external archive that has the closest Euclidean distance to the particle in the target space is selected.

[0060] Furthermore, the standard particle swarm optimization algorithm in existing technologies does not consider the differences in the degree of influence of different dimensions on the objective function when updating positions. In the xylose mother liquor chromatographic separation optimization problem, the five types of decision variables have significantly different sensitivities to the separation effect. For example, the sensitivity of valve switching time to xylose purity is usually much higher than that of feed concentration, because valve switching time directly determines the effective separation time window of the chromatographic column, while feed concentration mainly affects column load, and its impact on purity is relatively small when the column load is not overloaded. If the same update step size is used for all dimensions, the algorithm will waste a lot of iterative resources for fine-tuning in low-sensitivity dimensions, while the search accuracy in high-sensitivity dimensions may be insufficient. Therefore, this invention introduces a sensitivity-guided dimension-selective update mechanism. In the early stage of the algorithm, the sensitivity index of each decision variable dimension to the objective function is quantitatively evaluated, and the dimensions are divided into two categories: high sensitivity and low sensitivity. Different update step sizes are used for the two categories of dimensions, so that limited computing resources are preferentially allocated to the parameter dimensions that have the greatest impact on the optimization objective.

[0061] Specifically, in the initial iteration phase of the algorithm, this embodiment is set to execute within the fifth iteration. In this phase, the particle with the best overall fitness in the current population is taken as the baseline solution. This baseline solution contains normalized positional components in five dimensions, corresponding to feed flow rate, eluent flow rate, column temperature, feed concentration, and switching time, respectively. Sensitivity evaluation is performed sequentially for each decision variable dimension of the baseline solution. For each dimension, a small positive perturbation is applied to its normalized positional value. The perturbed solution is substituted into the multi-parameter, multi-objective optimization control model for chromatographic separation to calculate the xylose product purity and xylose yield after perturbation. The changes in the two objective functions after perturbation are subtracted from the objective function values ​​corresponding to the baseline solution before perturbation to obtain the changes in purity and yield. The squares of the two changes are summed, and the square root of the sum is taken to obtain the overall change in the objective function. This overall change is divided by the normalized perturbation to obtain the sensitivity index for each dimension. The physical meaning of the sensitivity index is: the combined change in two objective functions caused by a unit normalized perturbation in a certain dimension. The larger the sensitivity index value, the more significant the impact of a small change in that dimension on the optimization objective, and the more refined the handling of that dimension is required in the optimization search.

[0062] After obtaining the sensitivity indices for all five decision variable dimensions, the dimensions are divided into a high-sensitivity dimension set and a low-sensitivity dimension set, using the median of all dimension sensitivity indices as the dividing line. This allows for differentiated search strategies for the two types of dimensions during subsequent position updates. The specific implementation process is as follows: The calculated sensitivity index values ​​for the five dimensions are sorted in ascending order from smallest to largest. Since the total number of dimensions is odd, the median is taken as the sensitivity index value in the middle position (i.e., the third position) after sorting. If the total number of dimensions is even, the median is taken as the arithmetic mean of the two middle values ​​after sorting. This median is established as the sensitivity dividing threshold. Subsequently, the sensitivity index of each dimension is compared with this dividing threshold. Dimensions with sensitivity indices greater than or equal to the dividing threshold are assigned to the high-sensitivity dimension set, and dimensions with sensitivity indices less than the dividing threshold are assigned to the low-sensitivity dimension set.

[0063] During particle position updates, different update step sizes and amplitudes are used for high-sensitivity and low-sensitivity dimensions, respectively. This allows limited iterative computational resources to be prioritized for the parameter dimensions that have the greatest impact on the optimization objective, improving the overall optimization efficiency of the algorithm. The specific implementation process is as follows: When performing trigonometric function-enhanced position update operations on each particle in the population, the sensitivity category of the particle in each dimension is first determined. For dimensions in the high-sensitivity dimension set, the baseline position update formula is directly applied for updating. For dimensions in the low-sensitivity dimension set, the baseline update step size is multiplied by a preset step size amplification factor. The amplification factor ranges from 1.5 to 3; in this embodiment, it is set to 2. The amplified update step size increases the single movement amplitude of the particle in the low-sensitivity dimension to twice that in the high-sensitivity dimension, thus covering the effective value range of that dimension with fewer iterations. After each dimension update is completed, it is checked whether the updated position component exceeds the zero-to-one range of the normalized space. If it does, it reverts to the nearest boundary value. Clamping ensures that the particle positions in all dimensions are always within the normalized interval corresponding to the physically feasible range, avoiding particle out-of-bounds errors caused by step size amplification. The physical significance of this differentiated update strategy lies in the following: Small changes in high-sensitivity dimensions can elicit a significant response from the objective function; therefore, a small, baseline step size is used for fine-grained searching to accurately locate the optimal value in that dimension, preventing repeated oscillations around the optimal value due to excessively large step sizes. Conversely, larger changes in low-sensitivity dimensions are required to elicit an observable response from the objective function; therefore, a larger step size is used for rapid searching, avoiding excessive iterations in that dimension and allocating computational resources to high-sensitivity dimensions. Finally, the sensitivity relationships of each dimension to the objective function are not fixed throughout the optimization process but dynamically change as the population moves within the search space and the Pareto front gradually approaches. Therefore, every fixed number of iterations, the sensitivity of each dimension is recalculated, and the high- and low-sensitivity dimension partitioning results are updated.

[0064] Furthermore, in multi-objective optimization, an external archive is needed to store non-dominated solutions discovered during iterations to prevent high-quality solutions from being lost in subsequent population updates. Existing external archive maintenance strategies (such as pruning strategies based on crowding distance) have the following problems: crowding distance only measures the local density between adjacent solutions in the target space, lacking a direct measure of the overall convergence quality of the solution (i.e., the distance of the solution from the true Pareto front). This may lead to the retention of poorly converged but sparsely surrounded solutions while discarding well-converged but densely surrounded solutions. When the Pareto front is non-convex or has discontinuous regions, crowding distance can easily produce misleading density estimates at inflection points or discontinuities, leading to the incorrect removal of solutions in important regions. This invention uses the hypervolume (HV) index as the core evaluation criterion for external archive maintenance. The hypervolume index is defined as the size of the hypervolume enclosed by all non-dominated solutions in the external archive and a preset reference point in the target space. HV (Hyperdosing) is a known metric in multi-objective optimization that simultaneously measures convergence (the proximity of a solution to the true front) and diversity (the range of solutions distributed on the front). It possesses strict mathematical monotonicity, meaning that under Pareto dominance, the HV contribution of any solution dominated by a new solution is necessarily no greater than the HV contribution of the new solution. Specifically, at the end of each iteration of the improved leader-driven particle swarm optimization algorithm, newly discovered non-dominated solutions in the current population are stored in an external archive. The specific implementation process is as follows: For all particles in the current iteration population, their xylose product purity and xylose yield are evaluated using a chromatographic separation multi-parameter multi-objective optimization control model. Then, a non-dominated relationship determination is performed on the particles in the population: a particle is compared with all existing solutions in the population and the external archive for Pareto dominance. If the particle is not dominated by any existing solution, it is confirmed as a non-dominated solution. All newly discovered non-dominated solutions are added to the external archive. After joining, a non-dominated filtering operation is performed on the entire external archive. All solutions in the archive are compared pairwise for Pareto dominance, and solutions dominated by any other solution are removed, ensuring that all solutions in the archive are mutually non-dominated. Each solution in the filtered archive represents a point on the Pareto front in the target space that cannot be surpassed by any other solution simultaneously on two targets. The entire archive constitutes a complete description of the currently discovered Pareto fronts.

[0065] When the number of solutions in the external archive exceeds the preset maximum capacity, a pruning operation is required to control the archive size. First, a reference point is defined in the target space. The two components of the reference point are taken as the worst-case boundary values ​​for xylose purity and xylose yield, ensuring that the reference point is inferior to all feasible solutions. The overall hypervolume value of all non-dominated solutions in the current external archive relative to the reference point is calculated. In the two-dimensional target space, the hypervolume is calculated by sorting all solutions in the archive in ascending order of purity value, and then summing the areas of the rectangles formed by each solution and the reference point, minus the overlapping portions. Subsequently, each solution in the archive is temporarily removed one by one, and the hypervolume value of the remaining solution set after removal is calculated. The difference between the original overall hypervolume value and the removed hypervolume value is the removal contribution of that solution. The physical meaning of the removal contribution is the marginal contribution of the solution to the overall Pareto front coverage of the archive. Solutions with larger contributions are more important for supporting the convergence and diversity of the front and should be preferentially retained.

[0066] Among all solutions' removal contributions, the solution with the smallest contribution is selected for actual removal. This solution is permanently deleted from the external archive, reducing the archive size by one. Since the contributions of other solutions may change after removing one solution, the removal contributions of the remaining solutions need to be recalculated: that is, each remaining solution is temporarily removed one by one, and the decrease in the overall hypervolume value is recalculated. The solution with the smallest contribution is again selected for removal from the recalculated contributions. This iterative process is repeated until the number of solutions in the archive does not exceed the preset maximum capacity. The reason for using a greedy strategy of iteratively recalculating contributions instead of a one-time batch removal is that after each removal of a solution, the coverage area in the target space changes, and the marginal contributions of other solutions change accordingly. Iterative recalculation ensures that each removal minimizes the impact on the frontier quality under the current state, maximizing the preservation of the integrity and uniformity of the Pareto front.

[0067] Finally, the number of grid divisions is determined based on the current file capacity. The number of grid divisions is the square root of the smaller value between the file capacity and the preset maximum capacity, rounded down. When the maximum capacity is one hundred, the number of grid divisions is ten, meaning that the file is divided into ten intervals on both the purity axis and the yield axis, forming a ten-by-ten two-dimensional grid array. The grid boundaries are determined by the minimum and maximum values ​​of the objective function values ​​in the current file. The purity axis is evenly divided into ten segments from the minimum purity value to the maximum purity value, and the same applies to the yield axis. Each non-dominated solution in the file is assigned to its corresponding grid cell based on its purity and yield values. When multiple non-dominated solutions exist within the same grid cell, the local hypervolume contribution of each solution within that grid cell is calculated one by one. Only the solution with the highest contribution is retained, and the remaining solutions within that grid cell are discarded.

[0068] Finally, the improved leader-driven particle swarm optimization algorithm is used to iteratively solve the established multi-parameter, multi-objective optimization control model for chromatographic separation. After the iteration terminates, the solution with the highest fitness value in the Pareto optimal solution set from the external archive is selected as the optimal control scheme. The parameters in the optimal scheme are then inversely normalized to their physical parameter values, which are then converted into feed pump inverter frequency setpoints, elution water pump speed commands, column temperature control commands, and valve switching cycle setpoints. These are then transmitted to the various actuators of the chromatographic separation system via the industrial communication bus, achieving intelligent control of the xylose mother liquor chromatographic separation process.

[0069] Example 2: This invention provides a multi-parameter-based intelligent control system for xylose mother liquor chromatographic separation, comprising: a model building module, which establishes a multi-parameter, multi-objective optimization control model for chromatographic separation based on key operating parameters and separation quality objectives of the xylose mother liquor chromatographic separation process; an algorithm initialization module, which initializes the particle population using a hierarchical Sobol quasi-random sequence combined with a constraint-aware repair strategy, and generates an enhanced population through dynamic back-learning; an algorithm improvement module, which improves the leader-driven particle swarm optimization algorithm, including: introducing a role-adaptive multi-leader mechanism, dynamically dividing leader particles into three roles: scout leader, anchoring leader, and bridging leader; adopting a position update strategy enhanced by trigonometric functions for leaders of different roles; introducing a sensitivity-guided dimension-selective update mechanism; and adopting an external archive maintenance strategy based on hypervolume indices; and a control execution module, which uses the improved leader-driven particle swarm optimization algorithm to solve the multi-parameter, multi-objective optimization control model for chromatographic separation, and performs online parameter adjustment and intelligent control of the chromatographic separation system based on the solution results.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multi-parameter-based intelligent control method for chromatographic separation of xylose mother liquor, characterized in that, Includes the following steps: S1. Based on the key operating parameters and separation quality objectives in the chromatographic separation process of xylose mother liquor, establish a multi-parameter, multi-objective optimization control model for chromatographic separation; S2. Initialize the particle population using a hierarchical Sobol quasi-random sequence combined with a constraint-aware repair strategy, and generate an enhanced population through dynamic reverse learning. Select the best from the original population and the reverse population to form the initial iterative population. S3. Improvements to the leader-driven particle swarm optimization algorithm, specifically... This includes: introducing a role-adaptive multi-leader mechanism, dynamically classifying leader particles into three roles—reconnaissance leader, anchoring leader, and bridging leader—based on a fitness-distance balance index, with the number of each role adaptively adjusted according to the population convergence state; employing a trigonometric function-enhanced position update strategy for leaders of different roles, with different rotation angle update rules corresponding to different roles; introducing a sensitivity-guided dimension-selective update mechanism, using differentiated update step sizes for high-sensitivity and low-sensitivity dimensions based on the sensitivity differences of each decision variable dimension to the objective function; and using an external archive maintenance strategy based on a hypervolume index to store and manage non-dominated solutions, while ensuring the uniformity of solution set distribution in the target space through adaptive grid partitioning. S4. The improved leader-driven particle swarm optimization algorithm is used to solve the multi-parameter, multi-objective optimization control model of chromatographic separation, and the Pareto optimal control parameter solution set is obtained. The optimal solution that meets the actual production preference is selected from the solution set, and the corresponding control parameter adjustment command is sent to the actuator of the chromatographic separation system to realize intelligent control of the xylose mother liquor chromatographic separation process.

2. The intelligent control method for xylose mother liquor chromatographic separation based on multiple parameters according to claim 1, characterized in that, The key operating parameters in step S1 include feed flow rate, washing water flow rate, column temperature, feed concentration, and switching time. The optimization objectives include maximizing xylose product purity and maximizing xylose yield. The constraints include the physical feasibility range of the key operating parameters and the lower limit of xylose product quality indicators.

3. The intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters according to claim 1, characterized in that, The specific steps for initializing the particle population in step S2 are as follows: Based on the key operational parameter variable dimension m and the population size N, an m-dimensional Sobol quasi-random sequence is generated, producing N initial particle positions; For particles that do not meet the constraints, a repair strategy based on the constraint violation gradient direction is adopted, which projects the infeasible particles to the nearest feasible domain boundary along the direction of the fastest decrease in constraint violation. Dynamic back-learning is applied to the repaired population, and the offset weights of the back-learning points are determined based on the variance of the current population fitness distribution; the back-learning is... ,in, Let be the position of the reverse direction of the i-th particle. For weights in dynamic back-learning, These are the lower and upper bounds of the search space, respectively. Let be the current position of the i-th particle in the search space; The original population and the reverse population are merged, and the top N particles are selected according to fitness to form the initial iterative population.

4. The intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters according to claim 1, characterized in that, The specific steps for introducing the role-adaptive multi-leader mechanism in step S3 are as follows: Calculate the fitness-distance balance index for each particle in the population. The calculation method is as follows: ,in, Let i be the normalized rank of particle i in the fitness ranking. Let be the normalized ranking of particle i from its nearest neighbor particle in the decision space using Euclidean distance. This is the balance coefficient; In each iteration, the total number of leaders is set to... All particles in the population are sorted according to Sort in ascending order and select the first one. The particles in position constitute the leader set; The leader particle is divided into three roles: reconnaissance leader, anchoring leader, and bridging leader. Anchored Leader: Sort the particles in the leader set from best to worst fitness value, and select the top-ranked ones. The particles in the position guide the population to carry out localized and refined development in high-quality areas; Bridging Leader: After sorting the non-dominated solutions on the current Pareto front in ascending order along the first objective function value in the objective space, divide them equally into... Each segment contains the same number of solutions. The particle closest to the geometric center of the segment and belonging to the leader set is selected from each segment as the bridging leader of that segment, thus maintaining the diversity of the solution set distribution in the target space. Scout Leader: The remaining particles in the leader set that are not included in the anchored leader and bridging leader are regarded as scout leaders, guiding the population to conduct global exploration; The number of leaders in each role is based on population diversity indicators. Adjustment: When When the diversity threshold is exceeded, the particle with the worst fitness among the anchored leaders is reclassified as a scout leader; when When the diversity threshold is lower than the threshold, the particle with the best FDB value among the reconnaissance leaders is reclassified as an anchored leader; when the maximum normalized Euclidean distance between adjacent non-dominated solutions on the Pareto front is greater than the preset gap threshold, the number of bridging leaders is increased by 1, while the number of reconnaissance leaders is decreased by 1, ensuring that the total number of leaders remains unchanged.

5. The intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters according to claim 1, characterized in that, In step S3, a trigonometric function-enhanced position update strategy is applied to leaders with different roles. Different roles correspond to different rotation angle update rules, specifically including: For reconnaissance-type leaders, rotation angle ,in, The preset maximum rotation angle has a value range of [value range missing]. t is the current iteration number, and T is the total number of iterations; For anchored leaders, rotation angle ,in, The preset minimum rotation angle has a range of values. ; For bridging leaders, rotation angle ; Then, the particle position is updated using trigonometric functions: ,in, Let i be the updated position of the i-th particle. Let i be the position of the i-th particle before the update. Let r be the position vector of the particle guiding the direction of reference, and r be a random number uniformly distributed in [0,1]. Let be the position vectors of two different particles randomly selected from the population. The following rules apply: for a reconnaissance leader, the position of the particle with the best fitness in the current iteration is selected; for an anchoring leader, the position of the particle with the best fitness in the neighborhood of the decision space with the particle as the center and a preset neighborhood radius is selected; for a bridging leader, the position vector corresponding to the non-dominated solution in the external archive that has the closest Euclidean distance to the particle in the target space is selected.

6. The intelligent control method for chromatographic separation of xylose mother liquor based on multiple parameters according to claim 1, characterized in that, The specific steps for introducing the sensitivity-guided dimension-selective update mechanism in step S3 are as follows: In the early iteration stage of the algorithm, small positive perturbations are applied to each decision variable dimension in turn, the change magnitude of all objective functions before and after the perturbation is statistically analyzed, and the sensitivity index corresponding to each dimension is calculated. Calculate the median of all sensitivity indices and use this median as the dividing line to separate the high-sensitivity dimension set from the low-sensitivity dimension set. During particle position update, the high-sensitivity dimension uses the baseline update step size, while the low-sensitivity dimension uses an amplified update step size; the corresponding dimension update magnitude is performed in the position update operation. Every fixed number of iterations, the sensitivity of each dimension is recalculated, and the results of the high and low sensitivity dimension division are updated.

7. The intelligent control method for xylose mother liquor chromatographic separation based on multiple parameters according to claim 1, characterized in that, The specific steps of adopting the external archive maintenance strategy based on the super-volume index in step S3 are as follows: Store all non-dominated solutions obtained during the optimization process into an external archive. If the number of solutions in the file exceeds the preset maximum capacity, calculate the decrease in the overall over-volume index after removing a single solution as the contribution of that solution to the removal. The solution with the smallest contribution is removed in a loop until the number of solutions in the file does not exceed the maximum capacity. Adaptive grid partitioning is implemented for external archives, and the number of grid partitions is determined based on the archive capacity. When there are multiple non-dominated solutions in the same grid area, only the solution with the highest hypervolume contribution is retained, and the other solutions are removed to ensure that the Pareto solution set is uniformly distributed in the target space.

8. A multi-parameter-based intelligent control system for xylose mother liquor chromatographic separation, characterized in that, include: The model building module establishes a multi-parameter, multi-objective optimization control model for chromatographic separation based on the key operating parameters and separation quality objectives of the xylose mother liquor chromatographic separation process. The algorithm initialization module uses a hierarchical Sobol quasi-random sequence combined with a constraint-aware repair strategy to initialize the particle population, and generates an enhanced population through dynamic back learning. The algorithm improvement module improves the leader-driven particle swarm optimization algorithm by: introducing a role-adaptive multi-leader mechanism, dynamically dividing leader particles into three roles: scout leader, anchor leader, and bridging leader; adopting a position update strategy enhanced by trigonometric functions for leaders of different roles; introducing a sensitivity-guided dimension-selective update mechanism; and adopting an external archive maintenance strategy based on hypervolume indices. The control execution module uses an improved leader-driven particle swarm optimization algorithm to solve the multi-parameter, multi-objective optimization control model for chromatographic separation, and performs online parameter adjustment and intelligent control of the chromatographic separation system based on the solution results.

Images