Liquid chromatogram flow velocity real-time detection and optimal control method based on multi-point sensing

Abstract

The invention discloses a liquid chromatogram flow velocity real-time detection and optimal control method based on multi-point sensing. The method comprises the following steps: S1, constructing a distributed acquisition network; s2, constructing a fluid transmission topological graph and generating a multi-point flow velocity time sequence; s3, inputting the fluid transmission topological graph and the multi-point flow velocity time sequence into a FlowFormer model to generate a flow velocity evolution prediction map; s4, executing local gradient scanning, and generating an optimization control objective function; s5, calling a self-adaptive fox swarm algorithm controller, inputting a flow velocity evolution prediction map and an optimization control objective function, and generating an optimal adjustment strategy; s6, the control parameters of the liquid chromatography system are adjusted in real time and fed back to the FlowFormer model; and S7, when abnormal flow velocity change is detected, triggering an adaptive fox swarm algorithm controller to execute a local escape strategy. According to the invention, multi-point sensing and an intelligent optimization algorithm are fused, and real-time detection and optimal control of the liquid chromatogram flow velocity are realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence automatic control technology, and in particular to a method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing. Background Art

[0002] Liquid chromatography is an extremely important separation and quantitative analysis technology in modern analytical chemistry and is widely used in biomedicine, environmental monitoring, food safety, chemical processes and other fields. In practical applications, flow rate, as a key operating parameter of the liquid chromatography system, directly affects the separation efficiency of the chromatographic peak, the stability of the retention time and the stable reproducibility of the detection sensitivity. Therefore, how to achieve high-precision real-time monitoring and dynamic regulation of the flow rate has become one of the core technologies to ensure the performance of the liquid chromatography system.

[0003] In the prior art, liquid chromatography systems usually use single-point flow rate sensors in conjunction with fixed control logic to monitor and adjust the system flow rate. These single-point sensors are usually deployed at the inlet of the system or the front end of the chromatographic column, and mainly rely on the traditional pressure-flow rate conversion relationship to indirectly infer the liquid flow in the pipeline. At the same time, the control part relies on a controller based on a proportional-integral-differential (PID) algorithm or a preset adjustment curve to adjust the high-pressure pump speed or mixing ratio to maintain a relatively stable flow rate output. However, with the widespread application of complex sample components and nonlinear solvent gradient conditions, this control mode based on single-point perception and rule control has gradually shown obvious deficiencies in flow rate response speed, abnormal state handling capabilities and overall system stability.

[0004] First, in the actual operation process, since the flow state of the liquid in the pipeline is affected by many factors such as the mixing chamber, column packing, and pipe resistance of the connecting pipe section, the spatial distribution of the flow velocity shows obvious uneven characteristics. The single-point measurement method cannot fully reflect the dynamic behavior of the fluid inside the system, and it is easy to have the problem that the flow velocity fluctuation at the back end of the column or the front end of the detector is not detected in time, which in turn causes peak distortion or decreased detection sensitivity. In addition, single-point measurement does not have topological perception capabilities, and it is difficult to use it to analyze the actual distribution of fluids between multiple branch paths. It is particularly weak in judging the multi-segment flow state under complex structures (such as switching valves, bypass channels, etc.) in the chromatographic system.

[0005] Secondly, existing control methods often model the liquid phase system as a single-input single-output (SISO) system. The control logic lacks the global perception ability of the system state, and the control strategy relies on preset rules and lacks intelligent prediction functions. This results in a lag in the adjustment process response, a decrease in control accuracy, and problems such as under-adjustment or over-adjustment when the controller faces external disturbances such as variable sample loads, temperature fluctuations, and pressure disturbances, affecting the detection repeatability and stability. Especially under high-throughput experiments or continuous analysis conditions, this traditional regulation mechanism is difficult to meet the requirements of high-frequency stable operation of the system.

[0006] Thirdly, the response ability to abnormal flow rate events is also an important shortcoming of the existing technology. In practical applications, local flow rate mutations or pressure imbalances may occur due to reasons such as chromatographic column blockage, turbulent flow in the mixing chamber, pipeline loosening, or sudden changes in sample viscosity. Traditional systems often trigger alarms or shutdown processing only after the abnormality has caused obvious results (such as over-limit column pressure or detector signal drift), lacking an early prediction mechanism and local adjustment strategies, which not only cannot achieve dynamic risk mitigation but also increases the overall operation risk and maintenance cost of the system.

[0007] In addition, although some high-end liquid phase systems have begun to introduce multi-sensor collaborative control, data-driven modeling, and AI-based adjustment frameworks, these systems mostly exist in the form of commercial closed solutions, making it difficult to deeply customize and optimize algorithms, and there are engineering adaptability problems such as poor model generalization ability, opaque regulation strategies, and strong software-hardware coupling in the actual implementation process, and they still cannot meet the needs of scientific research and high-precision analysis for open, controllable, and evolvable systems.

[0008] Therefore, how to provide a real-time detection and optimization control method for liquid chromatography flow rate based on multi-point sensing is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0009] An object of the present invention is to propose a real-time detection and optimization control method for liquid chromatography flow rate based on multi-point sensing. The present invention integrates multi-point sensing technology, a structure-aware deep prediction model FlowFormer, and an adaptive fox swarm optimization algorithm to construct a closed-loop mechanism of flow rate prediction-control-feedback, realizing high-precision real-time detection and intelligent adjustment of the flow rate of the liquid chromatography system, having the advantages of high regulation accuracy, fast response speed, intelligent abnormal handling, and strong self-learning ability, significantly improving the operation stability and data reliability of the liquid chromatography system under complex working conditions, and being applicable to high-throughput, high-precision, and high-complexity sample analysis scenarios.

[0010] The real-time detection and optimization control method for liquid chromatography flow rate based on multi-point sensing according to an embodiment of the present invention includes the following steps: S1. Deploy flow rate sensors at key nodes of the liquid chromatography system to form a distributed acquisition network; S2. Construct a fluid transmission topology map based on the distributed acquisition network and simultaneously generate a structured multi-point flow rate time series; S3. Input the fluid transmission topology map and the multi-point flow rate time series into the FlowFormer model, use the position embedding encoding mechanism to encode the flow resistance relationship between topology nodes, calculate the dependency relationship between topology nodes, and generate a flow rate evolution prediction map; S4. Perform local gradient scanning based on the flow rate evolution prediction map, identify potential mutation regions or abnormal fluctuation nodes through curvature detection and high-order difference operations, and generate an optimized control objective function according to the minimum flow rate deviation tolerance and response time constraint set by the chromatographic operating parameters; S5. Invoke the adaptive fox swarm algorithm controller, input the flow rate evolution prediction map and the optimized control objective function, guide the virtual fox swarm to search in the control parameter space, and generate an optimal adjustment strategy; S6. Adjust the control parameters of the liquid chromatography system in real time according to the optimal adjustment strategy, and at the same time feedback to the FlowFormer model for closed-loop iterative optimization; S7. When the liquid chromatography system detects an abnormal flow rate change, trigger the adaptive fox swarm algorithm controller to execute a local escape strategy to achieve emergency control under abnormal conditions.

[0011] Optionally, the key nodes of the liquid chromatography system include the injection port, before and after the mixing chamber, the front end of the chromatographic column, the rear end of the chromatographic column, the front end and the end of the detector.

[0012] Optionally, the S2 specifically includes: S21. Collect the deployment positions and physical connection relationships of each flow rate sensor in the liquid chromatography system, and establish a node set , where represents the th node, represents the total number of nodes, and each node represents a flow rate sensor; S22. According to the connection structure and liquid flow direction of the liquid chromatography system, construct an edge set , where each edge represents that there is an effective flow-through relationship between the th node and the th node, forming a fluid transmission topology map , and label the pipe segment length, inner diameter, friction coefficient and initial flow resistance estimate for each edge; S23. Send a unified synchronous trigger signal to all flow rate sensor nodes, use the global time synchronization protocol to achieve clock calibration, enable each node to start sampling at the synchronous moment, and set the unified sampling frequency to ; S24. Each node collects the current flow velocity value at each sampling moment to form an initial raw sampling data sequence, and binds a local timestamp and a flow velocity sensor number to each sampling value; S25. Perform data alignment processing on all initial raw sampling data sequences, use the time window alignment algorithm to unify the sampling time sequence and correct the sampling error caused by hardware delay or communication difference, and generate a structured multi-point flow velocity time series; S26. Store the fluid transmission topology map and the multi-point flow velocity time series in the cloud, and mark the acquisition batch number, the start and end time of sampling, and the device status code.

[0013] Optionally, the S3 specifically includes: S31. Based on the structure and physical properties of the fluid transmission topology map, construct an embedding vector for structure modeling, and calculate the local structure-aware position tensor for each node: ; Wherein, represents the perception position tensor of the th node, , and represent the three-dimensional space coordinates of the th node in the liquid chromatography system, represents the set of direct adjacent nodes of the th node, represents the perceptron embedding function, represents the initial flow resistance coefficient of the edge , represents the th node and the th node S32. Collect the historical flow velocity values of each node at the nearest time step, and splice them with the perception position tensor to form a joint input tensor: ; Wherein, represents the joint input tensor of the th node, represents the splicing operation, represents the current moment, represents the time window length, represents the th node at time observation flow velocity value; S33. Input the combined input tensors of all nodes into the FlowFormer model, which includes a structural encoding module and a temporal prediction module. The structural encoding module models topological dependencies in the form of graph attention and introduces a structural consistency scoring function: ; where represents the structural consistency score between the -th node and the -th node, represents the historical flow rate change difference sequence between the -th node and the -th node, represents the flow resistance difference vector between the -th node and the -th node, represents the numerical stability term, represents the vector inner product, represents the L2 norm; S34. In the temporal prediction module of the FlowFormer model, use a one-dimensional convolutional network to perform step-by-step prediction on the node feature sequence after structural encoding, capture the dynamic trends of each node, and output the hidden state sequence. Then, use a linear transformation layer to map the hidden state sequence to the flow rate prediction values at future time points; S35. Combine the flow rate prediction values of all nodes at future time points into a flow rate evolution prediction map during the prediction period.

[0014] Optionally, the specific steps of S4 include: S41. Receive the flow rate evolution prediction map, which includes the flow rate prediction values of all nodes at future time points, and record the evolution trend of the flow rate over time; S42. Perform local window sliding processing on the time series of each node, and use a sliding window with a fixed width to extract the local flow rate change curve for gradient change scanning analysis; S43. Introduce a curvature sensitivity function to perform curvature change identification on the local flow rate change curve of each node within the time window: ; where represents the curvature index of the -th node at time , represents the flow rate prediction value of the -th node at time , represents the first-order time derivative, represents the second-order time derivative, Represents the adjustment parameter of the th node, represents the smoothing factor; S44. In each node time series, locate the set of time points of the local maximum curvature index as the mutation candidate point set, and perform high-order difference residual filtering on the mutation candidate point set: ; Among them, represents the fourth-order central difference mutation index of the th node at time , represents the predicted flow velocity value of the th node at time , represents the predicted flow velocity value of the th node at time , represents the predicted flow velocity value of the th node at time , represents the predicted flow velocity value of the th node at time , represents the absolute value operation; S45. Combine the curvature index and the fourth-order central difference mutation index and adopt a threshold double-threshold strategy to mark the mutation region or the abnormal fluctuation node set ; S46. Based on the recognition result and the current operating state, construct an optimal control objective function: ; Among them, represents the optimal control objective function, represents the set of control variables, , and represent the weighting coefficients, represents the target reference flow velocity of the th node, represents the predicted flow velocity value of the th node at time , represents the predicted time step, represents the time window length, represents the predicted flow velocity value of the th node at time , represents the response time required for the th node to reach the target state, represents the stability term to prevent the denominator from being zero.

[0015] Optionally, the threshold double-threshold strategy specifically includes setting a curvature index threshold and a fourth-order central difference variation index threshold. The curvature index threshold is used to judge the bending degree of the local flow velocity change curve, and the fourth-order central difference variation index threshold is used to judge the fluctuation intensity of the local flow velocity change curve; In the flow velocity evolution prediction map, filter out the time points where the curvature index exceeds the curvature index threshold as suspected mutation points, and further judge whether the fourth-order central difference variation index of the suspected mutation points exceeds the fourth-order central difference variation index threshold. Only retain the nodes that meet both threshold conditions to generate a mutation region or an abnormal fluctuation node set .

[0016] Optionally, the S5 specifically includes: S51. Obtain the flow velocity evolution prediction map and the optimized control objective function, and establish a control parameter space, where the control parameter space includes the high-pressure pump speed, the solvent ratio, and the valve opening; S52. Initialize the population using an adaptive fox swarm algorithm controller, and define the population as a set of fox swarm individuals , where represents the th fox individual, represents the total number of fox individuals. Each fox individual represents a set of candidate solutions in the control parameter space. The initial fox swarm individuals are generated by uniform random distribution, and the maximum number of generations of evolution and the convergence threshold are set; S53. According to the optimized control objective function, substitute the flow velocity evolution prediction map into the evaluation of the control performance and record the current optimal individual; S54. According to the distance difference between each fox swarm individual and the optimal individual, adaptively and dynamically update the position vector of the fox individual: ; where, represents the position vector of the th fox individual after update, represents the position vector of the th fox individual before update, represents the dynamic step size adjustment factor, represents the position vector of the optimal individual, represents the random perturbation coefficient, represents a Gaussian distribution random vector with a mean of 0, represents the L2 norm; S55. Record the optimal individuals generated during the evolution of the fox swarm in each generation and add them to the memory bank to assist in the generation of the next generation of fox swarm individuals; S56. Repeat steps S53 - S55 until the set number of evolution generations is reached or the convergence threshold of the optimization control objective function is met, and generate the optimal adjustment strategy.

[0017] Optionally, S6 specifically includes: S61. According to the optimal adjustment strategy, the central control unit sends control commands to each actuator of the liquid chromatography system in real time, including the high-pressure pump speed adjustment command, the proportion adjustment command of the solvent ratio electric control valve, and the valve opening control command; S62. After receiving the control commands, each actuator immediately implements the adjustment action. The high-pressure pump adjusts the operating frequency to reach the target flow rate setting, the proportion valve dynamically adjusts the proportion of different solvents and stabilizes the mixing proportion, and the valve adjusts the pipeline pressure and flow according to the opening degree; S63. During the adjustment execution process, the distributed acquisition network is used to collect the response flow rate data of each node of the liquid chromatography system in real time, and the adjusted response state is synchronously recorded to obtain the feedback data under the actual operating state; S64. Compare the feedback data with the flow rate prediction value of the FlowFormer model, calculate the deviation between the flow rate prediction value and the feedback data, and adaptively update the parameters of the FlowFormer model based on the deviation: ; Wherein, represents the updated FlowFormer model parameters, represents the FlowFormer model parameters before update, represents the adaptive learning rate, represents the gradient of the FlowFormer model parameters, represents the deviation between the flow rate prediction value and the feedback data, represents the L2 norm; S65. Use the updated FlowFormer model to regenerate the flow rate evolution prediction map for the next cycle, and further adjust and optimize the adjustment strategy for the next cycle to form a feedback closed loop.

[0018] Optionally, the local escape strategy includes that when the liquid chromatography system detects an abnormal flow rate change, the adaptive fox swarm algorithm controller limits the search range to the abnormal area where the abnormal node and the directly related upstream and downstream nodes are located, and preferentially generates a local optimal adjustment solution with the goal of reducing pressure and bypassing by compressing the control parameter space and introducing a fast iteration factor, which includes dynamically adjusting the valve opening of the abnormal path segment, temporarily reducing the output frequency of the high-pressure pump, or switching to a preset backup solvent channel to achieve flow rate unloading and pressure release for the abnormal node. At the same time, the liquid chromatography system continuously collects response data of the abnormal area, and records the input parameters and feedback results during the escape process to an offline learning pool.

[0019] The beneficial effects of the present invention are: First, the present invention deploys multiple high-frequency micro-flow rate sensors at key nodes of the liquid chromatography system to build a distributed acquisition network with comprehensive spatial coverage, breaking the technical bottleneck of single-point sampling and local observation in the prior art. The network not only covers multiple key positions such as the injection port, the front and back of the mixing chamber, the front and back of the chromatographic column, the front and end of the detector, but also realizes high-precision alignment of multi-point data through a unified time synchronization mechanism. Combined with the system topology, a fluid transmission directed graph is further constructed, so that a corresponding mapping relationship is established between the physical structure, pipeline connection relationship and sensor data inside the system, which significantly enhances the visualization and structural computability of the system operation status.

[0020] Secondly, the present invention introduces the structure-aware deep prediction model FlowFormer, which fuses the fluid topology map with the multi-point flow velocity time series, and encodes and models the flow resistance relationship between topological nodes through the position embedding mechanism, and uses the structural consistency scoring mechanism to mine the nonlinear dynamic dependencies between nodes. Compared with the traditional method of using only sequence data for sliding prediction, this model has stronger global state modeling capabilities and dynamic evolution trend capture capabilities, and can generate high-resolution system flow velocity evolution prediction maps, providing accurate decision-making basis for subsequent control strategy generation. At the same time, by introducing a convolutional time series prediction structure, the generalization ability and robustness of the model on the continuous time axis are improved.

[0021] Again, in terms of control strategy generation, the present invention uses an adaptive fox swarm algorithm controller to make a breakthrough in traditional control algorithms. The controller is based on a group intelligent behavior simulation mechanism, takes the flow rate prediction map and the optimization control objective function as input, guides the virtual fox swarm to search and iteratively optimize in the high-dimensional control parameter space, and dynamically generates the system-level optimal control strategy. This strategy covers key control variables such as high-pressure pump speed, solvent ratio, and microvalve opening. It has high control accuracy and fast convergence speed. It can also take into account optimization in multiple dimensions such as target deviation, system response time, and flow rate disturbance suppression, which significantly improves the control effect of the liquid chromatography system under complex working conditions.

[0022] In addition, the present invention also constructs an integrated closed-loop optimization mechanism of prediction-execution-feedback-update. During the execution of the adjustment strategy, the system collects real-time feedback flow rate data and compares the error with the predicted value of the FlowFormer model. Based on the prediction deviation, the model parameters can be adaptively updated, so that the prediction performance can be continuously improved during continuous use, thus having the learning ability and model evolution ability. This self-closed-loop mechanism effectively solves the problems of static models and fixed regulation in existing liquid phase systems, making the entire system highly adaptable and operationally robust.

[0023] Finally, the present invention proposes a local escape strategy to cope with abnormal states during system operation. When the system detects a flow rate mutation exceeding the safety threshold at any node, the controller can automatically identify the abnormal area, limit the search space within a local range, and quickly generate a local optimal adjustment strategy targeting pressure reduction, flow bypass, or bypass switching, so as to achieve rapid unloading of the abnormal area and maintain the stable operation of the overall system. At the same time, the data and strategy results in this process are stored in the offline learning pool for training future models and enhancing the response ability of the escape mechanism, further improving the safety and intelligence of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, but do not constitute a limitation to the present invention. In the drawings: Figure 1 is the overall flowchart of the real-time detection and optimization control method of liquid chromatography flow rate based on multi-point sensing proposed by the present invention; Figure 2 is the architecture diagram of the FlowFormer model of the real-time detection and optimization control method of liquid chromatography flow rate based on multi-point sensing proposed by the present invention; Figure 3 is the flowchart of generating the adjustment strategy of the adaptive fox swarm algorithm controller of the real-time detection and optimization control method of liquid chromatography flow rate based on multi-point sensing proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0026] Refer to Figures 1-3 , the real-time detection and optimization control method of liquid chromatography flow rate based on multi-point sensing includes the following steps: S1. Deploy flow rate sensors at key nodes of the liquid chromatography system to form a distributed acquisition network; S2. Construct a fluid transmission topology map based on the distributed acquisition network and synchronously generate a structured multi-point flow velocity time series; S3. Input the fluid transmission topology map and the multi-point flow velocity time series into the FlowFormer model, use the position embedding encoding mechanism to encode the flow resistance relationship between topological nodes, calculate the dependency relationship between topological nodes, and generate a flow velocity evolution prediction map; S4. Perform local gradient scanning based on the flow velocity evolution prediction map, identify potential mutation regions or abnormally fluctuating nodes through curvature detection and high-order difference operations, and generate an optimized control objective function according to the minimum flow velocity deviation tolerance and response time constraint set by the chromatographic operation parameters; S5. Invoke the adaptive fox swarm algorithm controller, input the flow velocity evolution prediction map and the optimized control objective function, guide the virtual fox swarm to search in the control parameter space, and generate an optimal adjustment strategy; S6. Adjust the control parameters of the liquid chromatography system in real time according to the optimal adjustment strategy, and at the same time feedback to the FlowFormer model for closed-loop iterative optimization; S7. When the liquid chromatography system detects an abnormal flow velocity change, trigger the adaptive fox swarm algorithm controller to execute a local escape strategy to achieve emergency control under abnormal conditions.

[0027] In the present invention, by constructing a multi-point distributed sensing network in the liquid chromatography system and combining structured modeling, a deep prediction model, and an intelligent optimization algorithm, a comprehensive perception, accurate prediction, and efficient regulation of the internal flow velocity state of the system are realized. Compared with the traditional single-point detection and static adjustment methods, the present invention can capture the spatial distribution of the flow velocity and its change trend in real time, and improve the response ability of the control system to the dynamic operating state. At the same time, the introduction of the FlowFormer model realizes the global flow velocity prediction with structure perception, and cooperates with the adaptive fox swarm algorithm controller to quickly search for the optimal adjustment strategy in the control parameter space, constructing a prediction-control-feedback closed-loop mechanism, enabling the system to have intelligent optimization and abnormal response capabilities while ensuring stable operation, significantly improving the repeatability, accuracy, and operation safety of chromatographic separation, and having broad practical application prospects and industrialization value.

[0028] In this embodiment, the key nodes of the liquid chromatography system include the injection port, before and after the mixing chamber, the front end of the chromatographic column, the rear end of the chromatographic column, the front end and the end of the detector.

[0029] By clarifying the location deployment strategy of the key nodes of the liquid chromatography system, such as before and after the injection port, mixing chamber, chromatographic column, and both ends of the detector, it ensures that the deployed sensor network is representative and covered, thus providing a basic guarantee for constructing a high-resolution and multi-angle flow rate monitoring system. This layout scheme not only improves the observability of the internal flow state of the system but also helps to identify the abnormal flow rate fluctuation areas caused by local structural characteristics, providing data support for subsequent modeling and optimization. The implementation of this strategy can effectively enhance the integrity and practicality of data collection, enable the system to have the ability to respond to multiple points under complex flow conditions, and contribute to the construction of a more refined and stable chromatographic control system.

[0030] In this embodiment, the S2 specifically includes: S21. Collect the deployment positions and physical connection relationships of each flow rate sensor in the liquid chromatography system, and establish a node set , where represents the th node, represents the total number of nodes, and each node represents a flow rate sensor; S22. According to the connection structure and liquid flow direction of the liquid chromatography system, construct an edge set , where each edge represents that there is an effective circulation relationship between the th node and the th node, forming a fluid transmission topology graph , and label the pipe segment length, inner diameter, friction coefficient, and initial flow resistance estimation value for each edge; S23. Send a unified synchronous trigger signal to all flow rate sensor nodes, and use the global time synchronization protocol to achieve clock calibration, so that each node starts sampling at the synchronous moment, and set the unified sampling frequency to ; S24. Each node collects the current flow rate value at each sampling moment, forms an initial raw sampling data sequence, and binds a local timestamp and a flow rate sensor number to each sampling value; S25. Perform data alignment processing on all initial raw sampling data sequences, use the time window alignment algorithm to unify the sampling time sequence and correct the sampling error caused by hardware delay or communication difference, and generate a structured multi-point flow rate time series; S26. Store the fluid transmission topology graph and the multi-point flow rate time series in the cloud, and label the acquisition batch number, sampling start and end times, and device status code.

[0031] By introducing a structured modeling method, the construction process of the topological graph, the synchronous sampling mechanism, and the standardization method of time series data are described in detail, effectively improving the acquisition efficiency and timing consistency of multi-point flow velocity information. This process ensures the integrity and accuracy of sensor data in terms of time and structure dimensions through operations such as unified synchronous triggering, high-frequency sampling, timestamp alignment, and interpolation compensation. The collected structured data not only has high availability but also has the ability to map graph structures, providing high-quality raw data for the input of subsequent deep learning models, fundamentally enhancing the perception ability and generalization effect of the prediction model for the flow state of the liquid phase system, and laying a solid foundation for achieving high-precision prediction and stable control.

[0032] In this embodiment, S3 specifically includes: S31. Based on the fluid transmission topological graph structure and physical properties, construct an embedding vector for structure modeling, and calculate the local structure perception position tensor for each node: ; Among them, represents the perception position tensor of the -th node, , and represent the three-dimensional space coordinates of the -th node in the liquid chromatography system, represents the set of direct adjacent nodes of the -th node, represents the perceptron embedding function, represents the initial flow resistance coefficient of the edge , represents the -th node and the -th node, S32. Collect the historical flow velocity values of each node at the nearest time step and splice them with the perception position tensor to form a joint input tensor: ; Among them, represents the joint input tensor of the -th node, represents the splicing operation, represents the current moment, represents the time window length, represents the -th node at time observed flow velocity value; S33. Input the combined input tensors of all nodes into the FlowFormer model, where the FlowFormer model includes a structural encoding module and a temporal prediction module. The structural encoding module models topological dependencies in the form of graph attention and introduces a structural consistency scoring function: ; where represents the structural consistency score between the -th node and the -th node, represents the historical flow rate change difference sequence between the -th node and the -th node, represents the flow resistance difference vector between the -th node and the -th node, represents the numerical stability term, represents the vector inner product, represents the L2 norm; S34. In the temporal prediction module of the FlowFormer model, use a one-dimensional convolutional network to perform step-by-step prediction on the node feature sequence after structural encoding, capture the dynamic trends of each node, and output the hidden state sequence. Then, use a linear transformation layer to map the hidden state sequence to the flow rate prediction values at future time points; S35. Combine the flow rate prediction values of all nodes at future time points into a flow rate evolution prediction map during the prediction period.

[0033] Through the detailed definition of the FlowFormer structure-aware model and the improved design of the node encoding method, the physical structure characteristics, flow resistance distribution, and flow rate change trends inside the liquid chromatography system are effectively mapped into a unified high-dimensional vector space. In particular, the structural consistency scoring function is introduced as the basis for attention allocation, enabling the model to consider both historical dynamic data and integrate topological structure and physical resistance factors when modeling the dependencies between nodes, thereby significantly improving the accuracy and robustness of the flow rate prediction results. This modeling method is more interpretable and adaptable than traditional temporal prediction models and is suitable for chromatographic analysis environments with complex flow structures and frequent perturbations, providing core support for realizing intelligent perception and precise control.

[0034] In this embodiment, the specific content of S4 includes: S41. Receive the flow rate evolution prediction map, which includes the flow rate prediction values of all nodes at future time points, and record the evolution trend of the flow rate over time; S42. Perform local window sliding processing on the time series of each node, and use a sliding window with a fixed width to extract the local flow velocity change curve for gradient change scanning analysis; S43. Introduce a curvature sensitivity function to perform curvature change identification on the local flow velocity change curve of each node within the time window: ; where, represents the curvature index of the th node at time , represents the flow velocity prediction value of the th node at time , represents the first-order time derivative, represents the second-order time derivative, represents the adjustment parameter of the th node, represents the smoothing factor; S44. In each node time series, locate the set of time points of the local maximum curvature index as the mutation candidate point set, and perform high-order difference residual filtering on the mutation candidate point set: ; where, represents the fourth-order central difference variation index of the th node at time , represents the flow velocity prediction value of the th node at time , represents the flow velocity prediction value of the th node at time , represents the flow velocity prediction value of the th node at time , represents the flow velocity prediction value of the th node at time , represents the absolute value operation; S45. Combine the curvature index and the fourth-order central difference variation index and adopt a threshold double-threshold strategy to mark the mutation region or the abnormal fluctuation node set ; S46. Based on the recognition result and the current operating state, construct an optimal control objective function: ; where, represents the optimal control objective function, represents the set of control variables, , and represent the weighting coefficients, represents the target reference flow velocity of the -th node, represents the predicted flow velocity value of the -th node at time , represents the predicted time step, represents the time window length, represents the predicted flow velocity value of the -th node at time , represents the response time required for the -th node to reach the target state, represents the stabilizing term to prevent the denominator from being zero.

[0035] By introducing the gradient scanning, curvature recognition, and high-order difference residual detection mechanisms, the sensitivity and positioning accuracy of the system to sudden changes or abnormal fluctuations in flow velocity are effectively improved. This strategy captures the changing trend of the minute perturbations of the flow velocity curve in the time domain through a mathematical model, and combined with the double-threshold determination strategy, it can achieve effective early warning at the initial stage of the abnormal state, providing a more timely and accurate abnormal recognition result for the controller. Furthermore, by constructing an optimized control objective function based on target deviation, fluctuation suppression, and response time constraints, the transformation of the control system from rule-driven to target-aware is realized, significantly enhancing the control effect on the non-linear dynamics under complex operating scenarios.

[0036] In this embodiment, the threshold double-threshold strategy specifically includes setting the curvature index threshold and the fourth-order central difference variation index threshold. The curvature index threshold is used to judge the bending degree of the local flow velocity change curve, and the fourth-order central difference variation index threshold is used to judge the fluctuation intensity of the local flow velocity change curve; In the flow velocity evolution prediction map, the time points with curvature indices exceeding the curvature index threshold are screened out as suspected mutation points, and the fourth-order central difference variation index of the suspected mutation points is further judged whether it exceeds the fourth-order central difference variation index threshold. Only the nodes that meet both threshold conditions are retained to generate the mutation region or abnormal fluctuation node set .

[0037] By introducing a double-threshold judgment mechanism in the mutation point recognition process, a joint analysis is carried out for the curvature of the flow rate change and the fluctuation intensity of the fourth-order central difference respectively, effectively solving the problem of false alarms or missed reports in single-index recognition. This double-threshold strategy ensures that only when the prediction results meet the mutation characteristics in both the change trend and the fluctuation degree dimensions at the same time, it is marked as a valid abnormal node, greatly improving the accuracy and discrimination efficiency of abnormal recognition. The use of this strategy provides a more reliable input basis for mutation events for the subsequent control logic, effectively avoiding the risks of misregulation and response lag, and ensuring the continuity and stability of the liquid chromatography system.

[0038] In this embodiment, the S5 specifically includes: S51. Obtain the flow rate evolution prediction map and the optimized control objective function, and establish a control parameter space, where the control parameter space includes the high-pressure pump speed, the solvent ratio, and the valve opening; S52. Initialize the population using an adaptive fox swarm algorithm controller, and define the population as a set of fox swarm individuals , where represents the th fox individual, represents the total number of fox individuals, and each fox individual represents a set of candidate solutions in the control parameter space. The initial fox swarm individuals are generated by uniform random distribution, and the maximum number of generations of evolution and the convergence threshold are set; S53. According to the optimized control objective function, substitute the flow rate evolution prediction map into the evaluation of the control performance, and record the current optimal individual; S54. According to the distance difference between each fox swarm individual and the optimal individual, adaptively and dynamically update the position vector of the fox individual: ; where, represents the position vector of the updated th fox individual, represents the position vector of the th fox individual before update, represents the dynamic step size adjustment factor, represents the position vector of the optimal individual, represents the random perturbation coefficient, represents a Gaussian distribution random vector with a mean of 0, represents the L2 norm; S55. Record the optimal individuals generated during the evolution process of each generation of fox swarms and add them to the memory bank for assisting in the generation of the next generation of fox swarm individuals; S56. Repeat steps S53 - S55 until the set number of generations of evolution is reached or the convergence threshold of the optimized control objective function is met, and generate the optimal adjustment strategy.

[0039] By designing an adaptive fox swarm algorithm search mechanism based on virtual agents and introducing historical optimal individual memory and multi-generation evolution strategies, the efficient solution of complex non-linear control objectives is achieved. This algorithm can dynamically adjust the search path in the control parameter space, and can find the global or approximate optimal solution in control problems with multiple local optima without relying on gradient information, with strong adaptability. Especially in the face of high-dynamic system environments such as sudden changes in flow rate, the adaptive fox swarm algorithm controller FoxOpt has adaptive jumping ability and path repair mechanism, which can effectively avoid falling into local optima, ensure the effectiveness of the adjustment strategy and the overall system control effect, and has obvious intelligent optimization advantages.

[0040] In this embodiment, the S6 specifically includes: S61. According to the optimal adjustment strategy, the central control unit sends control commands to each actuator of the liquid chromatography system in real time, including the high-pressure pump speed adjustment command, the proportional adjustment command of the solvent ratio electric control valve, and the valve opening control command; S62. After receiving the control command, each actuator immediately implements the adjustment action. The high-pressure pump adjusts the operating frequency to reach the target flow rate setting, the proportional valve dynamically adjusts the ratio of different solvents and stabilizes the mixing ratio, and the valve adjusts the pipeline pressure and flow according to the opening. S63. During the adjustment execution process, the distributed acquisition network is used to collect the response flow rate data of each node of the liquid chromatography system in real time, and the adjusted response state is synchronously recorded to obtain the feedback data under the actual operating state. S64. Compare the feedback data with the flow rate prediction value of the FlowFormer model, calculate the deviation between the flow rate prediction value and the feedback data, and adaptively update the parameters of the FlowFormer model according to the deviation: ; Among them, represents the updated parameters of the FlowFormer model, represents the parameters of the FlowFormer model before update, represents the adaptive learning rate, represents the gradient of the parameters of the FlowFormer model, represents the deviation between the flow rate prediction value and the feedback data, represents the L2 norm; S65. Use the updated FlowFormer model to regenerate the flow rate evolution prediction map of the next cycle, and further adjust and optimize the adjustment strategy of the next cycle to form a feedback closed loop.

[0041] This step constructs a closed-loop self-learning control process of optimal strategy execution-real-time feedback-model update, which not only realizes the precise real-time execution of control actions, but also enables the system to continuously improve the regulation performance based on historical feedback data. By introducing an adaptive update mechanism for model parameters, the FlowFormer model weights can be dynamically adjusted according to the prediction error, so that the prediction model maintains long-term effectiveness and high accuracy. This closed-loop structure improves the operational robustness and regulation intelligence of the liquid chromatography system in a dynamic environment, and realizes an essential leap from static control to dynamic perception and self-evolution regulation.

[0042] In this embodiment, the local escape strategy includes that when the liquid chromatography system detects abnormal flow rate changes, the adaptive fox swarm algorithm controller limits the search range to the abnormal node and the abnormal area where the directly related upstream and downstream nodes are located, and by compressing the control parameter space and introducing a fast iteration factor, a local optimal adjustment solution with pressure reduction and bypass as the goal is preferentially generated, which includes dynamically adjusting the valve opening of the abnormal path segment, temporarily reducing the output frequency of the high-pressure pump, or switching to a preset backup solvent channel to achieve flow rate unloading and pressure release of the abnormal node. At the same time, the liquid chromatography system continuously collects response data of the abnormal area, and records the input parameters and feedback results during the escape process to the offline learning pool.

[0043] By introducing a local escape strategy, when abnormal conditions such as sudden changes in flow rate are detected, the system no longer relies on full system readjustment or forced shutdown, but instead limits the abnormal area for local optimization control, effectively reducing overall system interference. This strategy can quickly unload abnormal pressure and restore normal flow by quickly generating emergency plans such as bypass, pressure reduction, or bypass switching in the abnormal area, significantly shortening the system's self-recovery time. At the same time, by recording data on the escape strategy process and introducing an offline learning mechanism, the system's empirical response capability to similar anomalies is enhanced, providing continuous support for subsequent iterative optimization of the model, and improving the system's adaptability and fault recovery levels.

[0044] Embodiment 1: In order to verify the feasibility of the present invention in implementation, the present invention was applied to the liquid chromatography experimental platform of a certain university analysis and testing center, which has long been used for the separation and analysis of complex Chinese medicine extracts and multi-component protein solutions. In practical applications, the sample sources are complex, the solution viscosity, particle content and gradient elution conditions change frequently, and the traditional chromatography system often faces technical bottlenecks such as drastic flow rate fluctuations, peak shape distortion, and detection signal drift, resulting in poor repeatability and insufficient stability of the analysis results, which seriously affects the reliability of experimental data and sample processing efficiency.

[0045] The present invention adopts a multi-point flow velocity sensing scheme, and a total of 6 high-precision micro flow velocity sensors are deployed at the sample inlet, the outlet of the mixing chamber, the front and rear ends of the chromatographic column, and both ends of the detector to form a distributed acquisition network. The system performs automatic calibration before daily operation to ensure that the sampling frequency of the sensors is uniformly set to 500 Hz, and the whole system is time synchronized through a high-precision clock module. The flow velocity data is uploaded to the central processing module in real time to construct a topological map of the liquid path and generate a structured flow velocity time series. The test sample used in the experiment is a high-concentration flavonoid extract, which is operated under acetonitrile-water gradient conditions. The system has a high operating pressure and strong requirements for flow velocity stability.

[0046] During the first experiment, the FlowFormer model successfully generated a predicted map of the flow velocity evolution within the next 15 seconds by learning the topological structure and historical flow velocity sequences. The control objective function is set to minimize the flow velocity fluctuation of the whole path, taking into account the trade-off between the target flow velocity and the response time. The adaptive fox swarm algorithm controller initializes 30 virtual fox swarm individuals in the control parameter space (high-pressure pump speed, solvent ratio, valve opening), sets the maximum number of iterations to 50, and the convergence threshold to 0.001. Convergence is achieved in the 12th generation, and the generated optimal adjustment strategy is sent to the high-pressure pump and proportional valve in real time to achieve fine control of the system.

[0047] During the continuous operation for 8 hours, the system identified and actively intervened in 9 flow velocity abnormal fluctuation events. Among them, 3 were short-term turbulences in the mixing chamber, 2 were mild blockages at the front end of the chromatographic column, and the remaining 4 were pulse disturbances caused by the instability of the high-pressure pump. Through the local escape strategy, the system switched to the standby solvent channel or temporarily reduced the pump speed in real time, avoiding peak shape distortion and system shutdown. The controller records all the adjustment process data, system response status, and flow velocity error into the learning pool to provide training samples for the subsequent adaptive update of the model.

[0048] Under traditional control conditions (using single-point detection + PID adjustment), the average time required for the platform to run a batch of complex sample tasks is 3 hours, the effective sample detection ratio is 83.6%, and the repeatability deviation (RSD) is about 7.8%. In the continuous three-day operation experiment after applying the present invention, the average task completion time is shortened to 2.1 hours, the effective detection ratio increases to 96.4%, the RSD decreases to 3.1%, and the number of system alarms decreases by 82.5%. In addition, the accuracy of the prediction model is improved from 78.5% at the beginning to 93.2% in the later stage of operation, reflecting that the model has good learning ability and real-time adaptation performance.

[0049] Through actual deployment and operation verification on the liquid chromatography platform in the analysis and testing center of universities, the present invention significantly improves the system's perception ability of flow rate dynamics, prediction accuracy, and intelligent control level. Based on the deployment of multi-point sensors, a structured fluid topology map is successfully constructed and integrated with the flow rate time series. Combining the structure perception model FlowFormer and the adaptive fox swarm algorithm controller FoxOpt, the full-process intelligent flow rate optimization control from real-time perception, prediction and early warning to closed-loop regulation is realized.

[0050] During the implementation process, the present invention effectively identifies and processes various abnormal flow rate events, such as turbulent flow in the mixing chamber, chromatographic column blockage, and pump fluctuations, avoiding experimental interference caused by peak shape distortion or system shutdown. The system prediction accuracy is improved from the initial 78.5% to 93.2%, and the average abnormal response time is shortened to within 3.5 seconds, far superior to the traditional PID control strategy.

[0051] This embodiment fully demonstrates that the present invention has strong adaptability and control effect in the actual complex environment of liquid chromatography. By integrating distributed perception, structure modeling, and intelligent control, the present invention solves the problems that traditional technologies cannot achieve multi-point monitoring, global prediction, and rapid response, and has significant technical advantages in improving sample detection efficiency, control accuracy, and system stability.

[0052] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing, characterized in that: The steps include: S1. Deploy flow rate sensors at key nodes of the liquid chromatography system to form a distributed acquisition network; S2, constructing a fluid transmission topology map based on the distributed acquisition network, and synchronously generating a structured multi-point flow velocity time series; S3, input the fluid transmission topology map and multi-point flow velocity time series into the FlowFormer model, use the position embedding coding mechanism to encode the flow resistance relationship between topological nodes, and calculate the dependency relationship between topological nodes to generate a flow velocity evolution prediction map; S4. Perform local gradient scanning based on the velocity evolution prediction map, identify potential mutation areas or abnormal fluctuation nodes through curvature detection and high-order differential operations, and generate an optimization control objective function based on the minimum velocity deviation tolerance and response time constraints set by the chromatographic operation parameters; S5, calling the adaptive fox swarm algorithm controller, inputting the velocity evolution prediction map and the optimization control objective function, guiding the virtual fox swarm to search in the control parameter space, and generating the optimal regulation strategy; S6. Adjust the control parameters of the liquid chromatography system in real time according to the optimal adjustment strategy, and feed back to the FlowFormer model for closed-loop iterative optimization; S7. When the liquid chromatography system detects an abnormal flow rate change, the adaptive fox swarm algorithm controller is triggered to execute a local escape strategy to achieve emergency control under abnormal conditions.

2. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The key nodes of the liquid chromatography system include the injection port, the front and rear of the mixing chamber, the front end of the chromatographic column, the rear end of the chromatographic column, and the front end and the end of the detector.

3. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The S2 specifically includes: S21. Collect the deployment positions and physical connection relationships of each flow velocity sensor in the liquid chromatography system and establish a node set ,in Indicates nodes, Represents the total number of nodes, each node represents a flow rate sensor; S22. Construct edge sets based on the connection structure and liquid flow direction of the liquid chromatography system , where each edge Indicates The node and There is an effective flow relationship between the nodes, forming a fluid transmission topology diagram , and mark each edge with the length of the pipe segment, inner diameter, friction coefficient and initial flow resistance estimate; S23, send a unified synchronization trigger signal to all flow velocity sensor nodes, use the global time synchronization protocol to implement clock calibration, so that each node starts sampling at the synchronization time, and sets a unified sampling frequency of ; S24, each node collects the current flow velocity value at each sampling time to form an initial original sampling data sequence, and binds a local timestamp and a flow velocity sensor number to each sampling value; S25, performing data alignment processing on all initial raw sampling data sequences, using a time window alignment algorithm, unifying the sampling timing and correcting the sampling error caused by hardware delay or communication difference, and generating a structured multi-point flow velocity time series; S26. Store the fluid transmission topology and multi-point flow velocity time series in the cloud, and mark the collection batch number, sampling start and end time, and device status code.

4. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The S3 specifically includes: S31. Based on the fluid transport topology structure and physical properties, an embedding vector for structural modeling is constructed, and the local structure-aware position tensor is calculated for each node: ； in, Indicates The perceived position tensor of nodes, , and Indicates The three-dimensional spatial coordinates of the nodes in the liquid chromatography system, Indicates The set of direct neighboring nodes of a node, represents the perceptron embedding function, Represents edge The initial flow resistance coefficient, Indicates The node and The spatial distance between nodes; S32, collect the historical flow velocity value of each node in the most recent time step, and compare it with the perceived position tensor Concatenate to form a joint input tensor: ； in, Indicates The joint input tensor of nodes, Represents a splicing operation, Indicates the current moment, represents the time window length, Indicates Nodes at time Observed flow velocity value; S33. The joint input tensor of all nodes is input into the FlowFormer model. The FlowFormer model includes a structure encoding module and a time series prediction module. The structure encoding module models the topological dependency in the form of graph attention and introduces a structure consistency scoring function: ； in, Indicates The node and The structural consistency score between nodes, Indicates The node and The historical flow rate change difference sequence of nodes, Indicates The node and The flow resistance difference vector between nodes is represents the numerical stability term, represents the vector inner product, represents the L2 norm; S34. In the time series prediction module of the FlowFormer model, a one-dimensional convolutional network is used to predict the node feature sequence after structural encoding step by step, capture the dynamic trend of each node, and output a hidden state sequence, which is then mapped to a flow velocity prediction value at a future time point by a linear transformation layer; S35. Combining the velocity prediction values ​​of all nodes at future time points into a velocity evolution prediction map within the prediction time period.

5. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The S4 specifically includes: S41, receiving a velocity evolution prediction map, wherein the velocity evolution prediction map includes velocity prediction values ​​of all nodes at future time points, and records the evolution trend of the velocity over time; S42, performing local window sliding processing on the time series of each node, using a sliding window with a fixed width to extract a local flow velocity change curve for gradient change scanning analysis; S43, introducing a curvature sensitivity function to perform curvature change identification on the local flow velocity change curve of each node within the time window: ； in, Indicates Nodes at time The curvature index when Indicates Nodes at time The predicted flow velocity at represents the first-order time derivative, represents the second-order time derivative, Indicates The tuning parameters of each node, represents the smoothing factor; S44. In each node time series, locate the time point set of the local maximum curvature index as a mutation candidate point set, and perform high-order difference residual filtering on the mutation candidate point set: ； in, Indicates Nodes at time The fourth-order central difference variation index when , Indicates Nodes at time The predicted flow velocity at Indicates Nodes at time The predicted flow velocity at Indicates Nodes at time The predicted flow velocity at Indicates Nodes at time The predicted flow velocity at Represents absolute value operation; S45, combining the curvature index and the fourth-order central difference variation index to adopt a double threshold strategy to mark mutation areas or abnormal fluctuation node sets ; S46. Based on the identification results and the current operating status, an optimization control objective function is constructed: ； in, represents the optimization control objective function, represents the set of control variables, , and represents the weighting coefficient, Indicates The target reference flow rate of each node is Indicates Nodes at time The predicted flow velocity at represents the prediction time step, represents the time window length, Indicates Nodes at time The predicted flow velocity at Indicates The response time required for a node to reach the target state, represents a stabilizing term that prevents the denominator from reaching zero.

6. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 5, characterized in that: The threshold double threshold strategy specifically includes setting a curvature index threshold and a fourth-order central difference variation index threshold, wherein the curvature index threshold is used to determine the curvature degree of the local flow velocity change curve, and the fourth-order central difference variation index threshold is used to determine the fluctuation intensity of the local flow velocity change curve; In the velocity evolution prediction map, the time points where the curvature index exceeds the curvature index threshold are selected as suspected mutation points, and the suspected mutation points are further judged whether the fourth-order central difference variation index exceeds the fourth-order central difference variation index threshold. Only nodes that meet both threshold conditions are retained to generate mutation areas or abnormal fluctuation node sets. .

7. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The S5 specifically includes: S51, obtaining a flow rate evolution prediction map and an optimization control objective function, and establishing a control parameter space, wherein the control parameter space includes a high-pressure pump speed, a solvent ratio, and a valve opening; S52, using the adaptive fox group algorithm controller to initialize the population, defining the population as a collection of fox group individuals ,in Indicates Individual foxes, represents the total number of fox individuals, each fox individual represents a set of candidate solutions in the control parameter space, the initial fox population individuals are generated by uniform random distribution, and the maximum evolutionary generation and convergence threshold are set; S53, according to the optimization control objective function, the velocity evolution prediction map is substituted into the evaluation control performance, and the current optimal individual is recorded; S54, according to the distance difference between each fox group individual and the optimal individual, adaptively and dynamically update the fox individual position vector: ； in, Indicates the updated The position vector of each fox, Indicates the number before the update The position vector of each fox, represents the dynamic step size adjustment factor, represents the position vector of the optimal individual, represents the random perturbation coefficient, represents a Gaussian distributed random vector with mean 0, represents the L2 norm; S55, record the best individuals produced in the evolution process of fox groups of all generations and add them to the memory bank to assist in the generation of individuals in the next generation of fox groups; S56, repeating steps S53-S55 until the set evolutionary generation is reached or the convergence threshold of the optimization control objective function is met, and generating the optimal adjustment strategy.

8. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The S6 specifically includes: S61, according to the optimal adjustment strategy, the central control unit sends control instructions to each actuator of the liquid chromatography system in real time, including a high-pressure pump speed adjustment instruction, a solvent ratio electronic control valve ratio adjustment instruction, and a valve opening control instruction; S62. After receiving the control command, each actuator immediately implements the control action. The high-pressure pump adjusts the operating frequency to achieve the target flow rate setting. The proportional valve dynamically adjusts the ratio of different solvents and stabilizes the mixing ratio. The valve adjusts the pipeline pressure and flow according to the opening; S63, during the adjustment execution process, the response flow rate data of each node of the liquid chromatography system is collected in real time through the distributed acquisition network, and the response state after adjustment is synchronously recorded to obtain feedback data under the actual operation state; S64, comparing the feedback data with the flow velocity prediction value of the FlowFormer model, calculating the deviation between the flow velocity prediction value and the feedback data, and adaptively updating the FlowFormer model parameters according to the deviation: ； in, represents the updated FlowFormer model parameters, Indicates the FlowFormer model parameters before updating, represents the adaptive learning rate, represents the gradient of the FlowFormer model parameters, Indicates the deviation between the flow rate prediction value and the feedback data, represents the L2 norm; S65. Use the updated FlowFormer model to regenerate the flow velocity evolution prediction map for the next cycle, and further adjust and optimize the regulation strategy for the next cycle to form a feedback loop.

9. The method for real-time detection and optimization control of liquid chromatography flow rate based on multi-point sensing according to claim 1, characterized in that: The local escape strategy includes that when the liquid chromatography system detects abnormal flow rate changes, the adaptive fox swarm algorithm controller limits the search range to the abnormal node and the abnormal area where the directly related upstream and downstream nodes are located, and by compressing the control parameter space and introducing a fast iteration factor, a local optimal adjustment solution with pressure reduction and bypass as the goal is preferentially generated, which includes dynamically adjusting the valve opening of the abnormal path segment, temporarily reducing the output frequency of the high-pressure pump or switching to a preset backup solvent channel to achieve flow rate unloading and pressure release of the abnormal node. At the same time, the liquid chromatography system continuously collects response data of the abnormal area, and records the input parameters and feedback results in the escape process to an offline learning pool.

Images