A method and system for adaptive flow control of a liquid chromatography pump
An adaptive flow control method combining a third-order nonlinear extended state observer and a lightweight temporal convolutional network with a Transformer model was developed to address the accuracy and stability issues of liquid chromatography pumps under varying operating conditions, achieving efficient and low-cost flow control.
Patent Information
- Application Number
- CN202611595979.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-15
- Publication Date
- 2026-11-10
AI Technical Summary
Existing liquid chromatography pump flow control systems struggle to maintain high accuracy and stability when operating conditions such as mobile phase viscosity, ambient temperature, and column back pressure change. Furthermore, high-precision flow sensors are expensive and increase system complexity, making it difficult to meet the demands of high-speed, high-sensitivity analysis.
A method based on dual-path network collaboration and soft measurement is adopted. By combining a third-order nonlinear extended state observer and a lightweight temporal convolutional network with a Transformer model, key physical quantities such as flow rate and viscosity are estimated, thereby achieving adaptive control of PID parameters and avoiding the use of high-precision flow sensors.
It reduces hardware costs and system complexity, improves the dynamic response speed and control accuracy of flow estimation, meets millisecond-level real-time control requirements, and enhances the system's adaptability and stability under different operating conditions.
Smart Images

Figure CN123067576A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of liquid chromatography pump flow control technology, and particularly relates to an adaptive flow control method and system for liquid chromatography pumps. Background Technology
[0002] The high-pressure constant flow pump is the power component of a liquid chromatography system, responsible for delivering the mobile phase to the chromatographic column and detector at a constant flow rate under high pressure conditions.
[0003] Traditional liquid chromatography pumps typically employ fixed-parameter PID control schemes for flow control. The control parameters rely on offline calibration under specific operating conditions. When operating conditions such as mobile phase viscosity ratio, ambient temperature, and column back pressure change, the system damping and response characteristics change accordingly. Fixed PID parameters are difficult to maintain optimal control performance, resulting in a significant decrease in control accuracy and stability. Furthermore, it is difficult to effectively suppress the periodic flow pulsations caused by the reciprocating motion of the plunger.
[0004] To achieve high-precision flow closed-loop control, some technical solutions use precision flow meters as direct feedback elements to construct the flow closed loop. However, high-precision flow sensors suitable for high-pressure conditions of tens of MPa are expensive to manufacture, and the sensors themselves have inherent response hysteresis. They also increase the dead volume of the system flow path, affecting chromatographic separation performance and making it difficult to meet the application requirements of high-speed, high-sensitivity liquid chromatography analysis.
[0005] Chinese patent application CN120065760A discloses a method for real-time detection and optimized control of flow rate in liquid chromatography based on multi-point sensing. It deploys distributed flow rate sensors at key nodes of the liquid chromatography system, uses the FlowFormer deep prediction model to generate flow rate evolution maps, and searches for optimal control parameters through an adaptive fox herd algorithm. Under abnormal operating conditions, it triggers a local escape strategy to achieve emergency control and model closed-loop iteration.
[0006] This solution combines deep learning with flow rate control, which improves the system's adaptability to certain operating conditions to some extent. However, this solution relies on multi-point distributed flow rate sensors to achieve flow perception, resulting in high hardware costs and increased system size and flow path complexity. In addition, it adopts a single deep learning model architecture, which does not consider the computing power limitations and real-time requirements of the embedded control terminal, making it difficult to directly apply to real-time control scenarios of high-pressure pumps with millisecond-level control cycles. Summary of the Invention
[0007] The purpose of this invention is to provide an adaptive flow control method and system for liquid chromatography pumps, which partially solves or alleviates the above-mentioned shortcomings in the prior art. Under existing embedded hardware conditions, it can achieve intelligent flow control of high-pressure pumps with real-time performance and adaptability without relying on expensive high-precision flow sensors.
[0008] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution: A first aspect of the present invention is to provide an adaptive flow control method for a liquid chromatography pump based on dual-pathway network collaboration and soft measurement, comprising the following steps: Measurable physical signals from a liquid chromatography pump are collected, and a third-order nonlinear extended state observer is constructed. Based on some measurable physical signals and the pressure change rate estimate and total system disturbance estimate output by the third-order nonlinear extended state observer, soft measurement values are estimated. The measurable physical signals and soft measurement values are combined into a multidimensional feature vector. The multidimensional feature vector is encoded using a temporal convolutional network to obtain a short-term feature vector; the Transformer model is trained using historical data and physical residuals, and the trained Transformer model outputs an attention weight matrix; the attention weight matrix and the short-term feature vector are weighted and fused to obtain a comprehensive feature vector. Based on the viscosity weighting coefficient, pressure pulsation intensity coefficient, and forward predicted flow deviation in the comprehensive feature vector, PID parameters are calculated according to the PID adaptive law to generate control quantities to drive the pump motor.
[0009] Furthermore, the measurable physical signals include pump outlet pressure, pressure change rate, cam phase angle, motor temperature, ambient temperature, and motor speed; the soft measurement values include instantaneous flow rate, equivalent viscosity of the mobile phase, and system damping coefficient.
[0010] Furthermore, the third-order nonlinear extended state observer is: ; ; ; ; Where e is the state observation error, z1 is the estimated value of the pump outlet pressure, z2 is the estimated value of the pressure change rate, z3 is the estimated value of the total system disturbance, P is the pump outlet pressure, u is the input control quantity, b0 is the system control gain coefficient, fal() is the nonlinear feedback coefficient, and δ is the threshold of the linear segment of the fal() function. 1. 2. 3 are the time derivatives of z1, z2, and z3, respectively, and β1-β3 is the error feedback gain coefficient.
[0011] Furthermore, the temporal convolutional network is a lightweight TCN, which contains four layers of extended convolutions with expansion rates of 1, 2, 4, and 8 respectively. Each layer has thirty-two filters and uses the ReLU activation function. Its input is the multidimensional feature vector of the most recent thirty-two time steps, and its output is a short-term feature vector of one hundred and twenty-eight dimensions.
[0012] Furthermore, the Transformer model is deployed in the cloud and adopts an encoder structure. The input is a long historical window sequence and the corresponding physical residual vector. The physical attention template is obtained from system sensitivity analysis. After training, the attention weight matrix and the convolution kernel parameters of the lightweight temporal convolutional network are extracted.
[0013] Furthermore, the loss function of the Transformer model is: ; Where Loss is the loss value, MSE(Q) pred Q true ) represents the mean squared error loss in flow forecasting, ||W att - W phy ||1 represents the L1 regularization term of the attention weights and the preset physical attention template, λ is the regularization coefficient, and Q pred Q is the predicted flow rate. true W represents the actual traffic value. att W is the attention weight matrix. phy This is the physical attention template matrix.
[0014] Furthermore, the adaptive law for the PID parameters is: ; ; ; Where Kp(t) is the real-time proportional gain coefficient at time t, Kp base This is the baseline value for the proportionality coefficient. w is the weighting coefficient for the viscosity correction term. η This is the viscosity weighting coefficient. This is the equivalent viscosity measurement of the mobile phase. w is the weighting coefficient for the pressure pulsation correction term. ΔP ΔP / Δt is the pressure pulsation intensity coefficient, and ΔP / Δt is the pressure change rate; Ki(t) is the real-time integral gain coefficient at time t. base The baseline value for the integral coefficient, T is the weighting factor for the motor temperature correction term. motor This is the actual measured value of the motor temperature. Let ∫e(t)dt be the integral value of the flow deviation, and Kd(t) be the weighting coefficient of the deviation integral correction term. base The base value for the differential coefficients, e is the weighting coefficient for the prediction deviation change rate correction term. q (t+1) represents the forward flow prediction deviation.
[0015] Furthermore, it also includes residual-driven event triggering steps: The absolute value of the residual between the predicted flow of the temporal convolutional network and the instantaneous flow in the soft measurement value is calculated online. The mean and standard deviation of the residual within a set time window are statistically analyzed. The sum of the mean and three times the standard deviation is used as a dynamic threshold. When the residual is greater than the dynamic threshold, it is determined to be a sudden change in operating condition.
[0016] Furthermore, upon determining a sudden change in operating conditions, the most matching set of pre-stored conservative PID parameters is selected immediately based on the current equivalent viscosity and its rate of change in the soft measurement values to temporarily cover the current PID parameters; the feature cache before and after the sudden change in operating conditions is used for incremental training of the Transformer model to update the attention weight matrix.
[0017] This invention also provides an adaptive flow control system for a liquid chromatography pump based on dual-path network collaboration and soft measurement, comprising: A multidimensional feature vector fusion module is used to acquire measurable physical signals from a liquid chromatography pump and construct a third-order nonlinear extended state observer. Based on some measurable physical signals and the pressure change rate estimate and total system disturbance estimate output by the third-order nonlinear extended state observer, soft measurement values are estimated. The measurable physical signals and soft measurement values are combined into a multidimensional feature vector. The comprehensive feature vector extraction module is used to encode the multidimensional feature vector using a temporal convolutional network to obtain a short-term feature vector; to train the Transformer model using historical data and physical residuals, and to output an attention weight matrix from the trained Transformer model; and to weight and fuse the attention weight matrix and the short-term feature vector to obtain a comprehensive feature vector. The PID control module is used to calculate PID parameters according to the PID adaptive law based on the viscosity weighting coefficient, pressure pulsation intensity coefficient and forward predicted flow deviation in the comprehensive feature vector, so as to generate control quantities to drive the pump motor.
[0018] Beneficial effects: This invention employs a third-order nonlinear expansion state observer, combining plunger kinematics and a fluid compressibility feedforward model. It can calculate key physical quantities across multiple dimensions, such as instantaneous flow rate, equivalent viscosity of the mobile phase, and system damping coefficient, in real time using only conventionally measurable physical signals like pump outlet pressure, cam phase, motor speed, and temperature, without requiring a high-pressure, high-precision flow sensor. This significantly reduces hardware costs and system complexity, avoids the additional dead volume introduced by flow sensors, and improves the separation performance of the chromatographic system. Furthermore, it eliminates the inherent response hysteresis of flow sensors, significantly improving the dynamic response speed of flow estimation and enabling accurate capture of the high-frequency pulsation characteristics of the plunger's reciprocating motion.
[0019] This invention employs a dual-path architecture combining real-time inference with a lightweight temporal convolutional network at the edge with offline optimization of a cloud-based physically enhanced Transformer. A lightweight TCN network is deployed at the edge, enabling short-term feature encoding within millisecond-level control cycles, meeting the stringent latency requirements of high-pressure pump real-time control and fully leveraging the computing power of the embedded system. A Transformer model is deployed in the cloud, learning global operational dependencies and long-term variation patterns from long-term historical data, compensating for the limited receptive field of the edge network. Simultaneously, through low-frequency parameter delivery and OTA update mechanisms, the accuracy of feature extraction at the edge is continuously improved without consuming real-time control bandwidth, achieving a balance between real-time performance and high accuracy. This solves the technical challenge of directly deploying a single complex deep network on an embedded control device.
[0020] The adaptive PID parameter law of this invention is based on physical laws, and the attention weights are generated by prior constraints from fluid dynamics templates. The adjustment of proportional, integral, and derivative parameters is directly and quantitatively related to physical quantities such as viscosity, pressure pulsation, and prediction deviation rate of change, rather than a purely data-driven black-box output. On the one hand, this gives the adjustment of control parameters clear physical meaning and causal logic, avoiding the safety hazard of pure deep learning control models outputting uncontrollable parameters under unseen operating conditions. On the other hand, the introduction of physical priors improves the model's generalization ability, enabling it to output control parameters that conform to fluid laws even in operating condition ranges not covered by training data, significantly improving the reliability and stability of system operation.
[0021] This invention constructs a dynamic residual threshold based on the 3σ statistical principle, which can adapt to the baseline deviation level under different operating conditions and accurately identify abnormal operating conditions such as sudden changes in mobile phase viscosity, air bubbles entering the flow path, and sudden changes in column back pressure. Compared with fixed threshold schemes, it has higher detection accuracy and lower false alarm rate. When operating conditions change abruptly, it quickly switches to a pre-calibrated conservative PID parameter set, using control parameters with higher safety margins to ensure operational stability under extreme conditions. At the same time, it triggers incremental training of the cloud model, continuously expanding the model's operating condition coverage using abnormal operating condition samples, realizing online iterative evolution of control performance. It takes into account both transient operational safety and long-term performance improvement, solving the problems of limited generalization ability and inability to evolve online in traditional gain scheduling schemes. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram illustrating the process of collecting sensor signals to the ESO in step S1, outputting an estimate, and splicing it with the measured value.
[0024] Figure 3 This is an architecture diagram of the interaction between the edge TCN and the cloud Transformer in step S2 through parameter download and attention matrix.
[0025] Figure 4 The mathematical flowchart for obtaining Kp, Ki, and Kd from the comprehensive feature vector in step S3 through adaptive law operation is shown.
[0026] Figure 5 This is a schematic diagram of the decision tree for residual calculation, threshold comparison, and action execution in step S4. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0028] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.
[0029] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0030] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0031] In this document, "and / or" includes any and all combinations of one or more of the listed related items.
[0032] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.
[0033] Example 1: This embodiment uses a binary high-pressure gradient liquid chromatography pump as an example. The embedded controller employs a digital signal processor and a field-programmable gate array (FPGA) architecture, and the cloud training server is equipped with a graphics processor. The present invention provides an adaptive flow control method for a liquid chromatography pump, as follows: Figure 1 As shown, the specific steps include: S1. Acquire measurable physical signals from the liquid chromatography pump and construct a third-order nonlinear extended state observer; based on some measurable physical signals and the pressure change rate estimate and total system disturbance estimate output by the third-order nonlinear extended state observer, estimate the soft measurement value; combine the measurable physical signals and the soft measurement value into a multidimensional feature vector.
[0034] The purpose of this step is to construct a third-order nonlinear extended state observer to estimate the system state and total disturbance by collecting limited measurable physical signals without configuring a high-precision direct flow sensor. Based on existing plunger kinematics and fluid compressibility equations, it calculates key physical quantities that cannot be directly measured, such as flow rate, viscosity, and damping, ultimately forming a multidimensional feature vector containing both measured and soft-measurement values. For example... Figure 2 As shown, it specifically includes: S11, capable of acquiring physical signals.
[0035] The purpose of this step is to complete the acquisition, filtering, range conversion, and basic derived quantity calculation of hardware signals, so as to provide raw measurable input for subsequent observers.
[0036] The system acquires raw physical signals from the liquid chromatography pump in real time and simultaneously receives the target flow rate setpoint from the system. Specifically, the pump outlet pressure is acquired by a high-pressure end pressure sensor, the cam phase angle is acquired by a photoelectric encoder at the pump motor end, the motor temperature is acquired by a temperature sensor built into the motor windings, the ambient temperature is acquired by an external environmental sensor of the pump chamber, the motor speed is obtained from feedback from the motor driver or calculated by encoder pulses, and the target flow rate is set by the user via a host computer.
[0037] The measurable physical signals collected include pump outlet pressure P, pressure change rate ΔP / Δt, cam phase angle θ, and motor temperature T. motor Ambient temperature T ambient Motor speed n. Target flow rate Q. target It is set by the user through the host computer.
[0038] S12, Construction of a third-order nonlinear extended state observer and real-time state estimation.
[0039] This sub-step constructs a third-order nonlinear extended state observer (ESO) based on the pressure dynamics characteristics of a liquid chromatography pump. Using measured pressure as the feedback benchmark, it estimates the system pressure state, pressure change rate state, and total system disturbance state in real time.
[0040] The pressure dynamics of the plunger pump are abstracted as a second-order system, with pressure as the system output and motor drive command as the system input. All internal and external uncertainties, such as changes in mobile phase viscosity, leakage, damping fluctuations, and unmodeled characteristics, are combined into a total system disturbance, which is then expanded into a third independent state variable, forming a third-order observer structure. The third-order nonlinear extended state observer is as follows: ; ; ; ; Where e is the state observation error, z1 is the estimated value of the pump outlet pressure, z2 is the estimated value of the pressure change rate, z3 is the estimated value of the total system disturbance, P is the pump outlet pressure, u is the input control quantity, b0 is the system control gain coefficient, fal() is the nonlinear feedback coefficient, and δ is the threshold of the linear segment of the fal() function. 1. 2. 3 are the time derivatives of z1, z2, and z3, respectively, and β1-β3 is the error feedback gain coefficient.
[0041] Within each control cycle, the measured pressure P and the motor drive command u are input into the observer. The three state variables are continuously updated through nonlinear feedback correction, so that z1 tracks the real pressure, z2 tracks the real pressure change rate, and z3 asymptotically converges to the real total disturbance of the system.
[0042] S13, soft measurement parameter solution based on cam phase feedforward model and disturbance estimation.
[0043] Using the plunger kinematic equations and fluid compressibility corrections in the cam phase feedforward model as the physical benchmark, and combining them with the disturbance estimates output by the expansion state observer, three types of soft-sensor values are calculated: instantaneous flow rate, equivalent viscosity of the mobile phase, and system damping coefficient. The specific calculation steps are existing technology and will not be elaborated here.
[0044] S14, Construction and normalization of multidimensional feature vectors.
[0045] This sub-step concatenates the measured signal, set value, and soft measurement value into a unified feature vector and performs normalization processing to adapt to the input requirements of the subsequent neural network.
[0046] The seven measurable physical signals are: pump outlet pressure P, pressure change rate ΔP / Δt, cam phase angle θ, and motor temperature T. motor Ambient temperature T ambient Motor speed n, and setpoint: target flow rate Q target , and 3D soft measurement values: real-time estimation of flow rate Q actual Equivalent viscosity flow The equivalent damping coefficient C is sequentially concatenated to form a 10-dimensional original feature vector.
[0047] The Z-score standardization method is adopted. Based on the mean and standard deviation of each dimension obtained by offline statistics, each dimension of the original feature vector is normalized to eliminate the differences in the dimensions and numerical ranges of different physical quantities, so as to ensure the stability of subsequent neural network training and inference.
[0048] S2, the multidimensional feature vector is encoded using a temporal convolutional network to obtain a short-term feature vector; the Transformer model is trained using historical data and physical residuals, and the trained Transformer model outputs an attention weight matrix; the attention weight matrix and the short-term feature vector are weighted and fused to obtain a comprehensive feature vector.
[0049] This step employs a dual-path collaborative architecture combining real-time lightweight inference at the edge with offline optimized training on the cloud. A lightweight temporal convolutional network is deployed on the embedded controller to achieve millisecond-level short-term temporal feature encoding, meeting real-time control latency requirements. A physically augmented Transformer model is deployed on the cloud server to learn long-period global operating condition features and output attention weights that conform to physical mechanisms. Edge-cloud feature fusion is achieved through low-frequency distribution of attention weights, balancing real-time performance and control accuracy. Specifically, this includes: S21, the edge-side temporal convolutional network encodes in real time to generate short-term feature vectors.
[0050] This sub-step is executed cycle by cycle in the local embedded controller to capture dynamic changes in operating conditions within a short time window.
[0051] Specifically, based on the normalized multidimensional feature vector output in step S1, a sliding time window mechanism is adopted to cache the feature vectors of the most recent 32 consecutive control cycles, forming a 32×10 dimension time-series feature sequence, corresponding to a 32ms working condition history; each control cycle, the sliding window moves forward one position, removing the features of the earliest time and incorporating the features of the latest time, thereby realizing real-time rolling updates of time-series data.
[0052] In this embodiment, a lightweight temporal convolutional network (TCN) is used as the edge feature extractor. The network employs a causal convolutional structure to ensure that inference only uses historical and current data and does not introduce future information, thus meeting the causal requirements of real-time control. The network consists of four extended convolutional layers with dilation rates of 1, 2, 4, and 8, respectively, fully covering the 32-step time window through exponentially expanded receptive fields. Each layer is configured with 32 convolutional filters, and the activation function is ReLU. (See [link to relevant documentation]). Figure 3 By using hierarchical feature extraction and temporal dependency modeling, the input temporal feature sequence is encoded into high-dimensional abstract features.
[0053] After four layers of extended convolution and global temporal pooling, the output is a 128-dimensional short-term feature vector F. fast It condenses multi-dimensional temporal dynamic features such as pressure, flow rate, viscosity, and temperature within nearly 32ms; simultaneously, a single-neuron prediction head is set at the end of the network to output the forward one-step predicted flow deviation e based on the encoded short-time features.q (t+1) is the predicted flow deviation value for the next control cycle, which is used for advance adjustment of the subsequent derivative stage.
[0054] S22, Cloud-based physical information augmented Transformer model training, output attention weight matrix.
[0055] This process is executed in the background of a cloud server. It adopts a mode that combines regular full training with event-triggered incremental training to learn the long-term global operating conditions and outputs attention weights that conform to the fluid mechanism through physical prior constraints.
[0056] Specifically, historical time-series data from long-term system operation is collected, including complete multi-dimensional feature sequences and corresponding physical residual vectors. When offline high-precision flow meter calibration data exists, flow truth labels are simultaneously accessed to form a training sample set. The time window for a single training sample is on the order of 10 minutes, covering the complete cycle of operating conditions and supporting the model in learning long-term dependencies.
[0057] In this embodiment, a Transformer Encoder structure is adopted, which uses a multi-head self-attention mechanism to capture the correlation between various feature dimensions and the global change pattern in long time series, making up for the deficiency of edge TCN in only capturing short-term dependencies, and realizing feature modeling across working conditions and long periods.
[0058] The model training uses a dual-objective loss function: ; Where Loss is the loss value, MSE(Q) pred Q true ) represents the mean squared error loss in flow forecasting, ||W att - W phy ||1 represents the L1 regularization term of the attention weights and the preset physical attention template, λ is the regularization coefficient, and Q pred Q is the predicted flow rate. true W represents the actual traffic value. att W is the attention weight matrix. phy This is the physical attention template matrix.
[0059] The first term is the mean squared error loss for flow prediction, which drives the model to accurately fit the flow change pattern; the second term is the physical attention regularization term, W. phy The pre-defined physical attention template matrix, based on the fluid dynamics sensitivity analysis of the pump system, characterizes the weight of each feature dimension on the actual physical impact of flow control; the attention weight matrix W is learned through the L1 norm constrained model. att It closely follows physical priors, avoiding false correlations and uncontrollability caused by purely data-driven approaches.
[0060] Under normal operating conditions, full training is performed at fixed intervals; when a sudden change in operating conditions is triggered in subsequent steps, incremental training is performed using the uploaded abnormal operating condition data.
[0061] After training, the output consists of two types of parameters: the attention weight matrix W. att It represents the importance of each feature dimension and the strength of the correlation between features; the updated TCN convolution kernel parameters are optimized to iteratively improve the accuracy of edge-side feature extraction.
[0062] S23, local attention weighted fusion, generates a comprehensive feature vector.
[0063] This sub-step is executed cycle by cycle in the local embedded controller, and the attention weights distributed from the cloud are used to weight and enhance the short-term features to obtain a comprehensive feature vector that incorporates global physical attention.
[0064] After training in the cloud, the attention weight matrix W is transmitted via Ethernet. att The updated TCN parameters are delivered to the embedded controller via OTA. In each control cycle, the short-time feature vector F of the edge TCN output is processed using the currently active attention weight matrix. fast Perform a feature weighting transformation to calculate the comprehensive feature vector F = W att ·F fast By using weighted operations, the contribution of feature dimensions that have a significant physical impact on flow control is enhanced, while the interference of irrelevant features is suppressed, so that the final features contain both short-term dynamic information and conform to global physical causal laws.
[0065] S3, based on the viscosity weighting coefficient, pressure pulsation intensity coefficient and forward predicted flow deviation in the comprehensive feature vector, calculates PID parameters according to the PID adaptive law to generate control quantities to drive the pump motor.
[0066] This step, based on the physical guidance coefficients extracted from the aforementioned comprehensive feature vector, dynamically adjusts the proportional, integral, and derivative gains of the PID controller in real time according to the parameter adaptive law that conforms to the laws of fluid dynamics. This generates a control quantity adapted to the current operating conditions to drive the pump motor, achieving high-precision closed-loop control of the flow rate. Specifically, this includes: S31, Analysis of control guidance coefficients.
[0067] This sub-step extracts three guiding variables from the integrated feature space and attention weights to drive PID parameter adjustment.
[0068] Viscosity weight coefficient extraction. Attention weight matrix W delivered from the cloud. attIn the process, the weight rows corresponding to the feature dimensions related to the viscosity of the mobile phase are extracted, the mean of the weights in that row is calculated, and then the normalized viscosity weight coefficient w is obtained through a sigmoid nonlinear mapping. η The value ranges from 0 to 1, representing the degree of influence of viscosity on flow control under the current operating conditions.
[0069] Pressure pulsation intensity coefficient extraction. From the comprehensive feature vector F, the dimensional components corresponding to the pressure pulsation features are analyzed to obtain the pressure pulsation intensity coefficient w. ΔP This indicates the intensity of the current pressure pulsation.
[0070] Forward prediction flow deviation acquisition. Retrieve the forward one-step prediction flow deviation e from the edge TCN network prediction head output in step S2. q (t+1) is the predicted flow deviation value for the next control cycle, which is used for advance adjustment of the derivative element.
[0071] S32, PID parameter adaptive law calculation.
[0072] This sub-step is based on three guiding coefficients, combined with real-time physical quantities and soft measurement values, and calculates the three parameters of the PID controller in real time according to the preset adaptive law formula, so as to realize the dynamic adaptive adjustment of parameters according to the operating conditions.
[0073] The adaptive law for the PID parameters is: ; ; ; Where Kp(t) is the real-time proportional gain coefficient at time t, Kp base This is the baseline value for the proportionality coefficient. w is the weighting coefficient for the viscosity correction term. η This is the viscosity weighting coefficient. This is the equivalent viscosity measurement of the mobile phase. w is the weighting coefficient for the pressure pulsation correction term. ΔP ΔP / Δt is the pressure pulsation intensity coefficient, and ΔP / Δt is the pressure change rate; Ki(t) is the real-time integral gain coefficient at time t. base The baseline value for the integral coefficient, T is the weighting factor for the motor temperature correction term. motor This is the actual measured value of the motor temperature. Let ∫e(t)dt be the integral value of the flow deviation, and Kd(t) be the weighting coefficient of the deviation integral correction term. base The base value for the differential coefficients, e is the weighting coefficient for the prediction deviation change rate correction term. q(t+1) represents the forward flow prediction deviation.
[0074] The five correction coefficients λ1-λ5 mentioned above were obtained through offline calibration and stored in the embedded controller in the form of a working condition lookup table. During runtime, the corresponding coefficient is called according to the current working condition range. The calculation can be completed with only simple arithmetic operations, ensuring millisecond-level real-time control efficiency.
[0075] S33, control quantity generation and limiting output.
[0076] This step calculates the final motor drive control quantity based on real-time PID parameters and flow deviation, performs limiting protection, and outputs it to the pump motor driver.
[0077] The deviation between the target flow rate and the real-time estimated flow rate is calculated and substituted into the standard PID control law formula to calculate the motor drive control quantity. The calculated control quantity is then limited to upper and lower limits to confine it within the rated input range of the motor driver, preventing overload of the motor or abnormal pump pressure caused by exceeding the control range, thus ensuring system safety.
[0078] The limited control quantity is output to the pump motor driver, which drives the motor to adjust its speed and phase to achieve closed-loop regulation of the flow rate. At the same time, the current control quantity is fed back to the third-order nonlinear extended state observer in step S1 as the control input of the observer to support the state estimation of the next cycle.
[0079] S4, Residual-driven event triggered, online calculation of the absolute value of the residual between the predicted flow of the temporal convolutional network and the instantaneous flow in the soft measurement value, statistical calculation of the mean and standard deviation of the residual within a set time window, using the sum of the mean and three times the standard deviation as a dynamic threshold, when the residual is greater than the dynamic threshold, it is determined as a sudden change in operating condition.
[0080] This step involves the system's anomaly monitoring and adaptive evolution. By comparing the predictions of the data-driven model with the soft measurement results of the physical model in real time, it dynamically generates anomaly thresholds based on statistical principles to accurately identify abrupt changes in mobile phase viscosity, air bubble ingress into the flow path, and sudden changes in column back pressure. Upon triggering, it executes a tiered response of transient safety fallback and long-term model evolution, ensuring operational reliability under extreme conditions while achieving continuous iterative improvement in model performance. Figure 5 As shown, it specifically includes: S41, Real-time residual calculation.
[0081] This sub-step is executed synchronously in each control cycle, calculating the deviation between the flow prediction value of the temporal convolutional network and the flow soft measurement value of the extended state observer, which serves as a characterization of abnormal operating conditions.
[0082] The current-time flow prediction value output by the edge-side temporal convolutional network is used as the expected output of the data-driven model, and the real-time estimated flow obtained by combining the third-order nonlinear extended state observer with the physical feedforward model in step S1 is used as the actual flow benchmark for physical observation. Under normal operating conditions, the deviation between model prediction and physical observation is small and stable; when unexpected changes occur in the operating conditions, exceeding the coverage of model training, the prediction deviation will increase significantly. Therefore, the residuals of the two can effectively characterize the degree of abnormality in the operating conditions.
[0083] Calculate the absolute value of the difference between the two flow rates, and use it as the residual at the current moment.
[0084] S42, Sliding time window residual statistics and dynamic threshold generation.
[0085] This sub-step uses a sliding window to statistically analyze the historical distribution characteristics of residuals, and adaptively generates anomaly detection thresholds based on the 3σ statistical principle to replace fixed thresholds and adapt to the baseline deviation levels under different operating conditions.
[0086] In this embodiment, a sliding time window of 100ms is used, corresponding to residual samples of 100 consecutive control cycles. The window slides forward one position in each control cycle, removing the earliest residual point and incorporating the latest calculated residual, so as to realize real-time rolling update of window data.
[0087] The statistical characteristics of all residual samples within the current sliding window are calculated in real time. The residual mean μ is the arithmetic mean of all residual values within the window, reflecting the baseline deviation level of the model prediction under the current operating conditions. The residual standard deviation σ is the standard deviation of all residual values within the window, reflecting the degree of fluctuation and dispersion of the residuals.
[0088] Based on the 3σ criterion of mathematical statistics, the threshold for determining sudden changes in operating conditions, Th = μ + 3σ, is calculated.
[0089] S43, Operating Condition Change Judgment and Operation Mode Branch.
[0090] This sub-step determines the current operating condition by comparing the real-time residual with the dynamic threshold and executes the corresponding control branch.
[0091] The real-time residual r at the current moment is compared with the dynamic threshold Th. If r ≤ Th, it is determined to be a normal operating condition. The system maintains the conventional adaptive control mode and continues to execute the real-time flow control process of steps S1 to S3 without triggering an event response.
[0092] If r > Th, it is determined that a sudden change in operating conditions has occurred. This corresponds to unusual operating conditions such as a sudden change in the mobile phase ratio leading to a sudden change in viscosity, the entry of air bubbles into the flow path, and an abnormal increase in the back pressure of the chromatographic column. A trigger signal is immediately output to start the event response mechanism.
[0093] S44, Execution of hierarchical response to mutation events.
[0094] This sub-step implements a two-level response to sudden changes in operating conditions, taking into account both transient operational safety and long-term model performance optimization.
[0095] Level 1: Conservative PID parameter switching.
[0096] Upon receiving an event trigger signal, the system immediately retrieves the pre-stored abnormal operating condition parameter table in the embedded controller. This table contains multiple sets of conservative PID parameters pre-calibrated for different viscosity and pressure ranges. Based on the current soft-sensor value of the equivalent viscosity of the mobile phase and its rate of change, the most suitable set of conservative PID parameters is matched (for example, the equivalent viscosity is pre-divided into several ranges: low viscosity, medium viscosity, and high viscosity; for each range, a set of conservative PID parameters that can stabilize the system is pre-determined experimentally), temporarily overriding the currently adaptively calculated PID parameters. By using conservative control parameters with higher stability margins, risks such as control oscillations, pressure overshoot, and flow runaway caused by model mismatch are avoided, ensuring the safe operation of the system under extreme conditions.
[0097] The second level is long-term response optimization.
[0098] Simultaneously with parameter switching, a data upload mechanism is triggered, uploading feature cache data (including multi-dimensional feature sequences, residual sequences, and operating condition parameter labels) for a preset duration before and after the abrupt change to the cloud server. Incremental training of the physically augmented Transformer model is initiated, utilizing abnormal operating condition samples to optimize the model's attention weight matrix and the convolutional kernel parameters of the temporal convolutional network, expanding the model's operating condition coverage. After training, the updated model parameters are pushed to the local embedded controller via OTA (Over-The-Air) updates, completing the model's online iterative evolution and improving prediction accuracy and control performance under similar operating conditions in subsequent iterations.
[0099] Example 2: This embodiment provides an adaptive flow control system for a liquid chromatography pump based on dual-path network collaboration and soft measurement, including: A multidimensional feature vector fusion module is used to acquire measurable physical signals from a liquid chromatography pump and construct a third-order nonlinear extended state observer. Based on some measurable physical signals and the pressure change rate estimate and total system disturbance estimate output by the third-order nonlinear extended state observer, soft measurement values are estimated. The measurable physical signals and soft measurement values are combined into a multidimensional feature vector. The comprehensive feature vector extraction module is used to encode the multidimensional feature vector using a temporal convolutional network to obtain a short-term feature vector; to train the Transformer model using historical data and physical residuals, and to output an attention weight matrix from the trained Transformer model; and to weight and fuse the attention weight matrix and the short-term feature vector to obtain a comprehensive feature vector. The PID control module is used to calculate PID parameters according to the PID adaptive law based on the viscosity weighting coefficient, pressure pulsation intensity coefficient and forward predicted flow deviation in the comprehensive feature vector, so as to generate control quantities to drive the pump motor.
[0100] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0101] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0102] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A method for adaptive flow control of a liquid chromatography pump, characterized in that, Includes the following steps: Measurable physical signals from a liquid chromatography pump are collected, and a third-order nonlinear extended state observer is constructed. The measurable physical signals include: pump outlet pressure, pressure change rate, cam phase, motor speed, motor temperature, and ambient temperature. The third-order nonlinear extended state observer is as follows: ; ; ; ; Where e is the state observation error, z1 is the estimated value of the pump outlet pressure, z2 is the estimated value of the pressure change rate, z3 is the estimated value of the total system disturbance, P is the pump outlet pressure, u is the input control quantity, b0 is the system control gain coefficient, fal() is the nonlinear feedback coefficient, and δ is the threshold of the linear segment of the fal() function.
1.
2. 3 are the time derivatives of z1, z2, and z3, respectively, and β1-β3 is the error feedback gain coefficient; Based on the measurable physical signals and the estimated pressure change rate and total system disturbance output by the third-order nonlinear expansion state observer, soft measurement values are estimated; the measurable physical signals and soft measurement values are combined into a multidimensional feature vector; the soft measurement values include: real-time estimated flow rate, equivalent viscosity of the mobile phase, and equivalent damping coefficient; The multidimensional feature vector is encoded using a temporal convolutional network to obtain a short-time feature vector; The Transformer model is trained using historical data and physical residuals, and the attention weight matrix is output by the trained Transformer model. The attention weight matrix and the short-term feature vector are weighted and fused to obtain a comprehensive feature vector; Based on the viscosity weighting coefficient, pressure pulsation intensity coefficient, and forward predicted flow deviation in the comprehensive feature vector, PID parameters are calculated by combining real-time acquired measurable physical signals and soft measurement values to generate control quantities to drive the pump motor. The PID parameters include: real-time proportional gain coefficient, real-time integral gain coefficient, and real-time derivative gain coefficient.
2. The adaptive flow control method for a liquid chromatography pump according to claim 1, characterized in that, The temporal convolutional network is a lightweight TCN, which contains four layers of expanded convolutions with expansion rates of 1, 2, 4, and 8 respectively. Each layer has thirty-two filters and uses the ReLU activation function. Its input is a multi-dimensional feature vector of the most recent thirty-two time steps, and its output is a short-term feature vector of one hundred and twenty-eight dimensions.
3. The adaptive flow control method for a liquid chromatography pump according to claim 1, characterized in that, The Transformer model is deployed in the cloud and adopts an encoder structure. The input is a long historical window sequence and the corresponding physical residual vector. The physical attention template is obtained from system sensitivity analysis. After training, the attention weight matrix and the convolution kernel parameters of the lightweight temporal convolutional network are extracted.
4. The adaptive flow control method for a liquid chromatography pump according to claim 3, characterized in that, The loss function of the Transformer model is: ; Where Loss is the loss value, MSE(Q) pred Q true ) represents the mean squared error loss in flow forecasting, ||W att - W phy ||1 represents the L1 regularization term of the attention weights and the preset physical attention template, λ is the regularization coefficient, and Q pred Q is the predicted flow rate. true W represents the actual traffic value. att W is the attention weight matrix. phy This is the physical attention template matrix.
5. The adaptive flow control method for a liquid chromatography pump according to claim 1, characterized in that, The calculation formulas for the PID parameter pairs are as follows: ; ; ; Where Kp(t) is the real-time proportional gain coefficient at time t, Kp base This is the baseline value for the proportionality coefficient. w is the weighting coefficient for the viscosity correction term. η This is the viscosity weighting coefficient. This is the equivalent viscosity measurement of the mobile phase. w is the weighting coefficient for the pressure pulsation correction term. ΔP ΔP / Δt is the pressure pulsation intensity coefficient, and ΔP / Δt is the pressure change rate; Ki(t) is the real-time integral gain coefficient at time t. base The baseline value for the integral coefficient, T is the weighting factor for the motor temperature correction term. motor This is the actual measured value of the motor temperature. Let ∫e(t)dt be the integral value of the flow deviation, and Kd(t) be the weighting coefficient of the deviation integral correction term. base The base value for the differential coefficients, e is the weighting coefficient for the prediction deviation change rate correction term. q (t+1) represents the forward flow prediction deviation.
6. The adaptive flow control method for a liquid chromatography pump according to claim 1, characterized in that, It also includes residual-driven event triggering steps: The absolute value of the residual between the predicted flow of the temporal convolutional network and the instantaneous flow in the soft measurement value is calculated online. The mean and standard deviation of the residual within a set time window are statistically analyzed. The sum of the mean and three times the standard deviation is used as a dynamic threshold. When the residual is greater than the dynamic threshold, it is determined to be a sudden change in operating condition.
7. The adaptive flow control method for a liquid chromatography pump according to claim 6, characterized in that, After determining that the operating condition has changed abruptly, based on the current equivalent viscosity and its rate of change in the soft measurement values, a set of pre-stored conservative PID parameters is selected to cover the current PID parameters; the feature cache before and after the operating condition change is used for incremental training of the Transformer model to update the attention weight matrix.
8. An adaptive flow control system for a liquid chromatography pump, characterized in that, include: A multidimensional feature vector fusion module is used to acquire measurable physical signals from the liquid chromatography pump and construct a third-order nonlinear extended state observer. Based on the measurable physical signals and the pressure change rate estimate and total system disturbance estimate output by the third-order nonlinear extended state observer, soft measurement values are estimated. The measurable physical signals and soft measurement values are combined into a multidimensional feature vector. The measurable physical signals include: pump outlet pressure, pressure change rate, cam phase, motor speed, motor temperature, and ambient temperature. The comprehensive feature vector extraction module is used to encode the multidimensional feature vector using a temporal convolutional network to obtain a short-term feature vector; to train the Transformer model using historical data and physical residuals, and to output an attention weight matrix from the trained Transformer model; and to weight and fuse the attention weight matrix and the short-term feature vector to obtain a comprehensive feature vector. The PID control module is used to calculate PID parameters according to the PID adaptive law based on the viscosity weighting coefficient, pressure pulsation intensity coefficient and forward predicted flow deviation in the comprehensive feature vector, so as to generate control quantities to drive the pump motor.
Citation Information
Patent Citations
Liquid chromatogram flow velocity real-time detection and optimal control method based on multi-point sensing
CN120065760A
