Component collection delay volume correction and optimization method based on artificial intelligence

By employing an AI-based delayed volume correction and optimization method, the valve switching time can be predicted and controlled in real time, solving the problem of inaccurate component collection timing in traditional methods, improving the purity and recovery rate of protein purification, and achieving automated operation.

CN121879130APending Publication Date: 2026-04-17CHANGZHOU SMART LIFESCI CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610006950.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-05
Publication Date
2026-04-17

AI Technical Summary

Technical Problem

Traditional methods cannot reflect changes in system status in real time, resulting in delayed or premature switching of component collection valves, which affects the purity and recovery rate of target components.

Method used

An AI-based component collection delay volume correction and optimization method is adopted. By acquiring chromatographic signals and system parameters in real time, a delay volume model is established. The AI ​​model is used to predict the peak arrival time and dynamically control the valve switching time. The model parameters are optimized by combining feedback learning.

Benefits of technology

It enables precise control of component collection, improves the purity and recovery rate of target protein components, supports automated operation, and reduces human adjustment errors.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121879130A_ABST
    Figure CN121879130A_ABST
Patent Text Reader

Abstract

The invention discloses a component collection delay volume correction and optimization method based on artificial intelligence. The method comprises the steps of S1, signal collection; s2, preprocessing the collected signals; s3, carrying out delayed volume modeling to form a training sample set; s4, training an AI prediction module by using the training sample set; s5, AI model training is carried out, and peak arrival time is predicted; s6, a control valve carries out accurate component collection; and S7, inputting the difference between the actual peak arrival time and the predicted time into an AI model for feedback optimization. According to the method, signals are collected and preprocessed, a delay volume model is established, AI model training is carried out, peak arrival prediction time is obtained, accurate collection of valve control is achieved, the difference between the prediction time and the actual peak arrival time is input into the AI model for feedback optimization, dynamic intelligent correction of the delay volume is achieved, manual adjustment errors are remarkably reduced, and the accuracy of valve control is improved. The collection opportunity can be optimized in real time according to flow velocity change and system drift, and the purity and recovery rate of target protein components are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of automated control technology for protein purification and chromatographic separation, and in particular to a method for component collection delay volume correction and optimization based on artificial intelligence. Background Technology

[0002] In protein purification, component collection is typically based on the peak position in chromatographic detection signals (such as UV, Vis, or fluorescence signals). Due to factors such as system volume, flow rate fluctuations, and temperature changes, there is a certain volume delay error between the appearance time of the detection signal peak and the time when the component arrives at the collector. Traditional methods often use fixed delay times or empirical correction methods to delay the volume. However, fixed delay times or empirical correction methods cannot reflect changes in system status in real time, leading to delayed or premature switching of the collection valve, which affects the purity and recovery rate of the target component. To solve the above problems, an artificial intelligence-based component collection delay volume correction and optimization method is proposed. Summary of the Invention

[0003] The purpose of this invention is to provide an artificial intelligence-based method for correcting and optimizing the volume of component collection delay, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a component collection delay volume correction and optimization method based on artificial intelligence, wherein the correction and optimization method includes the following operational steps: S1, perform signal acquisition, and acquire ultraviolet / visible light detection signals and flow rate, pressure and temperature parameters in real time through the chromatographic signal acquisition module; S2, perform preprocessing of the acquired signal, and use the data preprocessing module to filter, identify peaks and standardize the signals acquired by the acquisition module; S3, perform delay volume modeling, read historical peak data from the delay volume database, calculate the delay ΔV between the appearance time of the detection signal peak and the arrival time of the collector peak, and form a training sample set; S4 involves training the AI ​​model using the training sample set generated in S3. The AI ​​prediction module is trained with the current protein purification system status and real-time peak parameters as inputs, and the peak arrival time at the collector is predicted as the output. ; S5 performs real-time prediction and control. During the purification process, the AI ​​model receives newly acquired signal inputs in real time and predicts the peak arrival time. Then, the dynamic control module calculates the valve switching advance Δt based on the predicted time and sends control commands to the component collection control unit. S6 performs precise collection. Upon receiving the control command, the component collection control unit performs time-based collection. Switching is triggered at −Δt, synchronizing the arrival time of the peak signal and the component; S7 performs feedback optimization. The difference between the actual peak arrival time and the predicted time is input into the AI ​​model through the feedback learning module, and the AI ​​model parameters are automatically updated to perform model self-learning and accuracy optimization.

[0005] Preferably, the chromatographic signal acquisition module is connected to the chromatography system via an analog / digital data acquisition card or a PLC interface, and the acquired data is transmitted to the data preprocessing module in time series form.

[0006] Preferably, the preprocessing of the acquired ultraviolet / visible light detection signals and flow rate, pressure, and temperature parameters includes the following steps: S201, filtering and noise reduction, uses Savitzky-Golay smoothing filter or one-dimensional convolution smoothing algorithm to remove high-frequency noise from the signal; S202, Baseline Correction, uses moving average or polynomial fitting methods to perform dynamic baseline correction and eliminate drift; S203, Peak identification, uses the first derivative method or threshold search algorithm to locate the peak start point, peak apex and peak end point; S204, Feature Extraction and Standardization, calculates peak height, peak width, peak area, symmetry, and rise / fall time feature parameters for each detection peak, and performs normalization processing to provide stable input for the AI ​​model.

[0007] Preferably, the delayed volume database stores historical experimental data and physical parameters of the protein purification system, and the delayed volume database is used to support the retraining and adaptive optimization of the AI ​​model.

[0008] Preferably, the AI ​​prediction module takes an input feature vector X containing multi-dimensional features as input, processes it, and outputs the predicted arrival time point of the corresponding peak in the collector or the corresponding delay volume. .

[0009] Preferably, the AI ​​prediction module combines a long short-term memory network with a fully connected neural network to simultaneously capture time-dependent features and nonlinear mapping relationships.

[0010] Preferably, the network structure of the AI ​​prediction module adopts the Adam optimizer with a learning rate of 0.001 and incorporates an early stopping strategy. During the AI ​​model training process, supervised learning is performed by inputting data into a hybrid structure that combines a long short-term memory network and a fully connected neural network to obtain the parameter weights θ.

[0011] Preferably, the dynamic control module uses a time-triggered mechanism based on a real-time clock or PLC timing control logic to calculate the valve switching advance Δt.

[0012] Preferably, the component collection control unit performs the operation of switching target collection channels, controlling the sequential replacement or movement of collection tubes to precisely correspond to the outflow time of the target peak, based on the control signal output by the dynamic control module, thereby realizing the component graded collection operation.

[0013] Preferably, the operation steps of the feedback learning module are as follows: S701, comparing AI-predicted peak arrival times Compared with the actual detected ; S702, calculate the deviation ΔT= ; S703, write the deviation ΔT and the corresponding operating conditions into the delay volume database; S704 updates AI model parameters regularly or online to improve future prediction accuracy.

[0014] The technical effects and advantages of this invention are as follows: This invention preprocesses chromatographic detection signals and system operating parameters to establish a delay volume model. The training sample set generated from this delay volume model is then used to train an AI model to obtain the peak arrival prediction time. Precise collection is achieved by controlling the valve based on the valve switching advance. The difference between the predicted and actual peak arrival times is input into the AI ​​model for feedback optimization. This solution realizes dynamic intelligent correction of the delay volume, significantly reducing manual adjustment errors. It can optimize the collection timing in real time based on flow rate changes and system drift, improving the purity and recovery rate of target protein components. It also supports automated operation and reduces operator dependence. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the operational implementation of the delay volume correction and optimization method of the present invention. Detailed Implementation

[0016] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention provides, for example Figure 1 The illustrated method for correcting and optimizing component collection delay volume based on artificial intelligence includes the following steps: S1, perform signal acquisition, and acquire ultraviolet / visible light detection signals and flow rate, pressure and temperature parameters in real time through the chromatographic signal acquisition module; The chromatographic signal acquisition module is connected to the chromatography system via an analog / digital data acquisition card or PLC interface, and the acquired data is transmitted to the data preprocessing module in time series form.

[0018] Specifically, the acquired signals include: ultraviolet / visible absorption signals, fluorescence detection signals, and operating parameters of the protein purification system: flow rate, pressure, temperature, pump speed, and valve status; the sampling frequency of the chromatographic signal acquisition module is usually between 1 and 10 Hz to ensure sufficient time resolution to capture complete peak shape changes; S2, perform preprocessing of the acquired signal, and use the data preprocessing module to filter, identify peaks and standardize the signals acquired by the acquisition module; The preprocessing of the acquired UV / Vis detection signals and flow rate, pressure, and temperature parameters includes the following steps: S201, filtering and noise reduction, uses Savitzky-Golay smoothing filter or one-dimensional convolution smoothing algorithm to remove high-frequency noise from the signal; S202, Baseline Correction, uses moving average or polynomial fitting methods to perform dynamic baseline correction and eliminate drift; S203, Peak identification, uses the first derivative method or threshold search algorithm to locate the peak start point, peak apex and peak end point; S204, Feature Extraction and Standardization, calculates peak height, peak width, peak area, symmetry, and rise / fall time feature parameters for each detection peak, and performs normalization processing to provide stable input for the AI ​​model.

[0019] Normalization is performed using Z-score normalization or Min-Max normalization; S3, perform delay volume modeling, read historical peak data from the delay volume database, calculate the delay ΔV between the appearance time of the detection signal peak and the arrival time of the collector peak, and form a training sample set; The delayed volume database stores historical experimental data and physical parameters of the protein purification system. It supports the retraining and adaptive optimization of the AI ​​model. This database serves as both a training sample library for the AI ​​model and a long-term memory module for the system's self-learning. The delayed volume database includes: the type of mobile phase, viscosity, and gradient program for each run; Pipeline volume, connecting pipe length, pump type; detection signal peak appearance time and actual peak arrival time in the collector; protein purification requirements for system operating conditions (flow rate, temperature, pressure) and corresponding delay volume ΔV.

[0020] S4 involves training the AI ​​model using the training sample set generated in S3. The AI ​​prediction module is trained with the current protein purification system status and real-time peak parameters as inputs, and the peak arrival time at the collector is predicted as the output. ; Specifically, the input features of the model include: flow velocity F(t), pressure P(t), and temperature T(t); Peak width, peak height, and symmetry parameters of real-time ultraviolet / visible light signal S(t); historical delay volume ΔV historical sequence; model output: the estimated time T1 for peak arrival at the collector; The AI ​​prediction module takes an input feature vector X containing multidimensional features as input, processes it, and outputs the predicted arrival time of the corresponding peak at the collector or the corresponding delay volume. .

[0021] The feature vector X contains multidimensional features of the following categories: real-time operating parameters: flow rate F(t), pressure P(t), temperature T(t); signal features: peak height H, peak width W, peak area A, symmetry R, ​​signal slope dS / dt; historical state features: the delay volume ΔV1…ΔV of the most recent n runs. n System structural parameters: Piping volume Buffer type encoding The final input tensor is in the form of [T×F], where T is the time step and F is the feature dimension.

[0022] The AI ​​prediction module combines a long short-term memory network with a fully connected neural network to capture both time-dependent features and nonlinear mapping relationships. Its network structure is as follows: Input Layer: Receive time series input data X ∈ R^(T×F).

[0023] LSTM encoding layer (Feature Encoder): First layer LSTM: Extracts the temporal dynamic features of the signal, number of hidden units. ; The second LSTM layer further encodes long-term dependencies and the number of hidden units. ; The output is a hidden state vector. It is used to reflect the changes in peak shape and system response.

[0024] Feature concatenation layer: Compare the LSTM output with the system static parameters ( , The features are concatenated to form a comprehensive feature vector Z.

[0025] Fully connected prediction layer (MLP): Dense(64, ReLU) → Dense(32, ReLU) → Dense(1, Linear) Output prediction results or .

[0026] Loss function: Mean squared error (MSE) loss is used: (1) Where L is the final loss value, and N represents the total number of samples. This represents the predicted peak arrival time for the i-th sample. This represents the actual peak arrival time of the i-th sample, where i is the sample index.

[0027] Optimization algorithm: We used the Adam optimizer with a learning rate of 0.001 and incorporated an early stopping strategy to prevent overfitting.

[0028] S5 performs real-time prediction and control. During the purification process, the AI ​​model receives newly acquired signal inputs in real time and predicts the peak arrival time. Then, the dynamic control module calculates the valve switching advance Δt based on the predicted time and sends control commands to the component collection control unit. The main functions of the dynamic control module include: The system calculates the liquid transfer time from the detector to the collector in real time based on the flow rate F(t); a control signal is generated at time T1−Δt to trigger valve switching; ensuring that the component flow path is accurately established before the arrival of the main component peak, thus achieving synchronous collection. This module can employ a time-triggered mechanism based on a real-time clock (RTC) or PLC timing control logic.

[0029] Specifically, the component collection control unit executes the switching of target collection channels, the sequential replacement or movement of collection tubes, and the precise corresponding target peak outflow time based on the control signals output by the dynamic control module, thereby realizing the operation of component graded collection.

[0030] It should be noted that the component collection and control unit consists of a multi-channel solenoid valve, a stepper motor driver, and a pipeline distribution device.

[0031] S6 performs precise collection. Upon receiving the control command, the component collection control unit performs time-based collection. Switching is triggered at −Δt, synchronizing the arrival time of the peak signal and the component; S7 performs feedback optimization. The difference between the actual peak arrival time and the predicted time is input into the AI ​​model through the feedback learning module, and the AI ​​model parameters are automatically updated to perform model self-learning and accuracy optimization.

[0032] The feedback learning module verifies and dynamically corrects the AI ​​predictions in real time, forming a closed-loop control. The steps for operating the feedback learning module are as follows: S701, comparing AI-predicted peak arrival times Compared with the actual detected ; S702, calculate the deviation ΔT= ; S703, write the deviation ΔT and the corresponding operating conditions into the delay volume database; S704 updates AI model parameters regularly or online to improve future prediction accuracy.

[0033] Through this mechanism, the system possesses adaptive learning and self-evolution capabilities, and can automatically adjust the prediction model under different system and pipeline conditions to maintain high accuracy over a long period of time.

[0034] AI model training and inference process: Training phase: Read historical sample data from the delayed volume database; Perform data normalization and sliding window sampling; The LSTM-MLP hybrid model is input for supervised learning, and the parameter weights θ are obtained through training. The validation set is used to evaluate the model's prediction accuracy, and the optimal model is saved.

[0035] Online prediction phase: During protein purification, the AI ​​model receives newly acquired signal features in real time. The AI ​​model calculates the estimated arrival time T1 of the target peak based on the current system state. The predicted results are output to the dynamic control module for calculating the valve switching time Δt.

[0036] Self-learning and updating phase: After the actual data collection is completed, the feedback module compares the predicted T1 with the actual T1. Difference; Error samples are written into a database, and model parameters are periodically fine-tuned to enable online model learning.

[0037] Specifically, the dynamic control module uses a time-triggered mechanism based on a real-time clock or PLC timing control logic to calculate the valve switching advance Δt. The dynamic control module calculates the required valve switching advance Δt based on the target peak arrival time T1 output by the AI ​​prediction results. This invention preprocesses chromatographic detection signals and system operating parameters to establish a delay volume model. The training sample set generated from this delay volume model is then used to train an AI model to obtain the peak arrival prediction time. Precise collection is achieved by controlling the valve based on the valve switching advance. The difference between the predicted and actual peak arrival times is input into the AI ​​model for feedback optimization. This solution realizes dynamic intelligent correction of the delay volume, significantly reducing manual adjustment errors. It can optimize the collection timing in real time based on flow rate changes and system drift, improving the purity and recovery rate of target protein components. It also supports automated operation and reduces operator dependence.

[0038] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for correcting and optimizing the volume of component collection delay based on artificial intelligence, characterized in that, The correction and optimization method includes the following steps: S1, perform signal acquisition, and acquire ultraviolet / visible light detection signals and flow rate, pressure and temperature parameters in real time through the chromatographic signal acquisition module; S2, perform preprocessing of the acquired signal, and use the data preprocessing module to filter, identify peaks and standardize the signals acquired by the acquisition module; S3, perform delay volume modeling, read historical peak data from the delay volume database, calculate the delay ΔV between the appearance time of the detection signal peak and the arrival time of the collector peak, and form a training sample set; S4 involves training the AI ​​model using the training sample set generated in S3. The AI ​​prediction module is trained with the current protein purification system status and real-time peak parameters as inputs, and the peak arrival time at the collector is predicted as the output. ; S5 performs real-time prediction and control. During the purification process, the AI ​​model receives newly acquired signal inputs in real time and predicts the peak arrival time. Then, the dynamic control module calculates the valve switching advance Δt based on the predicted time and sends control commands to the component collection control unit. S6 performs precise collection. Upon receiving the control command, the component collection control unit performs time-based collection. Switching is triggered at −Δt, synchronizing the arrival time of the peak signal and the component; S7 performs feedback optimization. The difference between the actual peak arrival time and the predicted time is input into the AI ​​model through the feedback learning module, and the AI ​​model parameters are automatically updated to perform model self-learning and accuracy optimization.

2. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 1, characterized in that, The chromatographic signal acquisition module is connected to the chromatography system via an analog / digital data acquisition card or a PLC interface, and the acquired data is transmitted to the data preprocessing module in time series form.

3. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 2, characterized in that, The preprocessing of the acquired ultraviolet / visible light detection signals and flow rate, pressure, and temperature parameters includes the following steps: S201, filtering and noise reduction, uses Savitzky-Golay smoothing filter or one-dimensional convolution smoothing algorithm to remove high-frequency noise from the signal; S202, Baseline Correction, uses moving average or polynomial fitting methods to perform dynamic baseline correction and eliminate drift; S203, Peak identification, uses the first derivative method or threshold search algorithm to locate the peak start point, peak apex and peak end point; S204, Feature Extraction and Standardization, calculates peak height, peak width, peak area, symmetry, and rise / fall time feature parameters for each detection peak, and performs normalization processing to provide stable input for the AI ​​model.

4. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 1, characterized in that, The delayed volume database stores historical experimental data and physical parameters of the protein purification system. The delayed volume database is used to support the retraining and adaptive optimization of AI models.

5. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 1, characterized in that, The AI ​​prediction module takes an input feature vector X containing multi-dimensional features as input, processes it, and outputs the predicted arrival time of the corresponding peak at the collector. or corresponding delay volume .

6. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 5, characterized in that, The AI ​​prediction module combines a long short-term memory network with a fully connected neural network to capture time-dependent features and nonlinear mapping relationships.

7. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 6, characterized in that, The AI ​​prediction module uses the Adam optimizer with a learning rate of 0.001 and incorporates an early stopping strategy. During AI model training, supervised learning is performed by inputting data into a hybrid structure that combines a long short-term memory network and a fully connected neural network to obtain the parameter weights θ.

8. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 1, characterized in that, The dynamic control module uses a time-triggered mechanism based on a real-time clock or PLC timing control logic to calculate the valve switching advance Δt.

9. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 1, characterized in that, The component collection control unit performs the operation of switching target collection channels, controlling the sequential replacement or movement of collection tubes to precisely correspond to the outflow time of the target peak, based on the control signal output by the dynamic control module, thereby realizing the component graded collection operation.

10. The method for component collection delay volume correction and optimization based on artificial intelligence according to claim 1, characterized in that, The operation steps of the feedback learning module are as follows: S701, comparing AI-predicted peak arrival times Compared with the actual detected ; S702, calculate the deviation ΔT= ; S703, write the deviation ΔT and the corresponding operating conditions into the delay volume database; S704 updates AI model parameters regularly or online to improve future prediction accuracy.