Liquid chromatography method and system based on artificial intelligence and multi-modal sensing fusion
By employing a multimodal sensing fusion liquid chromatography method, which utilizes temporal convolutional networks and Transformer models, the problem of insufficient utilization of signal correlation in existing technologies is solved. This enables efficient and accurate determination of chromatographic separation status and adaptive control, reducing human error and sample loss.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGZHOU SMART LIFESCI CO LTD
- Filing Date
- 2026-01-19
- Publication Date
- 2026-04-28
AI Technical Summary
Existing liquid chromatography techniques cannot effectively utilize the correlation of multimodal signals, resulting in biased judgment of separation status, easy misjudgment or omission, and lack of self-learning ability and adaptive control, making it difficult to adapt to experimental changes, and subject to human interpretation error and time delay.
An AI-based multimodal sensing fusion method is adopted, which uses temporal convolutional networks and Transformer models to extract and fuse features of multimodal signals, constructs an anomaly detection model, realizes adaptive control and pump speed adjustment, and combines UV, conductivity and pH signals for comprehensive state judgment.
It improves the analytical accuracy and automation level of the chromatography process, enables early warning of anomalies such as peak overlap and baseline drift, reduces the risk of sample loss, and supports model self-learning and adaptive control.
Smart Images
Figure CN121935846A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a liquid chromatography method and system based on the fusion of artificial intelligence and multimodal sensing, belonging to the field of biological separation, purification and intelligent monitoring technology. Background Technology
[0002] Currently, liquid chromatography is a core separation and purification technology in the fields of biochemistry and pharmaceutical research, widely used for the separation and analysis of proteins, nucleic acids, peptides, and metabolites. Its technological development directly affects the purification efficiency of biological samples and the accuracy of subsequent studies. With the increasing demand for automation and intelligence in biological experiments, traditional chromatography systems that rely on manual or single-parameter monitoring can no longer meet the requirements of high precision and high efficiency. There is an urgent need for monitoring systems that integrate multi-sensor signals and intelligent algorithms to optimize the chromatography process.
[0003] The core principle of liquid chromatography is to separate different substances through the interaction between the stationary phase and the mobile phase within the chromatography column. The separation state needs to be determined by changes in physicochemical parameters such as ultraviolet absorption peak (UV), conductivity (Cond), and solution pH.
[0004] Multimodal sensing technology: Integrating an optical detector (monitoring UV signal), a conductivity meter (monitoring Cond signal), and a pH electrode (monitoring pH signal), it simultaneously acquires multidimensional physicochemical parameters during chromatography, providing comprehensive data signal support for subsequent analysis.
[0005] Intelligent algorithm applications: Utilizing deep learning models (such as temporal convolutional networks TCN and Transformer) to achieve temporal feature extraction, cross-modal fusion, and abnormal state identification of multimodal signals, replacing traditional manual interpretation or single threshold judgment, and improving the accuracy and real-time performance of analysis.
[0006] Current liquid chromatography control systems generally adopt a technical solution of "single-variable judgment + fixed threshold triggering", and the specific implementation method is as follows: Data is collected for single physicochemical parameters (mostly UV absorption peaks).
[0007] A preset empirical threshold for a single parameter (such as a UV signal intensity threshold or a conductivity change threshold) is used. When the collected signal exceeds this threshold, a valve switching or pump speed adjustment action is triggered. For example, when the UV signal intensity exceeds the set value, it is determined that the target substance has flowed out, and the valve is controlled to switch to the collection channel.
[0008] The original signal is filtered (e.g., low-pass filtered), and the signal curve is observed manually in real time to determine the separation status. If any abnormalities are found (e.g., peak overlap, baseline drift), manual intervention is performed.
[0009] However, the above solution still has the following drawbacks: 1. It is impossible to perform joint analysis of multi-modal signals such as UV, conductivity, and pH, and ignores the strong correlation between parameters (such as the correlation between the increase in pre-peak conductivity and the appearance of the UV peak), which leads to a one-sided judgment of the separation state and is prone to misjudgment or omission.
[0010] 2. Chromatography is essentially a time-varying nonlinear system. Fixed threshold models cannot capture the dynamic characteristics of signal time-series changes (such as slow pH drift and gradual conductivity transitions), and are difficult to adapt to changes in different experimental samples or chromatographic conditions.
[0011] 3. Relying on manual signal interpretation introduces subjective errors and time delays. When anomalies such as peak overlap or baseline drift occur, they cannot be detected and intervened in a timely manner, which can easily lead to sample loss and experimental failure.
[0012] 4. It lacks self-learning capabilities and cannot optimize its judgment logic based on historical experimental data signals. Furthermore, it lacks an adaptive control mechanism, requiring manual adjustment of parameters such as pump speed and valves after an anomaly occurs. Summary of the Invention
[0013] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a liquid chromatography method and system based on the fusion of artificial intelligence and multimodal sensing. Through multimodal data fusion processing, the analytical accuracy, operating efficiency and automation level of the chromatography process are significantly improved, and technical support is provided for the state judgment and process control of the chromatography separation process.
[0014] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing, which includes the following steps: Step S1: Acquire multimodal sensing data signals; Step S2: Preprocess the acquired multimodal sensing data signals; Step S3: Use a temporal convolutional network to extract local temporal features from the preprocessed multimodal sensing data signal; Step S4: Input the temporal feature sequence output by the temporal convolutional network into the Transformer model to perform cross-modal feature fusion (Transformer multimodal feature fusion). Step S5: Construct an anomaly detection model based on the fused feature vector to provide early warning of abnormal states, and dynamically adjust the pump speed according to the anomaly probability.
[0015] Furthermore, the multimodal sensing data signals include ultraviolet absorption intensity data signals, conductivity data signals, and pH value data signals; The expression for the multimodal sensing data signal is as follows: in, The absorbance signal intensity of the component that produces ultraviolet absorption at a wavelength of 280 nm or 260 nm in the eluate flowing out from the outlet of the chromatography column; The conductivity data signal of the eluent at the outlet of the chromatography column; This is the real-time pH value signal of the eluent at the outlet of the chromatography column.
[0016] Furthermore, in step S2, the acquired multimodal sensing data signals are preprocessed, specifically including the following steps: Step S21: Filter and denoise the acquired multimodal sensing data signals; Step S22: Normalize and synchronously interpolate the filtered and denoised multimodal sensor data signal; Step S23: Remove outliers from the multimodal sensing data signals.
[0017] Furthermore, in step S22, the normalization and synchronous interpolation operations are performed on the filtered and denoised multimodal sensing data signal, specifically including the following steps: All multimodal sensor data signals after filtering and denoising are normalized to the [0,1] interval using Min-Max, and then interpolated using a linear interpolation function to unify all multimodal sensor data signals to the same sampling time point. The expression of the linear interpolation function is as follows: in, It is a multimode signal matrix. This represents the original multimodal signal matrix after filtering and denoising, but which is not yet aligned on the time axis. For a uniform sampling interval.
[0018] Furthermore, in step S23, removing outliers from the multimodal sensing data signal specifically includes the following steps: The 3σ criterion is used to identify outliers in multimodal sensor data signals. Outliers exceeding three standard deviations are removed, and the mean within a local window centered on the outlier is used to replace the outlier.
[0019] Furthermore, the expression for the temporal convolutional network in step S3 is: in, The kernel size is [size]. These are the parameters of the l-th convolutional kernel; This is the time step output for the l-th layer.
[0020] Furthermore, the input sequence of the Transformer model in step S4 Represented as: in, This represents the local temporal feature vector extracted by the temporal convolutional network at the T-th time step. The calculation formula for the attention mechanism of the Transformer model is as follows: in, For query, As key, For value matrices, For feature dimensions; The expression for the fused feature vector F is: Furthermore, step S5 involves constructing an anomaly detection model based on fused feature vectors to provide early warning of abnormal states and dynamically adjusting the pump speed according to the anomaly probability. This specifically includes the following steps: The fused feature vector is input into a binary classification network, which employs fully connected layers (FC layers) and a softmax function. This network is used to determine the current tomography state, which includes normal and abnormal states. The binary classification network outputs the abnormality probability. ; If the abnormal probability , If the threshold is reached, an abnormal state is determined and an adaptive control mechanism is executed, based on the abnormal probability. The pump speed is dynamically adjusted, and the formula for calculating the pump speed is as follows: ; in, To control the gain coefficient, The adjusted pump speed, This is the pump speed before adjustment.
[0021] Another aspect of the present invention provides a system for liquid chromatography using a method based on the fusion of artificial intelligence and multimodal sensing, comprising: A multimodal sensing unit, wherein the multimodal sensing unit is used to acquire multimodal sensing data signals; A data acquisition card is used to receive multimodal sensing data signals acquired by the multimodal sensing unit and send the multimodal sensing data signals to the industrial-grade main control unit. The main control unit extracts local temporal features using a temporal convolutional network and combines it with a Transformer model to achieve global dependency fusion of cross-modal features. Then, it constructs an anomaly detection model based on the fused features to provide early warning of abnormal states. Finally, it adjusts the pump speed in real time through an adaptive control algorithm based on anomaly probability. The control execution unit is used to adjust the pump speed according to the adaptive control instructions of the main control unit. An alarm unit, specifically an audible and visual alarm device used to issue an alarm based on an abnormal state warning from the main control unit.
[0022] Furthermore, the multimodal sensing unit includes a UV detector, a conductivity meter, and a pH sensor.
[0023] By adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention realizes the joint modeling of the time-series characteristics of multi-source signals such as UV value, conductivity, and pH value, explores the potential correlation between various modal signals, and comprehensively judges the chromatographic separation state.
[0024] 2. This invention captures the local temporal dependence and global modal relationship of tomography signals through a deep learning model (TCN+Transformer), adapting to the time-varying nonlinear characteristics of the tomography process.
[0025] 3. The present invention constructs an anomaly detection model, which can realize early warning of abnormal trends such as peak overlap, baseline drift, and pH instability, and avoid sample loss caused by delays in manual interpretation.
[0026] 4. This invention establishes an adaptive control closed loop, adjusts the pump speed in real time based on anomaly detection results, and supports model self-learning (based on historical log optimization) to achieve fully automatic tomographic monitoring. Attached Figure Description
[0027] Figure 1 This is a flowchart of the liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to the present invention; Figure 2 This is a schematic diagram of the system of the present invention that applies a liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing. Detailed Implementation
[0028] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.
[0029] Example 1 like Figure 1 As shown in the figure, this embodiment provides a liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing, which includes the following steps: Step S1: Acquire multimodal sensing data signals of the liquid, including ultraviolet absorption intensity data signals, conductivity data signals, and pH value data signals. Specifically: An optical detector, conductivity meter, and pH sensor are installed at the outlet of the chromatography equipment to simultaneously acquire multi-channel signals from the chromatography column output. The system sampling period is set based on the dynamic response characteristics of the chromatography system, the sampling frequency of the multimodal sensors, and the real-time control requirements. In this embodiment, the preferred system sampling period is 0.5 s. A set of multimodal sensor data signals is acquired at each moment, and the expression for the multimodal sensor data signals is as follows: in, The absorbance signal intensity of the component that produces ultraviolet absorption at a wavelength of 280 nm or 260 nm in the eluate flowing out from the outlet of the chromatography column; The conductivity data signal of the eluent at the outlet of the chromatography column; This is the real-time pH value signal of the eluent at the outlet of the chromatography column.
[0030] Step S2: Preprocess the acquired multimodal sensing data signals. Specifically: Because different sensors differ in output frequency and response time, the raw modal sensing signals need to undergo the following preprocessing to ensure data signal alignment and validity: Step S21: Filter and denoise the acquired multimodal sensing data signals: The Savitzky-Golay smoothing filter was used to remove noise from the ultraviolet absorption intensity data signal; Median filtering is used to remove spike noise from conductivity data signals; Kalman filtering is used to correct the drift of the pH data signal.
[0031] Step S22: Perform normalization and synchronous interpolation operations on the filtered and denoised multimodal sensor data signal: After filtering and denoising, all multimodal sensor data signals are normalized to the [0,1] interval using Min-Max, and then interpolated using a linear interpolation function to unify all multimodal sensor data signals to the same sampling time point (Δt=0.5s), ensuring data synchronization. The expression of the linear interpolation function is as follows: in, This is the new multimodal signal matrix obtained after linear interpolation, with time uniformity and a sampling interval of Δt. This represents the original multimodal signal matrix after filtering and denoising, but which is not yet aligned on the time axis. To ensure a uniform sampling interval, The time is set between 0.1 s and 2 s, and is preferably 0.5 s in this embodiment, so as to reduce data redundancy and improve model calculation efficiency while ensuring the real-time response capability of the system.
[0032] Step S23: Remove outliers from multimodal sensing data signals: Outliers in multimodal sensor data signals are identified using the 3σ (3-Sigma) criterion. Outliers exceeding three standard deviations (3σ) are removed, and the mean value within a local window centered on the outlier is used to replace it. This ensures that the processed signal is smooth and continuous at outlier points, without abrupt jumps.
[0033] Step S3: Use a Temporal Convolutional Network (TCN) to extract local temporal features from the preprocessed multimodal sensing data signal. Specifically: The Temporal Convolutional Network (TCN) in this embodiment employs causal convolution and dilated convolution structures, enabling it to capture local dependencies and multi-scale temporal patterns in signals. Residual connections prevent gradient vanishing during training, enhancing model stability. The output of the TCN is a high-dimensional feature vector for each time step, reflecting dynamic trends in the signal such as peak shape changes, conductance transitions, and pH shifts.
[0034] The expression for a temporal convolutional network is: in, The kernel size is 3 in this embodiment; These are the parameters of the l-th convolutional kernel; This is the time step output for the l-th layer.
[0035] The temporal convolutional network (TCN) in this embodiment uses residual connections to prevent gradient decay and achieves multi-scale feature extraction through causal convolution and dilation factor setting (dilation factor = 2). It can effectively identify peak shape changes, conductivity transitions and pH drift trends in the tomography curve.
[0036] Step S4: Input the temporal feature sequence output by the temporal convolutional network into the Transformer model to perform cross-modal feature fusion (Transformer multimodal feature fusion). Specifically: The Transformer model in this embodiment utilizes a multi-head self-attention mechanism to dynamically learn the global dependencies between three signals: UV absorption intensity, conductivity, and pH. This enables the model to identify potential dependency structures among multiple signals (such as pre-peak conductivity increase and post-peak pH drift).
[0037] The output of the Transformer model is a fused feature vector, which can identify cross-modal correlation patterns, such as "an increase in pre-peak conductivity indicates the appearance of the UV peak" and other cross-modal correlation patterns.
[0038] Input sequence of Transformer model Represented as: in, This represents the local temporal feature vector extracted by the temporal convolutional network at the T-th time step. The formula for calculating the attention mechanism in the Transformer model is: in, For query, As key, For value matrices, For feature dimensions; The expression for the fused feature vector F is: Step S5, Anomaly Detection and Adaptive Control: An anomaly detection model is constructed based on fused feature vectors to provide early warning of abnormal states. Specifically: The fused feature vector is input into a binary classification network, which employs fully connected layers (FC layers) and a softmax function. The binary classification network is used to determine the current tomography state, which includes normal and abnormal states. The network outputs the abnormality probability. ; If the abnormal probability , If the threshold is set (0.8 in this embodiment), an abnormal state is determined and an adaptive control mechanism is executed based on the abnormal probability. The pump speed is dynamically adjusted, and the formula for calculating the pump speed is as follows: ; in, The gain coefficient is used to adjust the pump speed's response to abnormal probabilities. Its value is set according to the type of chromatography column, flow rate range, and system stability requirements. It can be any real number between 0 and 1. In this embodiment, it is preferred to use Set it to 0.2 to 0.5 to achieve a balance between control response speed and system stability; The adjusted pump speed, This is the pump speed before adjustment.
[0039] When an abnormal state occurs, the system automatically reduces the pump speed or stops the pump to prevent mixing of the separated peaks or loss of the sample. At the same time, it records the trigger time, abnormal type and control action to facilitate subsequent analysis and model optimization.
[0040] Upon detecting an abnormal state, an alarm can be triggered and a visual display can be provided: 1. Displayed in real time on the host computer software interface: Multichannel signal curves (UV value, conductivity, pH value); Model prediction status (normal / abnormal); Current pump speed and control commands.
[0041] 2. When an abnormal state is detected, the system will automatically trigger an audible and visual alarm to remind the operator to intervene.
[0042] 3. All operation logs (including timestamps, exception types, and control responses) are automatically saved, supporting post-event traceability and model retraining.
[0043] Example 2 like Figure 2 As shown, this embodiment provides a system that applies a liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing, which includes: The multimodal sensing unit is used to acquire multimodal sensing data signals. It includes a UV detector, a conductivity meter, and a pH sensor installed at the outlet of the chromatography apparatus. This unit synchronously acquires multi-channel signals from the chromatography column output. The system sampling period is set to 0.5 s, acquiring one set of multimodal sensing data signals at each moment. The UV detector monitors the ultraviolet absorption intensity at a wavelength of 280 nm or 260 nm, the conductivity meter acquires the conductivity data signal of the liquid, and the pH sensor acquires the real-time pH value data signal of the liquid.
[0044] The data acquisition card is used to receive multimodal sensing data signals collected by the multimodal sensing unit and send the multimodal sensing data signals to the industrial-grade main control unit to ensure the timeliness and accuracy of signal acquisition.
[0045] The main control unit utilizes a temporal convolutional network to extract local temporal features and combines it with a Transformer model to achieve global dependency fusion of cross-modal features. Then, based on the fused features, an anomaly detection model is constructed for abnormal state early warning. Finally, an adaptive control algorithm based on anomaly probability is used to adjust the pump speed in real time. In this embodiment, the main control unit adopts an embedded ARM or x86 industrial computer.
[0046] The control execution unit is used to adjust the pump speed according to the adaptive control instructions of the main control unit. In this embodiment, the pump can be a peristaltic pump, which is the execution mechanism of the liquid chromatography system and is responsible for pushing the buffer solution (mobile phase) through the chromatography column at a precise and adjustable flow rate.
[0047] The alarm unit is used to issue an audible and visual alarm based on the abnormal status warning of the main control unit.
[0048] The specific embodiments described above further illustrate the technical problems, technical solutions, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. 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 liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing, characterized in that, It includes the following steps: Step S1: Acquire multimodal sensing data signals; Step S2: Preprocess the acquired multimodal sensing data signals; Step S3: Use a temporal convolutional network to extract local temporal features from the preprocessed multimodal sensing data signal; Step S4: Input the temporal feature sequence output by the temporal convolutional network into the Transformer model to perform cross-modal feature fusion (Transformer multimodal feature fusion). Step S5: Construct an anomaly detection model based on the fused feature vector to provide early warning of abnormal states, and dynamically adjust the pump speed according to the anomaly probability.
2. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 1, characterized in that, The multimodal sensing data signals include ultraviolet absorption intensity data signals, conductivity data signals, and pH value data signals; The expression for the multimodal sensing data signal is as follows: in, The absorbance signal intensity of the component that produces ultraviolet absorption at a specified wavelength in the eluent flowing out from the outlet of the chromatography column; The conductivity data signal of the eluent at the outlet of the chromatography column; This is the real-time pH value signal of the eluent at the outlet of the chromatography column.
3. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 1, characterized in that, In step S2, the acquired multimodal sensing data signals are preprocessed, specifically including the following steps: Step S21: Filter and denoise the acquired multimodal sensing data signals; Step S22: Normalize and synchronously interpolate the filtered and denoised multimodal sensor data signal; Step S23: Remove outliers from the multimodal sensing data signals.
4. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 3, characterized in that, In step S22, the multimodal sensing data signal after filtering and denoising is normalized and synchronously interpolated, specifically including the following steps: All multimodal sensor data signals after filtering and denoising are normalized to the [0,1] interval using Min-Max, and then interpolated using a linear interpolation function to unify all multimodal sensor data signals to the same sampling time point. The expression of the linear interpolation function is as follows: in, It is a multimode signal matrix. This represents the original multimodal signal matrix after filtering and denoising, but which is not yet aligned on the time axis. For a uniform sampling interval.
5. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 3, characterized in that, In step S23, removing outliers from the multimodal sensing data signal specifically includes the following steps: The 3σ criterion is used to identify outliers in multimodal sensor data signals. Outliers exceeding three standard deviations are removed, and the mean within a local window centered on the outlier is used to replace the outlier.
6. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 1, characterized in that, The expression for the temporal convolutional network in step S3 is: in, The kernel size is [size]. These are the parameters of the l-th convolutional kernel; This is the time step output for the l-th layer.
7. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 1, characterized in that, The input sequence of the Transformer model in step S4 Represented as: in, This represents the local temporal feature vector extracted by the temporal convolutional network at the T-th time step. The calculation formula for the attention mechanism of the Transformer model is as follows: in, For query, As key, For value matrices, For feature dimensions; The expression for the fused feature vector F is: 。 8. The liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing according to claim 3, characterized in that, Step S5 involves constructing an anomaly detection model based on fused feature vectors to provide early warning of abnormal states and dynamically adjusting the pump speed according to the anomaly probability. Specifically, this includes the following steps: The fused feature vector is input into a binary classification network, which employs fully connected layers and a softmax function. This network determines the current tomography state, which includes normal and abnormal states. The binary classification network outputs the abnormality probability. ; If the abnormal probability , If the threshold is reached, an abnormal state is determined and an adaptive control mechanism is executed, based on the abnormal probability. The pump speed is dynamically adjusted, and the formula for calculating the pump speed is as follows: ； in, To control the gain coefficient, The adjusted pump speed, The pump speed before adjustment.
9. A system employing the liquid chromatography method based on the fusion of artificial intelligence and multimodal sensing as described in any one of claims 1 to 4, characterized in that, It includes: A multimodal sensing unit, wherein the multimodal sensing unit is used to acquire multimodal sensing data signals; A data acquisition card is used to receive multimodal sensing data signals acquired by the multimodal sensing unit and send the multimodal sensing data signals to the industrial-grade main control unit. The main control unit uses a temporal convolutional network to extract local temporal features and combines a Transformer model to achieve global dependency fusion of cross-modal features. Then, an anomaly detection model is built based on the fused features to provide early warning of abnormal states. Finally, the pump speed is adjusted in real time through an adaptive control algorithm based on anomaly probability. A control execution unit, which is used to adjust the pump speed according to the adaptive control command of the main control unit; An alarm unit, specifically an audible and visual alarm device used to issue an alarm based on an abnormal state warning from the main control unit.
10. The system according to claim 9, characterized in that: The multimodal sensing unit includes a UV detector, a conductivity meter, and a pH sensor.