An artificial intelligence-based automatic calibration method and system for a gas chromatograph

Abstract

The application belongs to the technical field of artificial intelligence, and discloses a gas chromatograph automatic calibration method and system based on artificial intelligence. The system constructs a general hardware interface protocol through a CNN-GRU fusion model, can automatically identify the hardware configuration and communication protocol of multiple brands and multiple models of instruments, realizes plug and play of the instruments, completes standard gas delivery regulation through intelligent flow control and standard gas intelligent switching technology, avoids operation errors caused by manual gas preparation and flow regulation, a multi-modal AI calibration module fuses multi-source data deep analysis, automatically generates and iterates optimal calibration parameters online, and improves calibration control precision; an environment intelligent compensation module collects laboratory multi-dimensional environmental parameters in real time based on a Transformer model, accurately analyzes the correlation between environmental factors and calibration results, dynamically calculates compensation coefficients and links the calibration module to adjust parameters, and eliminates the influence of environmental interference on calibration precision.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically an automatic calibration method and system for gas chromatographs based on artificial intelligence. Background Technology

[0002] Gas chromatographs are core analytical instruments in fields such as environmental monitoring, petrochemicals, and pharmaceutical research and development. The accuracy of their detection results depends on calibration precision, and calibration efficiency affects the timeliness of detection. Currently, gas chromatograph calibration technology mainly relies on manual operation, which presents the following technical problems: The calibration process is highly dependent on the operator's experience. Differences in operation between different personnel can easily lead to poor reproducibility of calibration results. In addition, there are concentration errors in the manual preparation of standard gases, and flow deviations can easily occur when adjusting the carrier gas flow rate, making it difficult to achieve precise control of calibration parameters. The calibration process is mostly a static offline operation, which cannot respond to dynamic changes in the operation of the instrument in real time. Faced with baseline drift caused by factors such as fluctuations in carrier gas flow, column aging, and sudden changes in environmental temperature and humidity, it cannot dynamically compensate for the effects of environmental and component fluctuations in real time. It can only be manually adjusted afterward, resulting in a continuous decline in long-term detection accuracy. Automated calibration systems are mostly single-function modules, lacking health management throughout the instrument's lifecycle. They can only perform basic parameter calibration and cannot predict potential problems such as column decay and core component failure. Furthermore, their calibration strategies are fixed and difficult to adapt to the personalized needs of different sample types and detection scenarios. The AI ​​algorithms used in the relevant calibration systems are mostly general-purpose fusion models, which are not specifically optimized for the detection characteristics and operating rules of the core components of gas chromatographs, and do not achieve deep coupling between the algorithm and the entire instrument calibration process. Summary of the Invention

[0003] The purpose of this invention is to provide an automatic calibration method and system for gas chromatographs based on artificial intelligence, so as to solve one or more problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an automatic calibration system for a gas chromatograph based on artificial intelligence, comprising: Preferably, the intelligent hardware interaction module is based on an adaptive recognition model that integrates convolutional neural networks and gated recurrent units, and constructs a generalized hardware interface protocol based on the CNN-GRU fusion model. It is compatible with the carrier gas control port, injection interface, and detector signal port of mainstream gas chromatographs. Through the model, it can autonomously learn the hardware signal characteristics of different instruments and automatically identify the instrument model, hardware configuration, and communication protocol. Simultaneously, it integrates a mass flow controller and an inert gas delivery pipeline, and combines a flow prediction and control model based on reinforcement learning to collect flow and pressure feedback data in real time. Through reinforcement learning algorithms, it dynamically adjusts the delivery speed and pressure of the standard gas to achieve adaptive optimization of flow control. It is equipped with a multi-channel standard gas storage unit and combined with a gas type recognition model based on neural networks to automatically identify the standard gas components required for calibration, realize intelligent switching of various standard gases for calibration, and achieve plug-and-play adaptation and standard gas delivery control for gas chromatographs of different brands and models.

[0005] Preferably, the instrument digital twin module is based on a fusion model of attention mechanism bidirectional long short-term memory network and graph neural network to construct a dedicated digital twin model for gas chromatograph. It collects the operating parameters and calibration process data of gas chromatograph transmitted by the intelligent hardware interaction module in real time, captures the temporal variation characteristics of parameters through Bi-LSTM model, and constructs the correlation map of various components of the instrument in combination with GNN model to automatically identify the core characteristic parameters of the instrument. A virtual calibration scenario based on deep learning is constructed to simulate the instrument's operating behavior under different testing conditions. The virtual simulation is used to predict possible parameter anomalies during the calibration process, and the initial values ​​of calibration parameters are optimized based on the simulation results. A real-time data interaction mechanism between the virtual model and the physical instrument is established. Data synchronization mapping is achieved based on neural networks, and the operating status of the physical instrument is simulated and predicted in real time through the virtual model.

[0006] Preferably, the multimodal AI calibration module is based on a fusion model of multi-scale convolutional neural network, long short-term memory network and deep reinforcement learning. It collects standard gas concentration data and instrument hardware parameters provided by the intelligent hardware interaction module, virtual simulation data provided by the instrument digital twin module and environmental parameters transmitted by the environmental intelligent compensation module, forming four types of multi-source heterogeneous data: standard gas concentration data, instrument hardware operating parameters, virtual simulation data and environmental parameters. MS-CNN extracts signal features at different scales, LSTM captures the temporal variation features of the signal, and the DRL model performs weight allocation and optimization on the fused features to achieve multi-dimensional data fusion analysis. The AI ​​model trained based on the fused features automatically generates the optimal combination of calibration parameters. At the same time, the DRL model realizes online iterative optimization of calibration parameters to adapt to the needs of different detection scenarios. During the calibration process, the model monitors the deviation of calibration data in real time. If an anomaly occurs, the MS-CNN-LSTM-DRL fusion model identifies the cause of the anomaly and dynamically adjusts the calibration parameters. The online calibration parameter iteration adopts a dual mode of regular timed iteration and emergency deviation iteration. Under normal conditions, the parameters are optimized by iteration at a fixed time period. When the calibration deviation exceeds the preset allowable range, an emergency iteration is triggered immediately to quickly correct the calibration parameters, thereby ensuring that the calibration accuracy is stable and meets the standard.

[0007] By employing model compression technology, the trained AI model is deployed on the embedded control terminal of the gas chromatograph, and the calibration results are simultaneously fed back to the instrument health management module and the calibration knowledge graph module.

[0008] Preferably, the environmental intelligent compensation module is based on the Transformer attention mechanism neural network model, combined with the environmental sensitivity characteristics of gas chromatograph calibration, to collect five types of parameters of the laboratory environment in real time: temperature, humidity, air pressure, vibration, and electromagnetic interference, to construct a time series dataset of environmental parameters, and to capture the dynamic change pattern of environmental parameters through the Transformer model to identify key parameters; Key environmental parameters are determined based on their weight influencing the calibration results. Environmental parameters with a weight higher than the system's set threshold are marked as key parameters. The system will increase the priority of key parameter acquisition and calculation, focus on monitoring their fluctuations, and prioritize compensation calculations.

[0009] The correlation weight between environmental parameters and calibration results output by the multimodal AI calibration module is analyzed by using the Transformer attention mechanism. The weight allocation is dynamically adjusted in combination with the operating status of the gas chromatograph, and the corresponding environmental compensation coefficient is calculated. Simultaneously, the compensation coefficient is synchronized in real time to the instrument's digital twin module and multimodal AI calibration module. The calibration parameters are adjusted in conjunction with the AI ​​model to achieve coordinated optimization of environmental compensation and the calibration process.

[0010] Preferably, the instrument health management module is based on a time-series graph neural network and an adversarial training model. It collects abnormal data during the instrument operation process transmitted by the intelligent hardware interaction module, the multimodal AI calibration module, and the instrument digital twin module. It constructs an instrument health status correlation map through the TGNN model, and combines the adversarial training model to remove abnormal data interference and assess the current health level of the instrument. The collected abnormal data undergoes two-stage filtering: threshold screening and continuity verification, eliminating invalid data caused by random electromagnetic interference and instantaneous signal jumps. The data is then input into the model for analysis and calculation, ensuring the authenticity and validity of the data used in health assessment and life prediction. The health assessment adopts a multi-source weighted fusion method, integrating hardware operation data, calibration deviation data, and virtual simulation data. Weights are assigned according to the importance of each data point, and the final health score is obtained after weighted calculation, avoiding the distortion of assessment results due to the bias of a single data source.

[0011] Based on the LSTM time series model and reinforcement learning algorithm, the system analyzes the operating data of core components, including the chromatographic column, flow controller, and detector. Combined with historical fault data and maintenance records, it predicts the remaining service life of the components and outputs the service life decay trend curve. Based on the health status and service life prediction results, it automatically generates dedicated maintenance suggestions. Simultaneously, it links with the multimodal AI calibration module to dynamically adjust calibration parameters and calibration cycles, thereby achieving health status assessment of the gas chromatograph, lifespan prediction of core components, and linkage optimization of calibration strategies with health status. Health assessment results and maintenance suggestions are synchronized to the intelligent scheduling and management module.

[0012] Preferably, the calibration knowledge graph module is based on graph neural network and knowledge graph technology, combined with deep learning knowledge mining algorithm, to collect and organize industry standards for gas chromatograph calibration, historical calibration data of multimodal AI calibration module, calibration specifications of different instrument models, and fault handling cases fed back by instrument health management module. Combined with knowledge extraction model, it automatically extracts multi-dimensional related information to construct dynamic calibration knowledge graph. By using GNN models and reinforcement learning algorithms, calibration patterns can be automatically extracted from historical calibration data and fault cases, and the optimal calibration strategy can be inferred in different scenarios. At the same time, conflicting information in the knowledge graph can be identified and automatically corrected. The optimal calibration strategy is output in a complete form, including calibration parameter combinations, standard operating procedures, and environmental requirements. It is directly synchronized to the multimodal AI calibration module as the initial reference for generating calibration parameters, thereby improving the efficiency and adaptability of the calibration strategy. The system automatically matches the corresponding scene tags in the knowledge graph based on the current instrument model, sample type, and laboratory environment conditions, and retrieves the most suitable optimal calibration strategy. If there is no perfectly matching scene, it automatically selects a similar scene strategy and performs adaptive fine-tuning.

[0013] The inference-generated calibration strategy is synchronized to the multimodal AI calibration module, while the deep learning model learns new calibration data and fault cases in real time, continuously updating the knowledge graph.

[0014] Preferably, the intelligent scheduling and management module is based on deep reinforcement learning and a multi-agent collaborative model. It combines the instrument usage frequency, detection task priority, instrument health status, and remaining standard gas quantity with the calibration knowledge provided by the calibration knowledge graph module. It generates the optimal calibration plan through the DRL model and automatically plans the calibration time, calibration items, and calibration cycle. Based on a multi-agent cooperative model, the working sequence of each module in the system is coordinated. A multi-objective optimization algorithm is adopted to optimize the allocation of hardware resources, computing resources and standard gas resources in the calibration process, and to prioritize the calibration needs of high-priority detection tasks, thereby realizing intelligent scheduling and resource optimization of calibration tasks.

[0015] When conflicts occur in standard gas, computing power, and hardware interface resources, the system sorts and processes them according to the urgency of the detection task, the health status of the instrument, and the timeliness of the calibration. Low-priority non-critical calibration tasks are suspended, and resources are allocated to high-priority critical calibration needs. After the conflict is resolved, the suspended tasks are automatically resumed to ensure the continuous and stable scheduling process.

[0016] This invention also provides an automatic calibration method for a gas chromatograph based on artificial intelligence, which, based on the above-described system, includes the following specific steps: First, hardware adaptation and standard gas preparation are carried out. A generalized hardware interface protocol is built through an adaptive recognition model that integrates CNN and GRU to automatically identify the gas chromatograph model, hardware configuration and communication protocol. At the same time, the flow rate and pressure of the standard gas are dynamically adjusted through the flow control component to complete the intelligent switching of multi-channel standard gases and instrument adaptation. Subsequently, multi-source data was collected and a virtual simulation scenario was constructed. Instrument hardware operating parameters and calibration data were collected in real time. A digital twin model was constructed through the fusion model of Bi-LSTM and GNN to simulate the instrument operating behavior under different detection conditions, predict calibration parameter anomalies and optimize initial values, and establish a virtual-real data synchronization mapping mechanism. Perform multimodal AI calibration and environmental compensation, collect various types of data from hardware, virtual simulation and environment, extract signal features and deeply fuse them through MS-CNN, LSTM and DRL fusion models, automatically generate optimal calibration parameters and realize online iteration, monitor calibration deviation in real time and adjust dynamically, and collect environmental parameters based on Transformer model and calculate compensation coefficients to adjust calibration parameters in a coordinated manner. Finally, health linkage and scheduling optimization are carried out. The health level of the instrument is evaluated through TGNN and adversarial training model, the lifespan of core components is predicted and maintenance suggestions are generated, and the parameters and cycle are optimized by linking the multimodal AI calibration module. Based on knowledge graph, calibration rules are extracted and strategy support is provided. The work of each module is coordinated through DRL and multi-agent model to generate the optimal calibration plan and optimize resource allocation.

[0017] The beneficial effects of this invention are as follows: 1. The system of this invention constructs a universal hardware interface protocol through a CNN-GRU fusion model, which can automatically identify the hardware configuration and communication protocol of instruments of multiple brands and models, enabling plug-and-play operation of instruments; through intelligent flow control and intelligent standard gas switching technology, it completes the standard gas delivery regulation, avoiding operational errors caused by manual gas mixing and flow adjustment; the multimodal AI calibration module integrates multi-source data for in-depth analysis, automatically generates and iterates the optimal calibration parameters online, improving calibration control accuracy and ensuring stable and reproducible calibration results.

[0018] 2. The intelligent environmental compensation module of this invention collects multi-dimensional environmental parameters of the laboratory in real time based on the Transformer model, accurately analyzes the correlation between environmental factors and calibration results, dynamically calculates the compensation coefficient and links the calibration module to adjust parameters, and eliminates the impact of environmental interference on calibration accuracy. The instrument digital twin module constructs a dedicated virtual calibration scenario through a fusion model, simulates the full-condition operation of the instrument, predicts abnormal calibration parameters and optimizes initial values, and establishes a real-time mapping mechanism between virtual and real data to ensure the long-term stability of the instrument's detection accuracy.

[0019] 3. The instrument health management module of this invention assesses the health level of the instrument through a dedicated model, predicts the remaining lifespan of core components, and generates dedicated maintenance suggestions, dynamically adjusting calibration parameters and cycles; the calibration knowledge graph integrates industry standards, historical data, and fault cases, continuously reasoning to adapt the optimal calibration strategy to different scenarios; the intelligent scheduling module optimizes resource allocation based on a multi-agent collaborative model, generates dedicated calibration plans based on task priorities, extends the instrument's lifespan, and reduces operation and maintenance costs. Attached Figure Description

[0020] Figure 1 This is a flowchart of the automatic calibration process for a gas chromatograph based on artificial intelligence, as described in this invention. Figure 2 This is a flowchart of the multimodal AI calibration parameter generation and iterative optimization process of this invention; Figure 3 This is a flowchart of the dynamic calibration process for intelligent environmental compensation and health linkage in this invention. Detailed Implementation

[0021] 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.

[0022] like Figures 1 to 3 As shown, this embodiment of the invention provides an automatic calibration system for a gas chromatograph based on artificial intelligence, comprising: In this embodiment of the invention, the intelligent hardware interaction module is based on an adaptive recognition model that integrates a convolutional neural network (CNN) and a gated recurrent unit (GRU). It constructs a generalized hardware interface protocol based on this CNN-GRU fusion model, compatible with the carrier gas control port, injection interface, and detector signal port of mainstream gas chromatographs. The model autonomously learns the hardware signal characteristics of different instruments, automatically identifying the instrument model, hardware configuration, and communication protocol without manual debugging. If the initial automatic recognition fails, the system automatically switches to a full command scan mode, iteratively matching and verifying the interface commands of mainstream instruments one by one. If recognition still fails after multiple attempts, it outputs an interface error message and fully records the fault type and troubleshooting information, ensuring a complete and traceable hardware adaptation process and rapid fault location.

[0023] The CNN-GRU fusion model consists of two convolutional layers, one max-pooling layer, and two fully connected GRU hidden layers. The convolutional layers extract spatial features of hardware signals, while the GRU layers capture temporal features of signals. The model inputs are instrument interface voltage and communication message sequences, and the outputs are instrument model and protocol label. Training uses the Adam optimizer with a learning rate of 1e-4, a batch size of 32, 50 training epochs, and cross-entropy loss as the loss function.

[0024] Simultaneously, it integrates a mass flow controller and an inert gas delivery pipeline, and combines a flow prediction and control model based on reinforcement learning to collect flow and pressure feedback data in real time. Through reinforcement learning algorithms, it dynamically adjusts the delivery speed and pressure of the standard gas to achieve adaptive optimization of flow control and ensure that the standard gas concentration error is controlled within ±0.08%. The flow control adopts a real-time closed-loop feedback adjustment mechanism, continuously collecting real-time pressure and flow signals of the delivery pipeline, quickly comparing them with preset standard values, and dynamically correcting the output control parameters. The inner wall of the inert delivery pipeline undergoes professional passivation treatment to effectively avoid the adsorption and residue of standard gas. The pipeline end integrates a real-time airtightness detection unit to monitor the pipeline sealing status throughout the process. Once a leakage risk is detected, an alarm is immediately triggered and delivery is suspended, ensuring the accuracy of concentration control from a hardware perspective.

[0025] The mass flow controller adopts a millisecond-level high-speed closed-loop control cycle, collects and responds to flow feedback signals in real time, and quickly fine-tunes the control output, effectively suppressing pipeline pressure fluctuations and flow drift problems, and ensuring continuous stability of flow rate and pressure throughout the standard gas delivery process.

[0026] It is equipped with a multi-channel standard gas storage unit and combined with a gas type recognition model based on neural networks to automatically identify the standard gas components required for calibration, realize intelligent switching of various standard gases for calibration, adapt to the calibration needs of samples with different components and concentration ranges, improve the module's adaptability flexibility, and realize plug-and-play adaptation of gas chromatographs of different brands and models as well as standard gas delivery control.

[0027] The multi-channel standard gas switching uses a high-precision electromagnetic control valve group to work together. The system automatically matches the corresponding standard gas channel according to the current calibration type. During the switching process, the residual gas in the pipeline is first emptied by negative pressure, and then the target standard gas is introduced at a constant speed. After the concentration is stable for 30 seconds, the calibration process is officially started to avoid cross-interference between different gases.

[0028] The universal hardware interface protocol incorporates communication command parsing libraries for mainstream brand gas chromatographs. Through adaptive level matching and signal conversion circuits, it adapts to the electrical characteristics of different instrument interfaces, automatically completing instrument handshake communication and core parameter matching. The entire process requires no manual configuration of interface type, communication rate, or other parameters, enabling rapid and stable hardware link connection and data exchange.

[0029] In this embodiment of the invention, the instrument digital twin module is based on a fusion model of Bidirectional Long Short-Term Memory Network (Bi-LSTM) and Graph Neural Network (GNN) to construct a dedicated digital twin model for gas chromatographs. It collects the gas chromatograph hardware operating parameters and calibration process data transmitted by the intelligent hardware interaction module in real time. The gas chromatograph hardware operating parameters include carrier gas flow rate, injection port temperature, column temperature, detector response value, etc. The Bi-LSTM model captures the temporal variation characteristics of the parameters, and the GNN model is combined to construct the correlation map of various instrument components (column, detector, flow controller) to automatically identify the core characteristic parameters of the instrument. The Bi-LSTM+GNN fusion model consists of a bidirectional LSTM layer, two graph convolutional layers (the number of nodes corresponds to the number of instrument components), and a fully connected layer connected in series. The input to the Bi-LSTM is the sequence of instrument time-series parameters, including carrier gas flow rate, temperature, and response value, with a time step of 60. The input to the GNN is the adjacency matrix of the instrument components, and the model output is the component state vector and the predicted parameter values. Training uses the SGD optimizer with a learning rate of 5e-4, a loss function of mean squared error, and 80 iterations.

[0030] The component correlation diagram is strictly constructed based on the physical connection relationship of the instrument and the logic of parameter coupling influence, clearly defining the parameter transmission path and mutual influence weight between the chromatographic column, flow controller and detector.

[0031] A virtual calibration scenario based on deep learning is constructed to simulate the instrument's operating behavior under testing conditions such as sample composition, environmental parameter fluctuations, and component aging. The virtual simulation can predict possible parameter anomalies during the calibration process, optimize the initial values ​​of calibration parameters based on the simulation results, reduce the number of trial and error steps in the actual calibration process, and reduce the consumption of standard gas. Virtual simulation comprehensively covers all aspects of operation, including carrier gas flow rate control, injection port and column temperature control, and detector signal acquisition. When the parameters calculated by simulation deviate from the normal operating range and exceed the set threshold, the system immediately determines that the calibration parameters are abnormal and issues a pre-judgment prompt, while simultaneously providing targeted parameter correction directions, effectively improving the success rate of physical calibration.

[0032] The virtual simulation parameter initialization is based on the current operating parameters of the physical instrument and preset using historical best calibration data. It is also compatible with the parameter requirements of different testing scenarios. After initialization, the simulation is automatically verified to ensure that the initial parameters meet the safe operating range of the instrument.

[0033] Establish a real-time data interaction mechanism between the virtual model and the physical instrument, realize data synchronization mapping based on neural networks, and perform real-time simulation and prediction of the operating status of the physical instrument through the virtual model to achieve coordinated linkage between virtual calibration and physical calibration.

[0034] The virtual-real data synchronization adopts a combination of timed triggering and event triggering modes. The timed synchronization period is fixed at 100 milliseconds to ensure the real-time performance of data under normal conditions. When the instrument parameters change suddenly, event synchronization is triggered immediately. All synchronized data is accompanied by a unique verification code. After verification, it is written into the virtual model to ensure that the virtual calibration scenario is completely consistent with the operating state of the physical instrument and there is no delay or deviation.

[0035] The key operating parameters of the virtual model and the physical instrument are accurately mapped one-to-one. Changes in parameters such as carrier gas flow rate, temperature control parameters, and detector response values ​​of the physical instrument are synchronously mapped to the corresponding parameters in the virtual model. The simulation predictions and optimization results of the virtual model are also synchronously mapped to the parameters to be adjusted in the physical instrument, ensuring complete consistency between the virtual and physical states.

[0036] In this embodiment of the invention, the multimodal AI calibration module is based on a fusion model of multi-scale convolutional neural network (MS-CNN), long short-term memory network (LSTM) and deep reinforcement learning (DRL). It collects standard gas concentration data and instrument hardware parameters provided by the intelligent hardware interaction module, virtual simulation data provided by the instrument digital twin module, and environmental parameters transmitted by the environmental intelligent compensation module, forming four types of multi-source heterogeneous data: standard gas concentration data, instrument hardware operating parameters, virtual simulation data, and environmental parameters. The four types of multi-source heterogeneous data are first normalized and standardized preprocessed to unify the data dimensions, and then feature-level fusion is performed according to parameter correlation and importance to integrate the effective features from different sources into a unified calibration analysis dataset.

[0037] By extracting signal features at different scales, such as baseline features, peak shape features, and noise features, using MS-CNN, capturing the temporal variation features of the signal using LSTM, and weighting and optimizing the fused features using the DRL model, multi-dimensional data fusion analysis is achieved. The AI ​​model trained based on the fused features automatically generates the optimal combination of calibration parameters, including carrier gas flow calibration values ​​and detector gain parameters. At the same time, the DRL model enables online iterative optimization of calibration parameters to adapt to the needs of different detection scenarios. During the calibration process, the model monitors the deviation of the calibration data in real time. If baseline fluctuations, peak shape anomalies, or other abnormalities occur, the MS-CNN-LSTM-DRL fusion model identifies the cause of the anomaly and dynamically adjusts the calibration parameters. The MS-CNN uses 3-scale convolutional kernels to extract multi-scale peak shape and baseline features, LSTM uses 2 hidden layers to capture temporal drift, DRL's state space includes calibration bias, environmental parameters, and component states, the action space includes calibration parameter adjustments, and the reward function is the inverse of the calibration error plus the stability score. The model input consists of 4 types of multi-source data, and the output is the optimal combination of calibration parameters. Training uses the PPO algorithm with a learning rate of 1e-5, an exploration rate of 0.1, and a training step size of 10000.

[0038] The calibration deviation is determined based on the peak area error of the standard sample and the retention time deviation of the target component. If it exceeds the allowable range, it is judged as a calibration deviation. The model can accurately identify four common abnormal causes: baseline fluctuation, abnormal carrier gas flow, environmental interference, and aging of core components, and output corresponding calibration parameter adjustment schemes.

[0039] By employing model compression technology (knowledge distillation and quantization), the trained AI model is deployed on the embedded control terminal of the gas chromatograph, ensuring that the response time for adjusting calibration parameters is controlled within 80ms, while reducing the computational power consumption of model operation. The calibration results are simultaneously fed back to the instrument health management module and the calibration knowledge graph module.

[0040] The model compression uses knowledge distillation technology to retain core calibration features, while quantizing the model into 8-bit fixed-point arithmetic. After compression, the model size is reduced by 70%, and it is lightweight and deployed on the instrument's embedded ARM control terminal. It receives multiple calibration data in real time through a high-speed serial port, and the output optimal calibration parameters are directly written into the instrument control register, achieving millisecond-level parameter adjustment response. Before writing calibration parameters, a validity check is performed to ensure that the parameters are within the safe operating range of the instrument. After the check is passed, the parameters are written to the control register according to the specified protocol. After the writing is completed, the parameter validity check is automatically performed. Only after the check is passed will the new calibration parameters be officially used, ensuring that parameter adjustment is safe and reliable.

[0041] In this embodiment of the invention, the environmental intelligent compensation module is based on the Transformer attention mechanism neural network model. Combined with the environmental sensitivity characteristics of gas chromatograph calibration, it collects five types of parameters of the laboratory environment in real time: temperature, humidity, air pressure, vibration, and electromagnetic interference. It constructs a time series dataset of environmental parameters and captures the dynamic change pattern of environmental parameters through the Transformer model to identify key parameters that have a significant impact on the calibration results. The Transformer employs a 6-layer encoder and an 8-head self-attention mechanism. The input is a time-series sequence of 5 types of environmental parameters, and the output is the associated weights and compensation coefficients of the environmental parameters. The position encoding uses sinusoidal encoding, layer normalization, and a feedforward network. The training uses the AdamW optimizer with a learning rate of 2e-4, weight decay of 1e-6, batch size of 16, and 60 rounds.

[0042] Environmental sensors are evenly deployed in the instrument's internal cavity, around the carrier gas pipeline, and in the core testing area of ​​the laboratory. They adopt a distributed multi-point acquisition mode to upload raw environmental data to the processing unit in real time, and the system automatically removes single-point abnormal sampling values.

[0043] The correlation weight between environmental parameters and the calibration results output by the multimodal AI calibration module is analyzed by the Transformer attention mechanism. The weight allocation is dynamically adjusted in combination with the operating status of the gas chromatograph (such as column temperature and carrier gas flow rate), and the corresponding environmental compensation coefficient is calculated to eliminate problems such as baseline drift caused by temperature and humidity, flow deviation caused by gas pressure fluctuation, and signal noise caused by electromagnetic interference. The environmental compensation coefficient is transmitted in real time to the instrument's digital twin module and multimodal AI calibration module via the instrument's internal data bus. After receiving the compensation coefficient, the calibration module corrects the calibration parameters according to the weighting of environmental factors. The digital twin module updates the virtual environment parameters synchronously to ensure that the environmental compensation of the virtual simulation and the physical calibration are completely consistent. The compensation coefficient is set with upper and lower limits to prevent the compensation coefficient from being too large or too small due to extreme fluctuations in environmental parameters. When the limit is exceeded, the system automatically locks the safety threshold and resumes normal compensation calculation after the environmental state stabilizes, preventing over-compensation from affecting the calibration results.

[0044] The compensation coefficient is dynamically updated in real time to keep up with changes in environmental parameters. The update cycle is completely synchronized with the environmental data acquisition cycle, ensuring that the compensation coefficient can be adjusted in real time to keep up with fluctuations in the laboratory environment.

[0045] Simultaneously, the compensation coefficient is synchronized in real time to the instrument's digital twin module and multimodal AI calibration module. Through the linkage of the AI ​​model, the calibration parameters are adjusted to achieve synergistic optimization of environmental compensation and calibration process, ensuring that the calibration results are not affected by environmental factors and improving the accuracy and stability of the calibration results.

[0046] In this embodiment of the invention, the instrument health management module, based on a temporal graphical neural network (TGNN) and an adversarial training model, collects abnormal data during instrument operation transmitted by the intelligent hardware interaction module, the multimodal AI calibration module, and the instrument digital twin module. The abnormal data includes overshoot of the temperature control system, detector noise value, column efficiency decay rate, etc. The TGNN model is used to construct an instrument health status correlation map, and the adversarial training model is used to eliminate interference from abnormal data and assess the current health level of the instrument. The health level includes normal, warning, and fault. The TGNN consists of three temporal graph convolutional layers. The node features are the component operation parameters, and the edges are the component relationships. The adversarial training consists of a generator (two-layer GRU) and a discriminator (two-layer fully connected layers). The generator takes normal operating data as input, and the discriminator distinguishes between real and abnormal data. The model takes abnormal data sequences as input and outputs health level and lifespan prediction values. The training uses Wasserstein loss, with 70 iterations and a gradient penalty coefficient of 10.

[0047] Health levels are classified according to the deviation range of operating parameters of core components. A parameter deviation within 5% is considered a normal state, 5%-15% is considered a warning state, and more than 15% is considered a fault state. At the same time, the component life decay rate is also considered for comprehensive judgment to avoid misjudgment caused by fluctuation of a single parameter and improve the accuracy of health assessment.

[0048] Based on LSTM time-series models and reinforcement learning algorithms, the system analyzes the operational data of core components such as chromatographic columns, flow controllers, and detectors. Combining historical fault data and maintenance records, it predicts the remaining lifespan of these components and outputs lifespan decay trend curves. When the remaining lifespan of a core component falls below the system's set safety threshold, the system automatically triggers a tiered lifespan warning, simultaneously pushing warning information to the control terminal and dynamically adjusting the calibration cycle. This reminds operators to prepare for component replacement in advance, preventing sudden instrument malfunctions from affecting normal testing operations. Based on the health status and lifespan prediction results, the system automatically generates dedicated maintenance recommendations, including component cleaning, calibration cycle adjustment, and replacement warnings. The remaining lifespan of core components is calculated by combining the cumulative operating time of the components, the rate of parameter decay, and the probability of historical failures. Maintenance recommendations are automatically generated according to the instrument's health level. In normal condition, regular routine checks are suggested; in early warning condition, component cleaning and maintenance are suggested; and in fault condition, immediate shutdown and component replacement are suggested.

[0049] Simultaneously, it links with the multimodal AI calibration module to dynamically adjust calibration parameters and calibration cycles. For example, it optimizes carrier gas flow and temperature control parameters to address column decay, thereby extending instrument lifespan, reducing maintenance costs, and enabling the coordinated optimization of gas chromatograph health status assessment, core component lifespan prediction, and calibration strategies with health status. Health assessment results and maintenance recommendations are synchronized to the intelligent scheduling and management module.

[0050] In this embodiment of the invention, the calibration knowledge graph module is based on graph neural network and knowledge graph technology, combined with deep learning knowledge mining algorithm, to collect and organize industry standards for gas chromatograph calibration, historical calibration data of multimodal AI calibration module, calibration specifications of different instrument models, and fault handling cases fed back by instrument health management module. Combined with knowledge extraction model, it automatically extracts multi-dimensional correlation information including instrument model, calibration parameters, environmental conditions, calibration results, and fault type to construct a dynamic calibration knowledge graph. The knowledge graph contains five categories of entities: instrument model, calibration parameters, environmental conditions, fault type, and industry standards. It adopts an incremental real-time update mode. Every time a calibration operation or fault handling is completed, the system automatically synchronizes the effective data to the graph, continuously enriching the knowledge content and ensuring the real-time nature of the graph information.

[0051] By using GNN models and reinforcement learning algorithms, calibration patterns can be automatically extracted from historical calibration data and failure cases, and the optimal calibration strategy can be inferred for different scenarios. For example, there are dedicated calibration schemes for high-boiling-point samples, complex component samples, and high-concentration samples. At the same time, it can identify conflicting information in the knowledge graph and automatically correct it, thereby improving the accuracy of inference. The conflict correction adopts a historical data voting and screening mechanism, comparing multiple sets of similar calibration data and failure cases, retaining high-confidence valid information, eliminating abnormal conflict data, and automatically updating the entity association relationship of the knowledge graph after the correction is completed, so as to continuously ensure the accuracy of the knowledge graph reasoning calibration strategy.

[0052] The inference-generated calibration strategy is synchronized to the multimodal AI calibration module, while the deep learning model learns new calibration data and fault cases in real time, continuously updating the knowledge graph.

[0053] The knowledge graph GNN consists of two graph attention layers. Entities include instrument model, parameters, faults, and standards, while relations include calibration associations and fault causality. The input consists of triplet data of head entity, relation, and tail entity, and the output is the optimal calibration strategy vector. Training uses the TransE algorithm with a learning rate of 1e-3, a negative sampling number of 5, a number of rounds of 100, and an automatic conflict correction threshold of 0.8.

[0054] In this embodiment of the invention, the intelligent scheduling and management module is based on deep reinforcement learning and a multi-agent collaborative model. According to the usage frequency of the gas chromatograph, the priority of the detection tasks, the instrument health status fed back by the instrument health management module, and the remaining amount of standard gas fed back by the intelligent hardware interaction module, combined with the calibration knowledge provided by the calibration knowledge graph module, the module generates an optimal calibration plan through the DRL model, automatically plans the calibration time, calibration items, and calibration cycle, and realizes dynamic adjustment of the calibration plan to avoid resource waste. The entire calibration process is equipped with a multi-level anomaly protection mechanism. When situations such as hardware compatibility failure, data acquisition interruption, or excessive calibration deviation occur, the system will automatically pause the current calibration step, record the anomaly information completely, and issue an audible and visual prompt. After troubleshooting, the calibration process can be resumed from the breakpoint without having to restart the calibration job from the beginning.

[0055] The calibration plan has the ability to be updated automatically in real time. When there are sudden changes in the health status of the instrument, adjustments to the priority of the detection task, or insufficient standard gas reserves, the system will immediately recalculate and update the calibration plan.

[0056] Based on a multi-agent collaborative model, the system coordinates the working sequence of each module to avoid conflicts between modules. The multiple agents coordinate their working sequence according to the logical order of the entire calibration process. After hardware interaction and adaptation are completed, the data acquisition module is triggered to work. After data preprocessing is completed, the calibration calculation module is started to avoid resource conflicts caused by multiple modules running at the same time. At the same time, a multi-objective optimization algorithm is adopted to optimize the allocation of hardware resources, computing resources, and standard gas resources during the calibration process. For example, when multiple instruments are calibrated in parallel, computing power and standard gas are automatically allocated to prioritize the calibration needs of high-priority detection tasks, thereby realizing intelligent scheduling and resource optimization of calibration tasks.

[0057] The multi-agent collaborative model comprises three agents: hardware, computing power, and gas. The DRL uses the MADDPG algorithm. The state space consists of task priority, health status, and gas balance. The action space consists of calibration time and resource allocation. The reward function consists of task completion rate and resource consumption rate. The input is the scheduling feature vector, and the output is the optimal calibration plan. The training learning rate is 3e-4, the discount factor is 0.95, and the training step size is 8000.

[0058] Multi-objective optimization focuses on calibration efficiency, resource consumption, and detection task priority. It prioritizes the calibration needs of high-priority detection tasks and allocates standard gas and computing resources according to the calibration needs of each instrument when calibrating multiple instruments in parallel, avoiding resource contention and waste and improving overall scheduling efficiency.

[0059] This invention also provides an automatic calibration method for a gas chromatograph based on artificial intelligence, which, based on the above-described system, includes the following specific steps: First, hardware adaptation and standard gas preparation are carried out. A generalized hardware interface protocol is built through an adaptive recognition model that integrates CNN and GRU to automatically identify the gas chromatograph model, hardware configuration and communication protocol. At the same time, the flow rate and pressure of the standard gas are dynamically adjusted through the flow control component to complete the intelligent switching of multi-channel standard gases and instrument adaptation. Subsequently, multi-source data was collected and a virtual simulation scenario was constructed. Instrument hardware operating parameters and calibration data were collected in real time. A digital twin model was constructed through the fusion model of Bi-LSTM and GNN to simulate the instrument operating behavior under different detection conditions, predict calibration parameter anomalies and optimize initial values, and establish a virtual-real data synchronization mapping mechanism. Perform multimodal AI calibration and environmental compensation, collect various types of data from hardware, virtual simulation and environment, extract signal features and deeply fuse them through MS-CNN, LSTM and DRL fusion models, automatically generate optimal calibration parameters and realize online iteration, monitor calibration deviation in real time and adjust dynamically, and collect environmental parameters based on Transformer model and calculate compensation coefficients to adjust calibration parameters in a coordinated manner. Finally, health linkage and scheduling optimization are carried out. The health level of the instrument is evaluated through TGNN and adversarial training model, the lifespan of core components is predicted and maintenance suggestions are generated, and the parameters and cycle are optimized by linking the multimodal AI calibration module. Based on knowledge graph, calibration rules are extracted and strategy support is provided. The work of each module is coordinated through DRL and multi-agent model to generate the optimal calibration plan and optimize resource allocation.

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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 process, method, article, or apparatus.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automated calibration system for a gas chromatograph based on artificial intelligence, characterized in that, include: Intelligent hardware interaction module: Based on the CNN-GRU fusion model, a generalized hardware interface protocol is constructed to automatically identify the hardware parameters and communication protocols of multiple brands and models of gas chromatographs. It integrates flow control components and multi-channel standard gas storage units to realize plug-and-play functionality for multiple instrument models and intelligent delivery and switching of standard gases. Instrument Digital Twin Module: Based on the fusion model of attention mechanism bidirectional long short-term memory network and graph neural network, a dedicated digital twin model is constructed to collect instrument operation and calibration data, build a virtual calibration scenario, and realize instrument operation simulation, calibration parameter prediction and real-time data interaction between virtual and real. Multimodal AI calibration module: Based on the MS-CNN+LSTM+DRL fusion model, it collects and analyzes multi-source heterogeneous data, automatically generates and iteratively optimizes calibration parameters online; Environmental intelligent compensation module: Based on the Transformer attention mechanism neural network model, it collects multiple environmental parameters in the laboratory, analyzes the correlation between environmental parameters and calibration results, and dynamically calculates compensation coefficients; Instrument health management module: Based on time-series graph neural network and adversarial training model, it collects abnormal instrument operation data, constructs health status correlation map, assesses health level, predicts the remaining life of core components and generates maintenance suggestions, and adjusts calibration parameters and cycles accordingly; Calibration Knowledge Graph Module: Based on graph neural network and knowledge graph technology, it collects and organizes calibration-related industry standards, historical data and fault cases, constructs a dynamic knowledge graph, extracts calibration patterns and infers the optimal calibration strategy; Intelligent scheduling and management module: Based on deep reinforcement learning and multi-agent collaborative model, combined with instrument health status, remaining standard gas quantity and calibration knowledge, it generates the optimal calibration plan and optimizes the allocation of hardware, computing power and standard gas resources.

2. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 1, characterized in that, The generalized hardware interface protocol of the intelligent hardware interaction module is compatible with the carrier gas control port, injection interface and detector signal port of mainstream gas chromatographs. The flow control component adopts a mass flow controller and inert delivery pipeline, and uses reinforcement learning algorithm to dynamically optimize the intelligent delivery flow rate and pressure of standard gas in real time.

3. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 2, characterized in that, The instrument digital twin module captures the temporal variation characteristics of parameters through a bidirectional long short-term memory network model with an attention mechanism, and constructs a graph of the relationship between instrument components by combining it with a graph neural network model. The virtual calibration scenario can simulate the instrument's operating behavior under different detection conditions, predict abnormal calibration parameters, and optimize initial values.

4. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 3, characterized in that, The multimodal AI calibration module collects multi-source heterogeneous data including standard gas concentration data, instrument hardware parameters, virtual simulation data, and environmental parameters. It extracts multi-scale signal features through a multi-scale convolutional neural network, captures temporal change features through a long short-term memory network, and realizes feature weight allocation and iterative optimization of calibration parameters through a deep reinforcement learning model.

5. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 4, characterized in that, The environmental intelligent compensation module collects environmental parameters including temperature, humidity, air pressure, vibration, and electromagnetic interference intensity. It analyzes the correlation weight between environmental parameters and calibration results through a Transformer attention mechanism neural network model, dynamically adjusts the weight allocation in real time, calculates the compensation coefficient, and synchronizes the compensation coefficient to the instrument's digital twin module and multimodal AI calibration module in real time.

6. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 5, characterized in that, The instrument health management module constructs a health status correlation map through a time-series graph neural network model, combines an adversarial training model to eliminate abnormal data interference, and predicts the remaining lifespan of core components based on a long short-term memory time-series model and reinforcement learning algorithm. The generated maintenance suggestions and health assessment results are synchronized to the intelligent scheduling and management module.

7. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 6, characterized in that, The calibration knowledge graph module extracts multi-dimensional correlation information through deep learning knowledge mining algorithms, extracts calibration patterns and infers the optimal calibration strategy through graph neural network models and reinforcement learning algorithms, and can automatically identify and correct conflicting information in the knowledge graph, learn new calibration data and fault cases in real time, and continuously update the knowledge graph dynamically.

8. The automatic calibration system for a gas chromatograph based on artificial intelligence according to claim 7, characterized in that, The intelligent scheduling and management module generates a calibration plan by combining instrument usage frequency, detection task priority, instrument health status, remaining standard gas quantity, and calibration knowledge. It coordinates the working sequence of each module through a multi-agent collaborative model and optimizes the allocation of hardware, computing power, and standard gas resources based on deep reinforcement learning and the multi-agent collaborative model.

9. An automatic calibration method for a gas chromatograph based on artificial intelligence, based on the system described in any one of claims 1-8, characterized in that, The specific steps include the following: Step 1: Hardware adaptation and standard gas preparation. The instrument parameters and communication protocols are automatically identified through a universal hardware interface protocol, and the intelligent delivery flow rate and pressure of the standard gas are dynamically adjusted to complete the adaptation of multiple instrument models and intelligent switching of multiple standard gas channels. Step 2: Multi-source data acquisition and virtual simulation. Collect instrument operation and calibration data, construct a digital twin model and virtual calibration scenario, predict calibration parameter anomalies and optimize initial values, and establish a virtual-real data synchronization mapping between the virtual model and the physical instrument. Step 3: Multimodal AI calibration and environmental compensation, which integrates multi-source data to generate and iteratively optimize calibration parameters, collects environmental parameters to calculate compensation coefficients, and adjusts calibration parameters accordingly; Step 4: Health linkage and scheduling optimization. Assess the instrument's health level, predict the lifespan of core components, and generate maintenance recommendations. Optimize calibration parameters and cycles in a coordinated manner. Combine knowledge graph reasoning with calibration strategies to generate the optimal calibration plan and optimize the allocation of hardware, computing power, and standard gas resources.

10. The automatic calibration method for a gas chromatograph based on artificial intelligence according to claim 9, characterized in that, The calibration parameter deviation monitoring in step 3 uses an MS-CNN+LSTM+DRL fusion model, which can automatically identify the cause of deviation and dynamically adjust the calibration parameters.

Images