Automatic liquefied natural gas sampling method convenient to operate

Abstract

The invention relates to the technical field of liquefied natural gas sampling, and discloses an automatic liquefied natural gas sampling method convenient to operate. The method comprises the following steps: acquiring historical sampling data in a preset sampling period and generating a data set; extracting historical sample characteristic parameters, and establishing a trend change characteristic matrix; constructing a multi-dimensional state space according to the matrix, and calculating the state aggregation degree of historical sampling abnormal events to determine a sample feature early warning index set; extracting real-time sample characteristic parameters and establishing vectors, and performing spatial position correlation calculation with the early warning index set to determine real-time risk correlation factors; and generating a risk probability prediction model in combination with the real-time risk association factor, the historical sample feature parameter set and the trend change feature matrix, predicting a sampling anomaly probability and generating an adjustment suggestion. According to the method, historical data and real-time analysis are utilized, the sampling accuracy and adaptability are improved, manual intervention is reduced, and a more reliable mode is provided for liquefied natural gas sampling.

Description

Technical Field

[0001] The present invention relates to the technical field of liquefied natural gas sampling, and in particular to an automatic sampling method for liquefied natural gas that is easy to operate. Background Art

[0002] During the production, transportation, and storage of liquefied natural gas (LNG), sampling and analysis are crucial for ensuring its quality and safety. Traditional LNG sampling methods rely heavily on manual labor, requiring operators to travel to sampling points at designated times, manually opening valves, and collecting samples. This method is not only labor-intensive but also prone to biased sampling results due to operational variability. For example, inaccurate sampling timing and inconsistent sampling volumes can adversely impact subsequent composition analysis and quality assessment.

[0003] The rapid development of the liquefied natural gas (LNG) industry has placed higher demands on sampling efficiency and accuracy. While existing automated sampling equipment reduces manual intervention to a certain extent, most employ fixed sampling cycles and patterns, lacking effective utilization of historical data. These automated sampling systems are unable to adjust sampling strategies based on actual LNG state changes. When encountering abnormalities, they often fail to issue timely warnings and adjust sampling parameters. This results in sampling data that fails to truly reflect the real-time state of LNG, compromising judgments on its quality and safety.

[0004] Existing sampling methods lack a systematic analysis and prediction mechanism for handling abnormal events. This makes it difficult to effectively extract features and analyze trends from abnormal information in historical sampling data, making it difficult to establish scientific early warning indicators. This makes it impossible to quickly identify potential risks during real-time sampling and to take appropriate corrective measures in advance. This, to a certain extent, increases safety risks in the production and transportation of liquefied natural gas. Summary of the Invention

[0005] The object of the present invention is to provide an automatic sampling method for liquefied natural gas that is easy to operate, so as to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides an easy-to-operate automatic sampling method for liquefied natural gas, the method comprising: Acquiring historical sampling data of liquefied natural gas within a preset sampling period, and generating a historical sampling data set of liquefied natural gas based on the historical sampling data; Extracting sample characteristic parameters from the historical sampling data set to obtain a historical sample characteristic parameter set of liquefied natural gas, and establishing a trend change characteristic matrix based on the temporal trend of the historical sample characteristic parameters; Constructing a multidimensional state space based on the trend change feature matrix, calculating the state aggregation degree of historical liquefied natural gas sampling abnormal events in the multidimensional state space, and determining a sample feature early warning indicator set based on the state aggregation degree; Extracting real-time sample characteristic parameters based on real-time sampling data of liquefied natural gas, establishing a real-time characteristic state vector, and performing spatial position correlation calculation with the sample characteristic warning indicator set in the multidimensional state space to determine a real-time risk association factor; A risk probability prediction model is generated based on the real-time risk association factor, the historical sample characteristic parameter set and the trend change characteristic matrix to predict the probability of abnormal sampling of liquefied natural gas and generate sampling adjustment suggestions.

[0007] Preferably, historical sampling data of liquefied natural gas within a preset sampling period is obtained, and a historical sampling data set of liquefied natural gas is generated based on the historical sampling data, specifically: Collecting historical sampling data of liquefied natural gas within the preset sampling period; Perform data validity screening based on the change range, change trend and fluctuation amplitude of each parameter data in the historical sampling data, and remove parameter data that exceeds a preset reasonable range; The historical sampling data after data validity screening are classified and integrated according to the parameter correspondence to generate a historical sampling data set of liquefied natural gas containing temperature parameters, pressure parameters and composition parameters.

[0008] Preferably, the sample characteristic parameters of the historical sampling data set are extracted to obtain a historical sample characteristic parameter set of liquefied natural gas, and a trend change characteristic matrix is ​​established according to the temporal trend of the historical sample characteristic parameters, specifically: Performing historical sample feature parameter extraction processing on the historical sampling data set; Establishing a historical sample feature parameter set according to different dimensions of the historical sample feature parameters; Calculating the change trend values ​​of different historical sample characteristic parameters in the historical sample characteristic parameter set within a specified time period; Arrange the change trend values ​​of the historical sample characteristic parameters in chronological order to form a trend change characteristic matrix; the rows of the trend change characteristic matrix represent different historical sample characteristic parameters, and the columns represent corresponding time series data.

[0009] Preferably, a multidimensional state space is constructed based on the trend change feature matrix, the state aggregation degree of the historical liquefied natural gas sampling abnormal events in the multidimensional state space is calculated, and a sample feature warning indicator set is determined based on the state aggregation degree, specifically: The dimension of the multidimensional state space is determined according to the number of historical sample characteristic parameters in the trend change characteristic matrix; the coordinate axes of the multidimensional state space are different historical sample characteristic parameters of the trend change characteristic matrix, and the characteristic position coordinates in the multidimensional state space are the change trend values ​​of the historical sample characteristic parameters; Calculating the aggregation degree of state points when abnormal sampling of liquefied natural gas occurred in the past based on the coordinates of each characteristic position in the multidimensional state space, and obtaining the state aggregation degree value of the abnormal sampling event of the liquefied natural gas in the multidimensional state space; A warning threshold is set according to the state concentration value, and historical sample feature parameters corresponding to feature position coordinates whose state concentration values ​​exceed the warning threshold are selected as a sample feature warning indicator set.

[0010] Preferably, real-time sample characteristic parameters are extracted based on real-time sampling data of liquefied natural gas, a real-time characteristic state vector is established, and spatial position correlation calculation is performed with the sample characteristic warning indicator set in the multidimensional state space to determine the real-time risk association factor, specifically: Real-time collection of liquefied natural gas sampling data under the current sampling status; extracting real-time sample characteristic parameters based on the real-time sampling data, and forming a real-time characteristic state vector according to the parameter sequence of the historical sample characteristic parameter set; The real-time characteristic state vector is mapped to the multidimensional state space, the spatial distance between the real-time characteristic state vector and the characteristic position coordinates in the sample characteristic warning indicator set is calculated, the spatial position correlation is calculated based on the spatial distance, and the real-time risk association factor of the current sampling state of the liquefied natural gas is determined based on the spatial position correlation.

[0011] Preferably, a risk probability prediction model is generated based on the real-time risk association factor, the historical sample characteristic parameter set, and the trend change characteristic matrix to predict the probability of abnormal sampling of liquefied natural gas and generate sampling adjustment suggestions, specifically: Using the real-time risk correlation factor and the historical sample characteristic parameter set as input data for a risk probability prediction model; Using the risk probability prediction model, the change trend values ​​of the historical sample characteristic parameters in the trend change characteristic matrix are calculated to predict the probability value of sampling anomaly under the current sampling state of liquefied natural gas; Generate liquefied natural gas sampling adjustment recommendations based on the predicted sampling anomaly probability value and the preset sampling adjustment threshold.

[0012] Preferably, the method further comprises constructing a digital twin model of the liquefied natural gas sampling process, collecting real-time sampling data of the liquefied natural gas through sensors deployed at the sampling site, and synchronizing the real-time sampling data into the digital twin model; Simulating actual sampling status in the digital twin model to predict sampling development trends and potential sampling anomalies; Set up an intelligent sampling control strategy. The intelligent sampling control strategy combines the degree of mutual influence between real-time sampling data and the execution conditions of the control action priority to determine whether there is a logical conflict in the rule engine. If so, optimize the intelligent sampling control strategy. Specifically: The intelligent sampling control strategy constructs a comprehensive feature vector based on the coupling relationship between the sampling quality prediction value and the component stability estimation value, combines the preset control action priority, and uses the sequence prediction model to predict the risk value of the logical conflict in the rule engine, and compares it with the predetermined threshold to determine whether there is a logical conflict; The sampling operation is automatically performed according to the optimized intelligent sampling control strategy, and the control behavior and feedback data are recorded in real time, and the digital twin model is continuously adjusted based on the recorded feedback data.

[0013] Preferably, the actual sampling state is simulated in the digital twin model to predict the sampling development trend and potential sampling anomalies, specifically including: The entire sampling area is divided into multiple monitoring sub-areas, each of which corresponds to a node in the sequence. The connection relationship between nodes is established based on the spatial adjacency relationship of the sampling positions or the parameter migration path. Each node is associated with multiple sensor data to form the node's feature vector. The initial sequence structure of the sampling area is constructed. The sampling state at each moment is represented as a sequence, and the continuous time forms the sequence data input; the label data is the target variable or sampling abnormal state mark of each node in a certain time window in the future; Use the sequence prediction model to extract the spatiotemporal features of the nodes; the sequence prediction model integrates the sequence at each time step to obtain the embedded representation of the node; The node representation is input into the subsequent time series prediction module; the temperature or pressure variables of each node in the future time step are predicted; the prediction results are scored for anomalies based on the actual historical distribution; if the predicted value of a node deviates from the normal change trend, it is marked as a potential sampling anomaly.

[0014] Preferably, setting an intelligent sampling control strategy and optimizing the intelligent sampling control strategy further includes: Based on the conversion of the sampling quality prediction value and the component stability estimation into a comprehensive feature vector, the comprehensive feature vector is used as the input of the polynomial regression model. In combination with the currently set control action priority, the polynomial regression model uses each set of comprehensive feature vectors to predict the risk value label of the logical conflict of the rule engine as the prediction target, and takes minimizing the sum of the prediction errors of the risk value labels of the logical conflict of all rule engines as the training target. The polynomial regression model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The risk value of the logical conflict of the rule engine is determined based on the model output result; The obtained risk value of the logical conflict in the rule engine is compared with a predetermined threshold. If the risk value of the logical conflict in the rule engine is greater than or equal to the predetermined threshold, it indicates that the rule engine has a logical conflict. At this time, an early warning signal is generated and the intelligent sampling control strategy is optimized. If the risk value of the logical conflict in the rule engine is less than the predetermined threshold, it indicates that the rule engine does not have a logical conflict. At this time, no early warning signal is generated and no additional adjustment is required.

[0015] Preferably, the digital twin model is continuously adjusted based on the recorded feedback data, specifically including: in the initial stage, constructing a digital twin model for simulating the liquefied natural gas sampling process, and at the same time establishing a set of data structures for recording, associating the correspondence between each control behavior, environmental response and model output; using historical feedback data to construct multiple small sample task sets, each task represents a specific sampling scenario, and each task contains a training set and a validation set; gradually fine-tuning the digital twin model in each task; optimizing the initial parameters of the model through repeated iterations of multiple tasks; in actual operation, after each control behavior is executed, the collected feedback data is immediately constructed as a new task; the new task is input into the trained meta-learning framework for iteration, the current twin model is locally fine-tuned, and the fine-tuned model is retained as the current scene sub-model; the system regularly verifies the prediction performance of the current twin sub-model; if the error decreases or meets the self-set threshold, the system will extract the fine-tuned parameter changes, feed them back to the main model parameter set, update the main model or add it to the experience model pool.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By acquiring historical sampling data within a preset sampling period and generating a dataset, we provide a wealth of fundamental information for subsequent analysis. Extracting characteristic parameters from historical samples and establishing a trend change matrix clearly demonstrates the state changes of LNG over different time periods, providing a deeper understanding of its inherent characteristics.

[0017] A multidimensional state space is constructed based on the trend change feature matrix. The sample feature early warning indicator set is determined based on the state aggregation of historical sampling anomalies. This makes the setting of early warning indicators more scientific and reasonable, and can more accurately capture potential anomalies. During the real-time sampling process, by extracting real-time sample feature parameters and establishing vectors, spatial correlation calculations are performed with the early warning indicator set to determine real-time risk association factors, enabling timely identification of potential risks in real-time sampling data.

[0018] A risk probability prediction model, generated based on real-time risk correlation factors, a set of historical sample characteristic parameters, and a trend change characteristic matrix, effectively predicts the probability of sampling anomalies, providing a reliable basis for generating sampling adjustment recommendations. This approach dynamically adjusts sampling strategies based on the actual state of liquefied natural gas, reducing the limitations of fixed sampling patterns and improving sampling flexibility and adaptability.

[0019] The entire process is automated, reducing potential errors caused by manual intervention and making sampling data more accurate and reliable, contributing to a more precise assessment of LNG quality and safety. By fully leveraging historical data and dynamically analyzing real-time data, potential risks can be identified in advance and sampling parameters adjusted promptly, providing stronger safeguards for LNG production, transportation, and storage. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a working principle diagram of the easy-to-operate automatic sampling method for liquefied natural gas according to the present invention; Figure 2 Flowcharts generated for historical sampling data; Figure 3 Flowcharts determined for the multidimensional state space and early warning indicator set; Figure 4 Flowchart generated for risk probability prediction and adjustment recommendations. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] See also Figure 1 The present invention provides an automatic sampling method for liquefied natural gas that is easy to operate, the method comprising: By establishing a multidimensional state space and risk probability prediction model, intelligent monitoring and adjustment of the sampling process are achieved. The method first obtains historical liquefied natural gas (LNG) sampling data within a preset sampling period to generate a historical sampling dataset. Sample feature parameters are extracted from the dataset to establish a historical sample feature parameter set. A trend change feature matrix is ​​constructed based on the temporal trend of the parameters. Based on this matrix, a multidimensional state space is constructed, and the state aggregation degree of historical sampling anomaly events is calculated to determine a sample feature warning indicator set. Real-time sampling data is extracted through feature extraction to form a real-time feature state vector. Spatial correlation between this vector and the warning indicator set is calculated within the multidimensional state space to obtain a real-time risk correlation factor. Combining the historical sample feature parameter set and the trend change feature matrix, a risk probability prediction model is generated to predict the probability of sampling anomalies and output adjustment recommendations.

[0023] Example 1: See Figure 2 The historical data collection process of the liquefied natural gas automatic sampling method is realized through a sensor network installed in storage tanks, transmission pipelines and sampling ports. The temperature sensor uses a platinum resistance element, and the measurement range covers the typical working conditions of liquefied natural gas. The installation location includes the liquid phase, gas phase and gas-liquid junction. The pressure sensor uses a piezoresistive transmitter. A measuring point is arranged every 50 meters in the main pipeline, and a high-frequency dynamic pressure sensor is added near the sampling port. The component analysis uses an online gas chromatograph, which completes a full component scan every 5 minutes, mainly detecting the content of methane, ethane, propane and nitrogen. All sensors transmit real-time data to the data acquisition system in the central control room through 4-20mA analog signals or Modbus digital protocols.

[0024] Data validity screening uses a multi-level verification mechanism. The raw data first undergoes a range check to eliminate outliers that clearly exceed the physically reasonable range. The lower limit of the reasonable range of temperature parameters is determined by the boiling point of liquefied natural gas, and the upper limit takes into account the influence of environmental heat transfer; the reasonable range of pressure parameters is determined by combining the pipeline design pressure and the safety valve setting value. The second-level verification is based on time continuity and uses a local outlier detection algorithm within a sliding window to identify and eliminate instantaneous jump data. The third-level verification is cross-validated through multi-sensor data. When there is a physical contradiction between the temperature and pressure measurements at the same location, a manual review process is triggered. After three levels of screening, the data is aligned by timestamp, and missing values ​​are supplemented by linear interpolation of the previous and next time points.

[0025] The construction of the structured data set follows the storage specifications of the time series database. Temperature parameters are stored by measurement location, including derivative indicators such as the average temperature of the liquid phase, the average temperature of the gas phase, and the maximum temperature difference. The pressure parameter distinguishes between static pressure and dynamic fluctuation components. The dynamic component is extracted through digital filtering to extract the 0.1-10Hz frequency band characteristics. In addition to the concentration of each component, the composition data also calculates process indicators such as the methane number and the Wobbe index. The data set adopts a hierarchical storage architecture. The original sampling values ​​are stored in the cache layer, and statistics such as the 5-minute mean and extreme values ​​are stored in the relational database. Aggregated data above the hour level are transferred to the time series database for long-term storage.

[0026] Sample feature extraction utilizes a multi-scale analysis approach. Within a 30-minute sliding window, temperature parameters are extracted at three levels of features: the statistical characteristics of the raw measurements reflect the overall distribution, the first-order difference sequence characterizes short-term variation patterns, and the frequency of extreme values ​​within the window indicates the risk of abnormal fluctuations. Feature extraction for pressure parameters focuses on dynamic characteristics, including the coefficient of variation of fluctuation amplitude, the decay time of the autocorrelation function, and the frequency domain energy distribution ratio. Feature design for compositional parameters considers interactions between components, primarily extracting the concentration change slope of key components, the eigenvalues ​​of the multi-component correlation matrix, and the moving standard deviation of specific component ratios.

[0027] Feature vector normalization utilizes a dynamic benchmark. The normalization coefficient for each feature parameter is dynamically calculated based on its historical distribution over the last 24 hours to avoid feature distortion caused by long-term operating condition drift. Temperature features are normalized using a quantile-based method, while pressure features are normalized using a logarithmic Z-score. Component features are compressed to the unit interval using a sigmoid function. The normalized feature vectors are organized into a matrix in chronological order, and missing time points are filled using cubic spline interpolation to ensure time series continuity.

[0028] The construction of the trend change feature matrix incorporates a joint time-frequency analysis. The row vectors of the matrix correspond to different feature parameters, while the column vectors represent consecutive time points. Each matrix element contains not only the normalized eigenvalue but also a marker indicating the change trend of that feature over the past six time windows. Trend markers are represented symbolically, with different encoding rules corresponding to rising trends, falling trends, and stable states. The matrix is ​​updated using a rolling window approach, removing the oldest time series segment as new data arrives, maintaining a constant time span.

[0029] The dimensionality reduction of the feature matrix utilizes a nonlinear manifold learning method. A locally linear embedding algorithm maps the 24-dimensional feature vectors into an 8-dimensional latent space, preserving over 90% of the variance in the original data. During the dimensionality reduction process, the local neighborhood size is set to 15 sample points, and the regularization parameter is automatically adjusted based on the sparsity of the matrix. The reduced feature matrix serves as the basis for subsequent state-space analysis, with its dimensions matching the critical degrees of freedom of the physical process.

[0030] Early identification of abnormal features relies on matrix pattern analysis. Singular value decomposition is performed on the feature matrix, and the first three principal components are extracted as monitoring indicators. When the Mahalanobis distance of the principal component scores exceeds the control limit, a feature backtracking mechanism is triggered. This mechanism searches backward along the timeline to locate the first occurrence of the feature anomaly and, combined with concurrent process parameters, determines the root cause of the anomaly. The feature matrix is ​​also used to construct a state transition probability graph, with nodes representing feature cluster centers and edge weights reflecting state transition frequencies, providing a visualization tool for analyzing anomaly propagation paths.

[0031] The comparison of real-time data with historical features utilizes a dynamic time warping algorithm. This algorithm automatically aligns the time axes of the real-time feature sequence with the historical template sequence and calculates the cumulative distance along the optimal matching path. The distance metric uses a modified Euclidean distance formula, applying variable weights to different feature dimensions. The comparison results generate a similarity score. When the score falls below a threshold, the system automatically retrieves the historical database for similar treatment records, providing the operator with a reference solution.

[0032] Example 2: See Figure 3 , the construction of the multidimensional state space uses the historical trend feature matrix as the input source. The number of dimensions of this space is determined by the number of rows of the feature matrix, and each dimension corresponds to a type of sample characteristic parameter. The spatial coordinate system adopts a standardized dimension, and the scale of each axis is normalized using the range of the characteristic parameter in the historical data. The generation of characteristic position coordinates is achieved through direct mapping of the matrix column vector, that is, the characteristic vector corresponding to each time point in the matrix is ​​a coordinate point in the state space. The spatial topology adopts a non-uniform partitioning strategy, and the regional density is proportional to the distribution frequency of the historical data points.

[0033] The calculation of the state aggregation of abnormal events adopts an adaptive grid division mechanism. First, the data quantiles of each dimension of the state space are calculated, and the space is divided into hypercube units according to the quantiles. The unit side length is dynamically adjusted according to the density of the data distribution. The frequency of historical abnormal events in each unit is counted, and the local aggregation is calculated by the ratio of the frequency to the unit volume. The aggregation value is logarithmically transformed to obtain the final index value, which reflects the concentration of abnormal events in a specific spatial area. The warning threshold is set using the quantile method, and the 85% quantile of the aggregation value is taken as the baseline. The determination of the sample feature warning indicator set is achieved by traversing the grid units: when the aggregation of a unit exceeds the threshold, all feature parameters adjacent to the center point of the unit are included in the warning indicator set.

[0034] The spatial mapping of the real-time feature state vector utilizes coordinate transformation technology. The raw real-time feature vector undergoes the same normalization as historical data, converting each eigenvalue into a state-space coordinate system. During the mapping process, rotational and translational transformations are applied to eliminate spatial offsets caused by measurement system deviations. The position of the real-time vector in state space is represented as a dynamic point mass with a directional property, and its motion trajectory is described by a velocity vector using a differential equation. The spatial position of this point mass is updated every 60 seconds, synchronized with the real-time data sampling period.

[0035] Spatial position correlation calculations use a hybrid distance measurement model. For each feature position coordinate in the early warning indicator set, the spatial distance between it and the real-time vector is calculated. Distance calculations use different measurement methods depending on the feature parameter type: Continuity parameters use the improved Minkowski distance, which introduces a parameter correlation matrix as a weighting factor:

[0036] in: represents the weighted Minkowski distance value; is the total number of state space dimensions; Represents the weight coefficient of the k-th dimension feature parameter, which is determined by the variance contribution rate of the parameter in the historical data; is the coordinate value of the real-time vector in the kth dimension; is the coordinate value of a feature position in the kth dimension in the early warning indicator set; The distance norm adjustment parameter is in the range of [1.5, 2.5] and is dynamically adjusted according to the spatial distribution topology of the early warning indicators. The discrete characteristic parameters are measured using a combination of Hamming distance and Jaccard coefficient.

[0037] The calculation of real-time risk association factors uses a multi-layer fusion architecture. First, the distance set between the real-time vector and each warning coordinate is calculated, and the distance set is standardized and sorted. Then, the distance weighting coefficient is set according to the concentration value of the warning indicator. The higher the concentration, the greater the influence weight of the warning point. At the spatial topology level, a multi-layer concentric sphere centered on the real-time vector is constructed, and the density gradient of the warning points within each spherical layer is calculated. The final risk factor is a weighted fusion of the output values ​​of three modules: the basic distance module outputs the minimum normalized distance value; the spatial distribution module outputs the directional concentration index of the warning point; and the time coupling module outputs the movement correlation between the current vector and the historical anomaly pattern.

[0038] The dynamic calibration of risk factors incorporates a self-correction mechanism. The system records the correspondence between risk factors and subsequent actual anomaly occurrences, forming a validation dataset. When prediction deviations occur repeatedly, a parameter calibration procedure is initiated: first, the weight distribution ratio in the spatial distance calculation is checked and the weight coefficients are optimized using a gradient descent algorithm. Second, the quantile intervals of the state space grid are adjusted, and the grid is further refined in areas with high anomaly incidence. Finally, the weight coefficients in the risk factor fusion formula are updated to enhance the decision-making influence of recent validation samples. The calibration process utilizes sample data within a sliding time window, with the window length adaptively adjusting to the fluctuations of the risk factor.

[0039] The robust design of the spatial analysis engine includes a three-dimensional fault-tolerance mechanism. First, multiple spatial mapping backup channels are set up, automatically switching to dimensionality-reduced subspace projection when data overflow occurs in the primary mapping path. Second, an anti-interference algorithm is built into distance calculations, implementing a temporary isolation strategy for abnormally deviated spatial points. Third, a virtual state point auxiliary system is established to generate reference clusters of points around real-time vectors to distinguish true anomalies from temporary interference. The system performs spatial topology integrity diagnosis every minute, calculating the geometric similarity between the theoretical boundaries of the feature space and the actual distribution.

[0040] Example 3: See Figure 4 The architectural design of the risk probability prediction model adopts a hybrid approach that combines deep neural networks with statistical learning. The model input layer receives two types of data: real-time risk correlation factors as dynamic feature inputs, and historical sample feature parameter sets as static feature inputs. After standardization, the input data enters the feature fusion module for spatiotemporal alignment. The fusion of dynamic and static features adopts the attention mechanism, and the fusion weight is determined by calculating the mutual information between features. The core processing unit of the model contains three parallel sub-networks: the temporal feature extraction network processes parameters with time dependence, the spatial feature extraction network analyzes the distribution characteristics in the multidimensional state space, and the association feature network mines the implicit relationship between dynamic and static features.

[0041] The structural parameters of the long-short-term memory network are adaptively configured based on the feature dimension. The activation functions for the input, forget, and output gates use the sigmoid transformation, and the activation function for candidate memory cells uses the hyperbolic tangent function. The network's hidden layer has 32 memory cells. The output of each time step is mapped to a 24-dimensional feature space through a fully connected layer, maintaining the same dimensionality as the original input. The forward and backward propagation results of the bidirectional LSTM are concatenated along the feature dimension to form a 64-dimensional joint feature representation. Network training uses a time-step expansion method, with the expansion length set to 10 time steps based on the update frequency of the real-time risk correlation factor.

[0042] The output layer design for anomaly probability prediction takes multi-task learning into consideration. The three output branches for temperature anomaly probability, pressure anomaly probability, and composition anomaly probability share the same feature extraction network and branch into independent structures at the final fully connected layer. The output of each branch is converted to a probability value using a softmax function, while the sum of the probabilities of the three branches remains constant. The loss function for probability prediction uses a modified cross-entropy form, introducing class weights to balance the impact of uneven sample distribution. The loss function is calculated as follows:

[0043] in: Indicates the total loss value; Indicates the anomaly category index (1 for temperature anomaly, 2 for pressure anomaly, 3 for composition anomaly); The weight coefficient of the c-th type of anomaly is determined by the frequency of occurrence of this type of anomaly in the training data; represents the total number of time steps; represents the true label of the c-th type of anomaly at the t-th time step; Represents the probability value of the c-th type of anomaly at the t-th time step predicted by the model. Weight coefficient The dynamic adjustment of follows the principle of minority class weighting and is recalculated every 100 iterations during training.

[0044] The logic for generating sampling adjustment recommendations adopts a hierarchical decision-making mechanism. The first-level decision is based on the absolute value of a single abnormality probability, setting three thresholds of 0.3, 0.7, and 0.85 to divide the risk level into four levels: attention, warning, serious, and critical. The second-level decision considers the combined effect of multiple abnormality probabilities, and triggers a combined warning when any two abnormal probabilities exceed 0.5 at the same time. The third-level decision introduces a time accumulation factor, calculates the probability sliding average of three consecutive time steps, and increases the adjustment strength when the average shows a monotonically increasing trend. The output of the decision engine includes three basic operating instructions: the flow adjustment instruction is achieved by changing the opening of the control valve, and the adjustment amplitude is in a piecewise linear relationship with the abnormality probability value; the pipeline switching instruction activates the backup sampling circuit, and at the same time marks the original circuit as requiring maintenance; the emergency stop instruction immediately cuts off the sampling process and starts the safety isolation procedure.

[0045] Control parameter optimization utilizes a model predictive control framework. The optimization objective function consists of three components: a suppression term for abnormality probability, a penalty term for operating costs, and a constraint term for system stability. Each optimization calculation is performed within a 15-second window, which is divided into five equally spaced control periods. The optimization solution utilizes a sequential quadratic programming algorithm, recalculating the optimal control sequence after each real-time risk correlation factor update. The control parameters are transmitted to the actuator via an industrial fieldbus. The transmission protocol incorporates a double-check mechanism to ensure command accuracy.

[0046] Feedback data collection and model updates form a closed-loop system. After each control operation is executed, the system records three cycles of response data: the sampling parameter change curve within 60 seconds of execution, the abnormality probability evolution trajectory within 120 seconds, and the system stability index within 180 seconds. After feature extraction, the feedback data is stored in the model update queue. The queue adopts a priority scheduling strategy, and cases with insufficient abnormality probability reduction are given higher processing priority. The online update of model parameters uses a small-batch gradient descent method. The batch size is dynamically adjusted based on the timeliness of the feedback data and is controlled between 16 and 64 samples.

[0047] Enhanced interpretability of predictive models is achieved through feature importance analysis. After each prediction cycle, the system calculates the contribution of input features to the output probability, quantifying this contribution using an integrated gradient method. The analysis results are displayed as a heat map, annotating key features and their impact. When significant changes in the distribution of feature contributions are detected, a model diagnostic process is triggered to check for potential feature drift or concept drift. The diagnostic results are used to guide feature engineering optimization and, if necessary, initiate feature reselection.

[0048] The temporal resolution of anomaly predictions utilizes multi-scale fusion technology. The basic prediction unit outputs probability values ​​at a 60-second interval. Three auxiliary predictors simultaneously run: a high-frequency predictor analyzes short-term fluctuation patterns at 15-second intervals, a low-frequency predictor captures trend changes at 5-minute intervals, and an event predictor specifically detects sudden anomalies. Multi-scale prediction results are fused using a Bayesian inference framework, ultimately outputting a comprehensive risk probability that considers different time scales. The fusion weights are dynamically adjusted every 24 hours based on the recent accuracy of each predictor.

[0049] The spatial visualization of risk probabilities utilizes 3D isosurface rendering technology. The anomaly probabilities of temperature, pressure, and composition serve as spatial coordinate axes, and the state at each sampling moment is represented as a point in the probability space. The system draws the probability isosurface in real time, reflecting the risk level through color gradients. When a probability point crosses the boundary of the isosurface, a dynamic warning signal is generated, and the angle and speed of the crossing are recorded. The visualization system supports time-reversal, reproducing the evolution of the probability space over any historical period.

[0050] A multi-dimensional evaluation system is established to continuously monitor model performance. Accuracy assessment compares forecast results within a rolling time window with actual conditions, calculating recall, precision, and F1 scores. Timeliness assessment measures the average latency from anomaly occurrence to alert issuance. Stability assessment measures the fluctuations in consecutive forecast results. Evaluation results are reported daily. Model retraining is initiated when any metric falls below a preset benchmark. Retraining data includes both recent real-world cases and historically representative scenarios to maintain a balanced data distribution.

[0051] The sampling adjustment strategy is validated using a digital twin simulation environment. Before actual control is implemented, the strategy is simulated in a virtual environment. The simulation model incorporates fluid dynamics, thermodynamic transfer, and component diffusion equations, accurately reflecting the dynamic characteristics of the sampling system. Deviations between simulation results and actual operation are used to modify the strategy parameters, with the correction coefficients determined through least squares fitting. After each strategy optimization, a stress test is performed in the simulation environment to simulate the control effect under extreme operating conditions.

[0052] Example 4: The construction of a digital twin model is based on the physical characteristics of the liquefied natural gas sampling system and real-time data fusion. The model input includes three types of data: real-time sensor data, equipment operation logs, and environmental parameters. Sensor data is collected by temperature sensors, pressure transmitters, and gas chromatographs deployed in storage tanks, pipelines, and sampling ports, with a sampling frequency of once per second. The equipment operation log records valve opening, pump status, and control instructions, with millisecond-level timestamp accuracy. Environmental parameters include atmospheric temperature, humidity, and air pressure, and are updated every 5 minutes. This data is transmitted to the digital twin engine via industrial communication protocols, and the operating status of the sampling system is synchronously reconstructed in a virtual environment.

[0053] The geometric modeling of the digital twin utilizes a parametric approach. The tank model includes an inner tank, an outer tank, and a vacuum interlayer. The piping system is meshed in three dimensions according to its actual routing, and independent fluid domains are set for the sampling loop. The physical field simulation is based on computational fluid dynamics principles, using unstructured grid cells and mesh refinement near the walls. Material properties are dynamically adjusted based on the actual composition of the liquefied natural gas (LNG), taking into account the varying proportions of major components such as methane and ethane. Boundary conditions include inlet flow, outlet pressure, and wall heat transfer coefficient, all of which are dynamically derived from real-time data.

[0054] The anomaly prediction module utilizes a dual-path analysis architecture. A physics-based numerical simulation path solves the governing fluid dynamics equations and calculates the distribution of temperature, pressure, and concentration fields. A data-based machine learning path analyzes historical anomaly patterns and extracts characteristic indicators. The results from these two paths are combined using confidence-weighted fusion to output a comprehensive anomaly score. The scoring system uses a percentage system, with higher scores indicating greater anomaly risk. When the score exceeds 70, an early warning mechanism is triggered, and the system automatically checks the evolutionary trends of relevant parameters.

[0055] The rule engine for the intelligent sampling control strategy contains 127 production rules, organized by anomaly type. The temperature anomaly rule set includes 35 rules to address scenarios such as sudden temperature rises and gradient anomalies; the pressure anomaly rule set includes 42 rules to address pressure fluctuations and pulsation anomalies; and the composition anomaly rule set includes 50 rules to address issues such as component deviations and impurity excesses. Each rule consists of three components: a precondition, a confidence factor, and an action. The preconditions are described using fuzzy logic. For example, "a rapid temperature rise rate" is defined as an increase of more than 0.5°C per minute. The confidence factor reflects the historical accuracy of the rule. The initial value is set based on expert experience and dynamically adjusted based on actual performance.

[0056] Model version management uses blockchain technology to ensure traceability. Each model update generates a block containing the update date and time, a description of the changes, an impact assessment, a summary of test results, and the signature of the responsible party. Blocks are linked via a hash chain, forming an immutable record of updates. The system retains copies of the last ten model versions, enabling rapid rollback to any historical version. Difference analysis between versions helps identify optimization areas and avoid duplicative modifications.

[0057] The field deployment architecture utilizes an edge computing model. The digital twin engine runs on an industrial edge server, close to the data source to reduce latency. Sensor data is preprocessed locally, including filtering, denoising, and feature extraction, before being transmitted to the twin model. Control commands are delivered directly to the actuators via hardwiring to ensure rapid response. A human-machine interface (HMI) is deployed in the central control room, displaying the virtual scene and real-time analysis results. Operators can use this interface to adjust model sensitivity and intervene in the execution of control strategies.

[0058] Knowledge accumulation for exception handling utilizes case-based reasoning. Each actual exception event is encoded as a case feature vector, containing information such as environmental parameters, device status, abnormal manifestations, and treatment measures. The case library is organized hierarchically, with the top layer categorized by exception type and the lower layers clustered by feature similarity. When a new anomaly alert is triggered, the system searches for similar historical cases and recommends a treatment plan. The case matching algorithm considers feature weights, assigning different impact factors to recent cases and historically representative cases.

[0059] The digital twin model and the prediction system collaborate via an event bus. The model's output, anomaly scores and virtual scenario data, are published to a message queue, which the risk probability prediction model consumes for comprehensive assessment. The prediction model can reverse-query the digital twin's historical simulation data to verify parameter relationships in specific scenarios. A buffering mechanism is used for data exchange between the two systems to avoid performance bottlenecks caused by high-frequency communication.

[0060] The system's security design incorporates multiple safeguards. Data transmission utilizes industrial encryption protocols to prevent man-in-the-middle attacks. User access is controlled through role-based permissions, with operators, engineers, and administrators having distinct functional permissions. Key control instructions require dual authentication, and routine operations record operator information. The system undergoes regular penetration testing, and discovered vulnerabilities are addressed. All security events are recorded in a dedicated log for audit analysis.

[0061] The maintenance management module monitors system health. Parameters such as hardware resource utilization, software performance indicators, and communication quality are monitored in real time. Maintenance alerts are automatically triggered when potential issues are detected. Preventive maintenance plans are developed based on equipment operating hours, and tasks include sensor calibration, software updates, and hardware inspections. Maintenance records are linked to the digital twin model to facilitate analysis of the impact of equipment degradation on system performance.

[0062] The training system is integrated into the digital twin platform. New employees can use virtual reality devices to operate the digital twin and learn sampling processes and exception handling methods. Training scenarios include typical failure cases and emergency drills, and operational results are recorded in individual competency profiles. The training system is regularly updated to maintain consistency with the actual system. Training effectiveness is evaluated through a combination of simulated operational assessments and theoretical knowledge tests.

[0063] Example 5: The monitoring sub-area division of the liquefied natural gas sampling area adopts an improved spatial clustering algorithm. The entire sampling area is decomposed into multiple continuous units based on the physical layout and fluid characteristics, and each unit corresponds to an independent node in the sequence structure. The node division principle takes into account the spatial proximity of the position, and the adjacent distance threshold is dynamically calculated based on the pipeline diameter to keep the total area of ​​the unit constant. The connection relationship of the node is determined by the parameter migration path in the historical data, and a bidirectional connection edge is established for the node whose migration probability exceeds the preset threshold. Each node is configured with a unique spatial code, and the code value contains information about the area type and functional level. The connection weight between nodes is dynamically adjusted according to real-time data, and the weight update frequency is synchronized with the sampling period.

[0064] The node's feature vector construction utilizes a multi-source information fusion strategy. Temperature measurement point data undergoes adaptive filtering to eliminate ambient temperature interference. Pressure measurement point data distinguishes steady-state components from dynamic fluctuations, and frequency-domain energy distribution is calculated. Component measurement point data is synchronously correlated with gas chromatography analysis results to identify key component concentrations. Feature vector standardization utilizes a sliding window mechanism, with the window length proportional to the process change rate. Data validity for each feature dimension is verified in real time using a redundant sensor cross-comparison mechanism. Feature vector length remains constant, and newly added measurement point data is compressed and fused using principal component analysis.

[0065] The input data structure of the sequence prediction model is organized as a spatiotemporal tensor. The sampling states of consecutive time points are sampled at a fixed frequency, forming an equally spaced time series. Preprocessing of the input data includes filling in missing values, using an algorithm that incorporates the dual constraints of spatial neighborhood and temporal continuity. Label data is generated using a forward-looking time window mechanism, with the target variable consisting of two output types: temperature prediction and abnormal state markers. Multiple time window length settings are available, with a basic window of 10 minutes and auxiliary windows including 5-minute short-term predictions and 30-minute trend analysis. Data augmentation during the model training phase utilizes time warping techniques to generate rate-varying derivative sequences.

[0066] The extraction process of spatiotemporal features implements a hierarchical abstraction mechanism. The underlying feature extractor uses convolution kernels for local pattern recognition, with kernel size adaptively optimized based on node spacing. Edge processing in convolution operations utilizes a virtual node padding strategy, with feature values ​​derived from symmetrically positioned physical nodes. Spatial feature aggregation utilizes a multi-head attention mechanism, with the number of heads dynamically adjusted based on the number of nodes, and each head's attention range constrained by a neighborhood layer limit. Temporal feature extraction is performed in two stages: short-term dependencies are captured using causal convolution, and long-term dependencies are modeled using recurrent neural networks. Spatiotemporal features are fused using a gated crossover mechanism, with fusion weights automatically learned via a parametric network.

[0067] Node embeddings are generated using multiple layers of nonlinear transformations. The embedding space dimension is set as a logarithmic function of the number of nodes to ensure balanced information density. The initial representation of each node consists of three components: the feature vector at the current moment, the historical state at the previous time step, and the interaction information of neighboring nodes. The aggregation function for this interaction information is designed to be adjustable, adaptively switching between mean pooling, maximum pooling, and attention pooling. The embedding vector generation process preserves the complete backpropagation path, supporting end-to-end model training. The geometric structure of the embedding space is orthogonalized to avoid interference from redundant information between dimensions.

[0068] The generation of prediction results implements a multi-task parallel output mechanism. The temperature prediction branch uses a fully connected regression network to output a sequence of temperature values ​​at each future time point. The pressure prediction branch adds a gradient constraint term to optimize prediction smoothness in areas of rapid fluctuation. The abnormal state prediction branch uses an independent feature extraction channel, focusing on transient patterns that deviate from the steady state. Coordination between multiple tasks utilizes a shared underlying network, with independent optimization performed at higher-level network branches. Each prediction task is equipped with a dedicated loss function component: Huber loss for temperature prediction, logarithmic loss for pressure prediction, and focal loss for abnormality prediction.

[0069] The anomaly scoring system implements dynamic threshold management. The construction of the normal change trend template adopts the mixed density estimation method to distinguish the state distribution under different working conditions. The deviation calculation of the prediction results combines shape difference and numerical difference. The shape difference uses the dynamic time-warped distance metric, and the numerical difference uses the standardized Euclidean distance. The synthetic weight of the deviation is automatically adjusted according to seasonal factors, with an emphasis on numerical difference in winter and shape difference in summer. The normalization processing of the anomaly score adopts local rank transformation to eliminate the interference of baseline drift. The marking of potential anomalies sets dual confirmation conditions: the deviation exceeds the local threshold and the trigger rule is met for three consecutive forecast cycles.

[0070] The predictive model's self-correction mechanism implements online incremental updates. The correction source data includes the error sequence between the predicted results and the actual measured values, anomaly verification flags, and a new model for parameter migration paths. Model parameter adjustments are calculated using a constrained optimization algorithm, with constraints preventing excessive parameter drift. The correction effect is verified through a split-stream comparison test: the corrected model predicts a portion of the real-time data stream, while retaining the original model as a reference. The verification cycle is terminated with adaptive end conditions, with the comparison terminating when the error rate decreases or the fluctuation amplitude converges.

[0071] The local sampling enhancement mechanism is designed as a graded response model. The temporary measurement point addition scheme includes three configurations: the basic configuration adds one temperature and one pressure measurement point; the enhanced configuration adds two temperature and two pressure measurement points; and the complete configuration adds a combination of temperature and pressure measurement points. Configuration selection is based on the anomaly score level and node importance coefficient. The location of additional measurement points is determined based on the uncertainty distribution of spatial interpolation, with priority given to areas with the greatest gradient variation. A flexible mechanism is implemented for the service life of temporary measurement points, automatically revoking the additional points when the anomaly score returns to normal levels.

[0072] The operating mechanism of the rolling forecast window implements pipeline management. The forecast task is divided into four stages: data preparation, feature calculation, model inference, and result output. Each stage is executed in parallel to form a processing pipeline. The window movement process uses a time-sliced ​​overlay technique, with the new window containing the overlapping area at the end of the previous window. After each forecast, a model correction operation is performed, with the correction coefficient set as a negative exponential function of the forecast error. The correction operation's impact range is isolated and protected, with only a specified proportion of adjustable parameter layers open. The window switching process is seamless, and the time series of forecast results remains continuous and smooth.

[0073] Dynamic optimization of the sequence structure implements incremental reconstruction. When newly deployed temporary measurement points are operating stably, the system automatically evaluates their information contribution. Measurement points whose contribution exceeds the threshold are upgraded to permanent nodes, assigned new spatial codes, and incorporated into the sequence structure. The adjacency relationships of the newly added nodes are reestablished through data analysis, and the migration path probabilities are recalculated. The dimensionality of node features is expanded using principal component fusion, without changing the length definition of the original feature vector. The evolution history of the sequence structure is fully recorded, supporting historical state retrospective analysis. Node relationships are visualized using a three-dimensional force-directed graph, with spatial position and connection line color used to indicate parameter migration strength.

[0074] Multi-scale analysis technology is used to improve the spatial positioning accuracy of anomaly processing. The coarse positioning phase identifies the regional blocks where the anomalous node resides, with regional division taking into account the characteristics of fluid dynamics. The fine positioning phase deploys a high-density sensor grid around the anomalous node, with grid accuracy exceeding the division scale of the basic monitoring unit. Multi-dimensional verification of positioning results utilizes physical simulation to replicate anomaly patterns in the digital twin model. The output format of spatial positioning information includes node code, relative position coordinates, and the radius of the anomaly's impact range. The automatic positioning accuracy assessment module calculates the overlap between the target area boundary and the actual anomaly area.

[0075] Example 6: The polynomial regression model input feature construction for rule engine conflict detection utilizes a composite expansion strategy. The sampled quality prediction value sequence undergoes moving average filtering, extracting low-frequency trend components as fundamental features. The component stability estimation sequence calculates the coefficient of variation and autocorrelation function values ​​to characterize the fluctuation characteristics. Interactive features are generated through Cartesian product expansion to form polynomial combinations of the original features, with the highest-order terms constrained to a reasonable range. Feature vector dimensionality reduction utilizes a principal component preservation method, retaining components with information exceeding a set threshold. Input data standardization implements an online update mechanism, with statistical parameters dynamically adjusted with new samples.

[0076] The model training process implements a phased optimization strategy. Initial training utilizes historical rule engine execution records, covering a wide range of operating scenarios. The training objective function includes regularization constraints to prevent overfitting caused by higher-order terms. Parameter optimization utilizes an adaptive learning rate algorithm, automatically increasing the learning rate to escape local optima when the loss function plateaus. Training termination criteria are dual: a decrease in the loss value below a threshold or a continuous increase in the validation set error. After training, model parameters are permanently stored, while a continuous learning interface is maintained.

[0077] A robustness mechanism is introduced for calculating the prediction error of risk value labels. The error weights for abnormal samples are dynamically attenuated to reduce the impact of outliers. Truncation is implemented in the calculation of error gradients to prevent numerical explosion during backpropagation. A momentum term is added to the optimization direction of the loss function to accelerate convergence. Confidence assessment of model predictions utilizes an ensemble learning approach, constructing multiple sub-models to vote and output the final result. Predictions with confidence levels below a threshold trigger a manual review process.

[0078] The rule engine implements a multi-level verification mechanism for conflict determination. When the risk value exceeds a predetermined threshold, the system automatically retrieves historical execution records for similar cases. The virtual simulation environment simultaneously recreates the conflict scenario, testing the effectiveness of the rule combination within the digital twin model. A time reversal analysis module verifies the logical consistency of the rule trigger sequence and identifies potential race conditions. Final conflict determination requires positive results from at least two verification channels to prevent false positives from disrupting normal operations.

[0079] Intelligent sampling control strategies are optimized using rule recombination technology. After identifying conflicting rule sets, the system analyzes the overlapping areas of rule preconditions. Rules in these overlapping areas are merged to generate new composite rule entries. The confidence factor of the composite rule is recalculated based on historical data, and the execution action is determined using a weighted voting mechanism. The optimized rule set is topologically sorted to ensure that the execution path is free of circular dependencies. A detailed change log is recorded for each optimization operation, including replaced rules, new rules, and adjusted parameters.

[0080] The digital twin model's continuous adjustment framework builds a meta-learning infrastructure. The meta-learner is designed as a two-layer loop structure: the inner loop is responsible for rapid adaptation to specific tasks, while the outer loop optimizes the model's generalization capabilities. The task construction module divides historical feedback data into independent scenario units, each containing a complete sequence of control actions and environmental response records. The task training and validation sets are divided using time series cross-validation to prevent data leakage.

[0081] The model fine-tuning process implements a parameter space constraint strategy. The inner loop updates are restricted to a subset of model parameters, protecting the core feature extraction layer from excessive modification. The fine-tuning step size sets the annealing schedule, with a larger initial step size for accelerated convergence and a smaller step size later for improved accuracy. Projection processing is implemented in the gradient update direction to ensure compatibility with the main model parameter space. Fine-tuned sub-model parameters are stored in a sparse format, recording only the difference relative to the main model.

[0082] The processing of new tasks is streamlined. Feature alignment and timestamp synchronization are performed immediately upon receipt of feedback data. The data augmentation module generates derivative samples of rotation and translation transformations to improve generalization capabilities in scenarios with small sample sizes. The task builder automatically annotates input and output correspondences to form a standardized training format. The task priority scheduler allocates computing resources based on scenario urgency, prioritizing critical tasks.

[0083] The model validation mechanism implements a distributed testing architecture. The validation process is performed in isolated computing containers to avoid interfering with online systems. Validation metrics include prediction accuracy, response speed, and resource consumption. Accuracy verification utilizes a multi-metric comprehensive evaluation approach to balance the impact of different types of errors. Response speed testing simulates high-load scenarios and measures processing latency under extreme conditions. Resource consumption monitoring records peak CPU, memory, and graphics memory usage. Validation reports are automatically generated, identifying performance bottlenecks and suggesting improvements.

[0084] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0085] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A convenient automatic sampling method for liquefied natural gas, characterized in that: The following steps are involved: Acquiring historical sampling data of liquefied natural gas within a preset sampling period, and generating a historical sampling data set of liquefied natural gas based on the historical sampling data; Extracting sample characteristic parameters from the historical sampling data set to obtain a historical sample characteristic parameter set of liquefied natural gas, and establishing a trend change characteristic matrix based on the temporal trend of the historical sample characteristic parameters; Constructing a multidimensional state space based on the trend change feature matrix, calculating the state aggregation degree of historical liquefied natural gas sampling abnormal events in the multidimensional state space, and determining a sample feature early warning indicator set based on the state aggregation degree; Extracting real-time sample characteristic parameters based on real-time sampling data of liquefied natural gas, establishing a real-time characteristic state vector, and performing spatial position correlation calculation with the sample characteristic warning indicator set in the multidimensional state space to determine a real-time risk association factor; A risk probability prediction model is generated based on the real-time risk association factor, the historical sample characteristic parameter set and the trend change characteristic matrix to predict the probability of abnormal sampling of liquefied natural gas and generate sampling adjustment suggestions.

2. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 1, characterized in that: Acquire historical sampling data of liquefied natural gas within a preset sampling period, and generate a historical sampling data set of liquefied natural gas based on the historical sampling data, specifically: Collecting historical sampling data of liquefied natural gas within the preset sampling period; Perform data validity screening based on the change range, change trend and fluctuation amplitude of each parameter data in the historical sampling data, and remove parameter data that exceeds a preset reasonable range; The historical sampling data after data validity screening are classified and integrated according to the parameter correspondence to generate a historical sampling data set of liquefied natural gas containing temperature parameters, pressure parameters and composition parameters.

3. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 2, characterized in that: The sample characteristic parameters of the historical sampling data set are extracted to obtain a historical sample characteristic parameter set of liquefied natural gas. A trend change characteristic matrix is ​​established according to the temporal change trend of the historical sample characteristic parameters, specifically: Performing historical sample feature parameter extraction processing on the historical sampling data set; Establishing a historical sample feature parameter set according to different dimensions of the historical sample feature parameters; Calculating the change trend values ​​of different historical sample characteristic parameters in the historical sample characteristic parameter set within a specified time period; Arrange the change trend values ​​of the historical sample characteristic parameters in chronological order to form a trend change characteristic matrix; the rows of the trend change characteristic matrix represent different historical sample characteristic parameters, and the columns represent corresponding time series data.

4. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 3, characterized in that: A multidimensional state space is constructed based on the trend change feature matrix, the state aggregation degree of the historical liquefied natural gas sampling abnormal events in the multidimensional state space is calculated, and a sample feature warning indicator set is determined based on the state aggregation degree, specifically: The dimension of the multidimensional state space is determined according to the number of historical sample characteristic parameters in the trend change characteristic matrix; the coordinate axes of the multidimensional state space are different historical sample characteristic parameters of the trend change characteristic matrix, and the characteristic position coordinates in the multidimensional state space are the change trend values ​​of the historical sample characteristic parameters; Calculating the aggregation degree of state points when abnormal sampling of liquefied natural gas occurred in the past based on the coordinates of each characteristic position in the multidimensional state space, and obtaining the state aggregation degree value of the abnormal sampling event of the liquefied natural gas in the multidimensional state space; A warning threshold is set according to the state concentration value, and historical sample feature parameters corresponding to feature position coordinates whose state concentration values ​​exceed the warning threshold are selected as a sample feature warning indicator set.

5. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 4, characterized in that: Real-time sample characteristic parameters are extracted from real-time liquefied natural gas sampling data to establish a real-time characteristic state vector. Spatial position correlation calculation is performed with the sample characteristic warning indicator set in the multidimensional state space to determine the real-time risk association factor, specifically: Real-time collection of liquefied natural gas sampling data under the current sampling status; extracting real-time sample characteristic parameters based on the real-time sampling data, and forming a real-time characteristic state vector according to the parameter sequence of the historical sample characteristic parameter set; The real-time characteristic state vector is mapped to the multidimensional state space, the spatial distance between the real-time characteristic state vector and the characteristic position coordinates in the sample characteristic warning indicator set is calculated, the spatial position correlation is calculated based on the spatial distance, and the real-time risk association factor of the current sampling state of the liquefied natural gas is determined based on the spatial position correlation.

6. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 5, characterized in that: Based on the real-time risk correlation factor, the historical sample characteristic parameter set, and the trend change characteristic matrix, a risk probability prediction model is generated to predict the probability of abnormal sampling of liquefied natural gas and generate sampling adjustment suggestions, specifically: Using the real-time risk correlation factor and the historical sample characteristic parameter set as input data for a risk probability prediction model; Using the risk probability prediction model, the change trend values ​​of the historical sample characteristic parameters in the trend change characteristic matrix are calculated to predict the probability value of sampling anomaly under the current sampling state of liquefied natural gas; Generate liquefied natural gas sampling adjustment recommendations based on the predicted sampling anomaly probability value and the preset sampling adjustment threshold.

7. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 1, characterized in that: The method also includes constructing a digital twin model of the liquefied natural gas sampling process, collecting real-time sampling data of the liquefied natural gas through sensors deployed at the sampling site, and synchronizing the real-time sampling data into the digital twin model; Simulating actual sampling status in the digital twin model to predict sampling development trends and potential sampling anomalies; Set up an intelligent sampling control strategy. The intelligent sampling control strategy combines the degree of mutual influence between real-time sampling data and the execution conditions of the control action priority to determine whether there is a logical conflict in the rule engine. If so, optimize the intelligent sampling control strategy. Specifically: The intelligent sampling control strategy constructs a comprehensive feature vector based on the coupling relationship between the sampling quality prediction value and the component stability estimation value, combines the preset control action priority, and uses the sequence prediction model to predict the risk value of the logical conflict in the rule engine, and compares it with the predetermined threshold to determine whether there is a logical conflict; The sampling operation is automatically performed according to the optimized intelligent sampling control strategy, and the control behavior and feedback data are recorded in real time, and the digital twin model is continuously adjusted based on the recorded feedback data.

8. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 7, characterized in that: The actual sampling status is simulated in the digital twin model to predict the sampling development trend and potential sampling anomalies, including: The entire sampling area is divided into multiple monitoring sub-areas, each of which corresponds to a node in the sequence. The connection relationship between nodes is established based on the spatial adjacency relationship of the sampling positions or the parameter migration path. Each node is associated with multiple sensor data to form the node's feature vector. The initial sequence structure of the sampling area is constructed. The sampling state at each moment is represented as a sequence, and the continuous time forms the sequence data input; the label data is the target variable or sampling abnormal state mark of each node in a certain time window in the future; Use the sequence prediction model to extract the spatiotemporal features of the nodes; the sequence prediction model integrates the sequence at each time step to obtain the embedded representation of the node; The node representation is input into the subsequent time series prediction module; the temperature or pressure variables of each node in the future time step are predicted; the prediction results are scored for anomalies based on the actual historical distribution; if the predicted value of a node deviates from the normal change trend, it is marked as a potential sampling anomaly.

9. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 8, characterized in that: Setting an intelligent sampling control strategy and optimizing the intelligent sampling control strategy further includes: Based on the conversion of the sampling quality prediction value and the component stability estimation into a comprehensive feature vector, the comprehensive feature vector is used as the input of the polynomial regression model. In combination with the currently set control action priority, the polynomial regression model uses each set of comprehensive feature vectors to predict the risk value label of the logical conflict of the rule engine as the prediction target, and takes minimizing the sum of the prediction errors of the risk value labels of the logical conflict of all rule engines as the training target. The polynomial regression model is trained until the sum of the prediction errors reaches convergence, and the model training is stopped. The risk value of the logical conflict of the rule engine is determined based on the model output result; The obtained risk value of the logical conflict in the rule engine is compared with a predetermined threshold. If the risk value of the logical conflict in the rule engine is greater than or equal to the predetermined threshold, it indicates that the rule engine has a logical conflict. At this time, an early warning signal is generated and the intelligent sampling control strategy is optimized. If the risk value of the logical conflict in the rule engine is less than the predetermined threshold, it indicates that the rule engine does not have a logical conflict. At this time, no early warning signal is generated and no additional adjustment is required.

10. The easy-to-operate automatic sampling method for liquefied natural gas according to claim 7, characterized in that: The digital twin model is continuously adjusted based on the recorded feedback data, specifically including: in the initial stage, a digital twin model is constructed to simulate the liquefied natural gas sampling process, and a set of data structures for recording is established to associate the correspondence between each control behavior, environmental response and model output; multiple small sample task sets are constructed using historical feedback data, each task represents a specific sampling scenario, and each task contains a training set and a validation set; the digital twin model is gradually fine-tuned in each task; the initial parameters of the model are optimized through repeated iterations of multiple tasks; in actual operation, after each control behavior is executed, the collected feedback data is immediately constructed as a new task; the new task is input into the trained meta-learning framework for iteration, the current twin model is locally fine-tuned, and the fine-tuned model is retained as the current scenario sub-model; the system regularly verifies the prediction performance of the current twin sub-model; if the error decreases or meets the self-set threshold, the system will extract the fine-tuned parameter changes, feed them back to the main model parameter set, update the main model or add it to the experience model pool.

Images