Deep learning-based liquefied oil product whole-process quality monitoring method

Abstract

The invention discloses a liquefied oil product whole-process quality monitoring method based on deep learning. The method comprises the following steps: S1, collecting and preprocessing parameter data of a liquefied oil product; s2, constructing a differential equation set based on physical constraints, and designing a PINNs network of a dual residual mechanism; s3, a dynamic weighting factor is introduced for adaptive adjustment, a loss function is defined, and a PINNs network is trained; s4, embedding a Shengness differential equation unit in the PINNs network, continuously reasoning an evolution trajectory of the physical state of the liquefied oil product on a time axis, and introducing gradient regularization to perform stability constraint; s5, performing error comparison on the evolution trajectory and the standardized data set, and generating a quality monitoring report; and S6, the quality change trend of the liquefied oil product is judged, and a process adjustment instruction and early warning information output are triggered. According to the method, physical modeling and deep learning are combined, the residual error and stability are optimized, the quality monitoring precision and timeliness are improved, and the traditional error problem is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil product process and quality management, and particularly relates to a liquefied oil product whole-process quality monitoring method based on deep learning. BACKGROUND

[0002] As an important basic material in energy and chemical industries, the liquefied oil product is in a dynamic change in physical state and chemical composition in the whole process of liquefaction, transportation, storage and distribution. A slight carelessness will lead to quality deterioration, energy efficiency decline and even safety accidents. Therefore, the continuous monitoring of the quality of the liquefied oil product is of great significance to ensure the stability of the industrial process, improve economic efficiency and maintain environmental safety. The traditional liquefied oil product quality monitoring mainly relies on manual sampling and laboratory detection, which has a long detection cycle, a lagging response and a limited coverage, and is difficult to meet the real-time monitoring needs of modern large-scale, continuous and automated production modes. In addition, due to the influence of complex thermodynamics, multiphase flow and phase change process, the state evolution of the liquefied oil product is highly nonlinear and frequently changes dynamically. The traditional experience model and simple statistical analysis method have the problems of insufficient precision and poor adaptability in practical application.

[0003] With the development of industrial Internet of Things and intelligent manufacturing, more and more liquefied oil product transportation and storage facilities are equipped with a large number of physical sensors, which can collect multi-dimensional data such as temperature, pressure, flow rate, density and component concentration in real time. However, these data are huge in quantity, complex in dimension and strongly time-correlated. The traditional data analysis technology often cannot effectively extract deep feature relationships, resulting in insufficient sensitivity of anomaly detection and weak trend prediction ability. In recent years, deep learning technology has made significant progress in the field of industrial prediction and state monitoring. Through large-scale data-driven end-to-end learning mechanism, it shows great potential in modeling nonlinear dynamic systems and extracting complex spatio-temporal features. However, the existing deep learning-based methods mostly only aim at data fitting optimization, ignoring the physical laws followed by the liquefied oil product, such as the basic physical laws of momentum conservation, energy conservation and mass diffusion, resulting in the risk of overfitting of the model, and insufficient robustness and poor generalization performance when facing changes in working conditions.

[0004] As a new deep learning framework that combines physical constraints and data-driven, physical information neural network (PINNs) effectively improves the consistency and interpretability of the model by introducing the residual error of the physical differential equation into the loss function. However, traditional PINNs still face multiple challenges in modeling complex multi-physical processes. On the one hand, most existing PINNs use a single residual optimization strategy, which lacks independent modeling and dynamic weighting of data reconstruction error and physical consistency error, making it difficult to achieve a reasonable balance between data fitting and physical constraints. On the other hand, the full process of liquefied oil involves the interaction of liquid and gas phases, multiphase interface heat transfer, and non-equilibrium phase change, and the physical equations are complex and strongly coupled. Existing methods often simplify the modeling of differential equations, making it difficult to truly reflect the evolution of liquefied oil under actual working conditions, affecting the accuracy and reliability of the final quality monitoring.

[0005] Neural ordinary differential equations (Neural ODEs) propose a new paradigm for continuous-time modeling, which can naturally model the continuous state evolution of liquefied oil by parameterizing the differential process of state variables over time, making up for the lack of accuracy in time reasoning of traditional discrete step deep networks. However, the standard neural ordinary differential equation model is prone to gradient explosion or gradient disappearance during training, especially when facing complex dynamic systems, the reasoning stability and training convergence become major bottlenecks. Existing work lacks a systematic stability control mechanism for this problem, often relying on manual parameter tuning or increasing network width and depth to alleviate the problem, resulting in large computational overhead, long training period, and incomplete stability guarantee.

[0006] In the field of liquefied oil quality monitoring, there is currently a lack of a systematic deep learning method that can balance the complexity of multi-physical processes, multi-source data fusion, continuous time evolution reasoning, and training stability. Traditional PINNs methods fail to introduce complete modeling for complex coupled physical processes such as interactive damping, multiphase heat transfer correction, and non-equilibrium phase change, and the residual mechanism is single, making it difficult to dynamically adapt to changes in data features and physical constraint strength at different stages. Existing neural ordinary differential equation applications fail to meet the special dynamic evolution needs of liquefied oil, and do not introduce effective gradient regularization constraints during reasoning, resulting in problems such as trajectory drift and loss of physical consistency in long-time reasoning. In addition, there is currently no deep integration of PINNs and neural ordinary differential equations to form an integrated intelligent quality monitoring framework that can reason, dynamically alert, and guide process optimization throughout the liquefaction, transportation, storage, and distribution processes.

[0007] Therefore, how to provide a deep learning-based liquefied oil full-process quality monitoring method is a problem that those skilled in the art need to solve. SUMMARY

[0008] One purpose of the present application is to provide a liquefied oil product full-process quality monitoring method based on deep learning, which fully utilizes the PINNs network and neural ordinary differential equation technology, combines the dynamic physical properties of liquefied oil products, and designs a new quality monitoring framework. The framework builds a set of differential equations of multiple physical processes including momentum conservation, energy conservation and mass diffusion, introduces a double residual mechanism to optimize data errors and physical consistency, and uses neural ordinary differential equation units to infer the time evolution of liquefied oil products, ensuring real-time monitoring and accurate prediction of liquefied oil product quality in the whole process.

[0009] According to the liquefied oil product full-process quality monitoring method based on deep learning, the method comprises the following steps: S1, collecting parameter data of liquefied oil products and preprocessing to form a standardized data set; S2, constructing a differential equation set based on physical constraints, and designing a PINNs network with a double residual mechanism, wherein the first residual mechanism calculates the data reconstruction residual of the standardized data set, and the second residual mechanism calculates the physical consistency residual of the differential equation set; S3, introducing a dynamic weighting factor to adaptively adjust the double residual weight, defining a loss function based on the double residual mechanism, and training the PINNs network using the dynamically weighted loss function to output the parameter prediction result of the liquefied oil product; S4, embedding a neural ordinary differential equation unit in the PINNs network, taking the parameter prediction result as the initial condition, continuously inferring the evolution trajectory of the physical state of the liquefied oil product on the time axis, and introducing gradient regularization to constrain the stability of the neural ordinary differential equation unit to suppress the gradient explosion and gradient disappearance phenomenon; S5, comparing the evolution trajectory with the standardized data set, calculating the physical state deviation, and generating a liquefied oil product quality monitoring report; S6, based on the liquefied oil product quality monitoring report, determining the quality change trend of the liquefied oil product in the liquefaction, transportation, storage and distribution stages, and triggering process adjustment instructions and early warning information output.

[0010] Optionally, the parameter data includes temperature, pressure, flow rate, density and chemical composition concentration.

[0011] Optionally, the preprocessing includes data cleaning, missing value interpolation, outlier removal and data standardization.

[0012] Optionally, the system of differential equations includes a momentum conservation equation, an energy conservation equation, and a mass diffusion equation, the momentum conservation equation introduces an interaction damping term considering the interaction between the liquid phase and the gas phase, the energy conservation equation introduces a multiphase interface heat transfer correction term, and the mass diffusion equation introduces a non-equilibrium phase change rate correction term, the three types of equations are constructed by sharing state variables and cross-coupling terms to form a unified system of differential equations suitable for the whole process of liquefaction, transportation, storage, and distribution.

[0013] Optionally, S2 specifically includes: S21, based on the standardized data set, a system of differential equations suitable for the whole process of liquefaction, transportation, storage, and distribution is constructed, and interaction damping, multiphase interface heat transfer correction, and non-equilibrium phase change rate correction are introduced, the system of differential equations is composed of a momentum conservation equation, an energy conservation equation, and a mass diffusion equation; S22, based on the system of differential equations, a PINNs network with a double residual mechanism is designed, the double residual mechanism includes a data reconstruction residual and a physical consistency residual: ; ; Wherein, is the first heavy data reconstruction residual, is the number of samples in the standardized data set, is the number of physical quantities in a single sample, is the network prediction value of the first sample, is the network prediction value of the first sample, is the standardized true value of the first physical quantity in the first sample, is the second heavy physical consistency residual, is the number of sampling points of the differential equation physical constraint, , are the residual amounts of the momentum conservation equation, the energy conservation equation, and the mass diffusion equation at the first sampling point .

[0014] Optionally, the momentum conservation equation is expressed as: ; Wherein, is the local density of the liquefied oil product, is the velocity component along the direction, is the velocity component along the direction, is the time, for local pressure, and are spatial coordinates, is a component of viscous stress tensor, is an inter-damping term caused by the interaction between liquid and gas phases, is an inter-damping coefficient, is a liquid velocity vector, is a gas velocity vector.

[0015] Optionally, the energy conservation equation is expressed as: ; where, is the total energy per unit mass, defined as , is the internal energy per unit mass, is the velocity component along direction, is the enthalpy, defined as , is the effective thermal conductivity, is the absolute temperature, is a multiphase interface heat exchange correction term, is the interfacial heat transfer coefficient, is the interfacial area per unit volume, and are the liquid and gas temperatures, respectively.

[0016] Optionally, the mass diffusion equation is expressed as: ; where, is the mass fraction of the th component, is the diffusion flux of the th component in direction, defined as , is the diffusion coefficient of the th component, is the non-equilibrium phase change rate correction term of the th component, is the phase change kinetics constant, is the mass fraction of the th component in equilibrium state.

[0017] Optionally, the S3 specifically includes: S31, introducing a dynamic weighting factor related to time step based on data reconstruction residual and physical consistency residual and Construct an adaptive comprehensive loss function: ; in, For the first The total loss value of the PINNs network. For the first Step data reconstruction residual weighting coefficients, For the first Step-by-step physical consistency residual weighting coefficient This is the gradient regularization weight factor. and They are respectively and Regarding PINNs network parameters The gradient vector; S32. Define a residual dynamic adjustment function that adaptively updates the weighting factors based on the trend of residual size changes in the current training step: ; in, For the first The ratio of the step data reconstruction residual to the physical consistency residual. To prevent small positive numbers with a denominator of zero from being converted to a fixed value, the value is set to 0. , and The updated weighting factor. To adjust the rate control constant for weighting, The equilibrium point for the expected residual ratio. This represents the natural exponential function with base e; S33. A dual-branch separation structure is introduced into the backbone of the PINNs network. The first branch outputs the data reconstruction prediction and the second branch outputs the physical consistency prediction, which correspond to the data reconstruction residual and the physical consistency residual, respectively. A differential consistency supervision unit is added after each hidden layer. The first and second spatial derivatives of the prediction function are forcibly constrained by the local derivative matching technique to enhance the modeling ability of the continuity and smoothness of the state variables of liquefied oil products. S34. Based on the updated weighting factors, jointly synthesize the loss function. The PINNs network training process employs an adaptive momentum optimizer to optimize network parameters. Iterative updates are performed, and the update rules are defined as follows: ; ; in, For the first Step network parameter vector, For learning rate, For the current step, the adaptive momentum direction, The momentum decay factor, This represents the gradient of the total loss function with respect to the parameters; S35. Continuously iteratively optimize until the comprehensive loss function is reached. Once convergence or the number of training rounds reaches the preset upper limit, the final output will be the parameter prediction results for liquefied oil products at each node of the entire process.

[0018] Optionally, S4 specifically includes: S41. Based on the parameter prediction results, set the initial conditions for the physical state of the liquefied oil. ,in, For liquefied oil products in time The predicted density, For liquefied oil products in time along Predicted values ​​of velocity components in the direction. For liquefied oil products in time The predicted temperature For liquefied oil products Components in time The predicted quality score; S42. Embedding a neural ordinary differential equation unit in the PINNs network, defining a continuous inference function for the evolution of the physical state of liquefied oil products over time, wherein the inference process satisfies the following neural ordinary differential equation: ; in, For time Predicted physical state vector of liquefied oil at any given time. This represents the dynamic evolution function in continuous time. Let be the set of trainable parameters in the unit of the constant differential equation, with the initial conditions being: ; S43. To suppress gradient explosion and gradient vanishing phenomena, a gradient regularization mechanism is introduced to constrain the stability of the units of the neural frequent differential equation. The gradient regularization term is defined as follows: ; in, For the gradient regularization loss of the units of the God-like ordinary differential equations This is the global weighting factor for gradient regularization. The end of the reasoning time interval, Let be the vector of the second derivative of the physical state of liquefied oil with respect to time. It is a vector of first-order derivatives. The first derivative regularization of the relative weight coefficients; S44. Regularize the gradient loss. with the comprehensive loss function superimposed, forming the total training loss after embedding the neural ordinary differential equations, and serving as the objective function, jointly optimizing the PINNs network parameters with the neural ordinary differential equation unit parameters , until the total training loss converges.

[0019] The beneficial effects of the present application are: Firstly, the liquefied oil product whole-process quality monitoring method based on deep learning of the present application solves the problem of the split between physical modeling and data-driven analysis in traditional methods by combining PINNs network and neural ordinary differential equations. By combining the physical constraints and differential equations of liquefied oil products with the training process of deep learning network, the quality prediction model can be automatically extracted and optimized from large-scale monitoring data while ensuring the physical consistency of the model. This method significantly improves the accuracy of liquefied oil product quality monitoring and overcomes the limitations of traditional experience-based and statistical analysis techniques in dynamic complex environments.

[0020] Secondly, the present application introduces a double residual mechanism and a dynamic weighting strategy, which flexibly adjusts the weights of data reconstruction error and physical consistency residual under different working conditions for the complex dynamic behavior of liquefied oil products. This innovative design can adaptively adjust the residual weight, so that the model can more effectively balance the relationship between data-driven and physical constraints in real-time monitoring, thereby enhancing the adaptability and accuracy of the model to quality changes of different stages and different types of oil products. This feature significantly improves the ability to run stably in complex environments for a long time, making quality prediction more reliable and faster to adapt to various abnormal changes.

[0021] Finally, the present application uses neural ordinary differential equation units to perform continuous time reasoning on the quality evolution process of liquefied oil products, so that the model can efficiently handle the dynamic quality changes of liquefied oil products, avoiding the precision loss of traditional discrete time step methods. In addition, by introducing gradient regularization mechanism, the gradient explosion and gradient disappearance problems in the neural network training process are effectively suppressed, ensuring the stability and convergence of the model training. This technical innovation ensures that the model can maintain good performance in long-time sequence prediction, providing more accurate quality monitoring and trend prediction, and reducing the production safety risks caused by quality fluctuations. BRIEF DESCRIPTION OF DRAWINGS

[0022] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings: Fig. 1 A flowchart of a liquefied oil product whole-process quality monitoring method based on deep learning according to the present application is shown; Fig. 2 A differential equation set construction and PINNs network design schematic diagram of a liquefied oil product whole-process quality monitoring method based on deep learning is proposed for the present application. Fig. 3 An evolution trajectory prediction and quality monitoring report generation schematic diagram of a liquefied oil product whole-process quality monitoring method based on deep learning is proposed for the present application. DETAILED DESCRIPTION

[0023] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0024] REFERENCE Figs. 1-3 A liquefied oil product whole-process quality monitoring method based on deep learning, comprising the following steps: S1, collecting parameter data of liquefied oil products and performing preprocessing to form a standardized data set; S2, constructing a differential equation set based on physical constraints, and designing a PINNs network with a double residual mechanism, wherein the first double residual mechanism performs data reconstruction residual calculation on the standardized data set, and the second double residual mechanism performs physical consistency residual calculation on the differential equation set; S3, introducing a dynamic weighting factor to adaptively adjust the double residual weight, defining a loss function based on the double residual mechanism, and training the PINNs network using the dynamically weighted loss function to output the parameter prediction result of the liquefied oil product; S4, embedding a neural ordinary differential equation unit in the PINNs network, taking the parameter prediction result as the initial condition, continuously inferring the evolution trajectory of the physical state of the liquefied oil product on the time axis, and introducing gradient regularization to constrain the stability of the neural ordinary differential equation unit to suppress the gradient explosion and gradient disappearance phenomenon; S5, comparing the evolution trajectory with the standardized data set, calculating the physical state deviation, and generating a liquefied oil product quality monitoring report; S6, based on the liquefied oil product quality monitoring report, determining the quality change trend of the liquefied oil product at each stage of liquefaction, transportation, storage and distribution, and triggering process adjustment instructions and early warning information output.

[0025] The present application collects multi-dimensional parameter data of liquefied oil products and performs standardization processing to generate a standardized data set, providing reliable input data for subsequent quality prediction. This method introduces a differential equation set based on physical constraints and a PINNs network with a double residual mechanism, combining the double optimization strategy of data reconstruction residual and physical consistency residual, effectively improving the accuracy and stability of liquefied oil product quality prediction, overcoming the limitations of traditional methods that cannot simultaneously satisfy data-driven and physical model consistency.

[0026] In this embodiment, the parameter data includes temperature, pressure, flow rate, density and chemical component concentration.

[0027] The present application clearly indicates the key parameter data required for liquefied oil quality monitoring, including temperature, pressure, flow rate, density and chemical component concentration, etc., providing accurate guidance for subsequent data collection and preprocessing, ensuring multi-dimensional coverage and comprehensiveness of quality monitoring. This approach not only enhances the representativeness and accuracy of the data, but also helps to improve the adaptability of the model under different physical conditions.

[0028] In this embodiment, the preprocessing includes data cleaning, missing value interpolation, outlier removal and data standardization.

[0029] The present application further specifies the preprocessing steps, including data cleaning, missing value interpolation, outlier removal and data standardization, etc., ensuring the quality of the data set, so that the deep learning model can better model the quality characteristics of liquefied oil. This preprocessing step ensures the accuracy and consistency of the data, reducing the negative impact of noise on model training and prediction results.

[0030] In this embodiment, the system of differential equations includes momentum conservation equations, energy conservation equations and mass diffusion equations, the momentum conservation equations introduce interaction damping terms considering the interaction between liquid and gas phases, the energy conservation equations introduce multiphase interface heat transfer correction terms, and the mass diffusion equations introduce non-equilibrium phase transition rate correction terms. The three types of equations are constructed by sharing state variables and cross-coupling terms, forming a unified system of differential equations suitable for the whole process of liquefaction, transportation, storage and distribution.

[0031] By constructing a system of differential equations including momentum conservation equations, energy conservation equations and mass diffusion equations, and combining interaction damping, multiphase interface heat transfer correction and non-equilibrium phase transition rate correction in the physical process of liquefied oil, the quality monitoring of liquefied oil is not only limited to data fitting, but also fully utilizes physical laws for constraint. This physical constraint mechanism effectively improves the stability and precision of the model, making the whole process quality monitoring of liquefied oil more reliable and scientific.

[0032] In this embodiment, S2 specifically includes: S21, based on the standardized data set, a system of differential equations suitable for the whole process of liquefaction, transportation, storage and distribution is constructed, and interaction damping, multiphase interface heat transfer correction and non-equilibrium phase transition rate correction are introduced, the system of differential equations is composed of momentum conservation equations, energy conservation equations and mass diffusion equations; S22. Based on the system of differential equations, design a PINNs network with a dual residual mechanism, wherein the dual residual mechanism includes data reconstruction residual and physical consistency residual: ; ; in, To reconstruct residuals from the first layer of data, To standardize the number of samples in the dataset, This refers to the number of different types of physical quantities in a single sample. For the first The first sample Network prediction values ​​for each physical quantity For the first The first sample The standardized true value of a physical quantity This is the second level of physical consistency residual. The number of sampling points for the physical constraints of the differential equation. , , The momentum conservation equation, energy conservation equation, and mass diffusion equation are respectively in the 1st, 2nd, 3rd, 4th, 5th, 6th, 7th, 8th, 9th, 10th, 11th, 12th, 13th, 14th, 15th, 16th, 17th, 18th, 19 ... sampling points The residual at the location.

[0033] This invention designs a PINNs network with a dual residual mechanism based on a system of differential equations. This network achieves more accurate quality monitoring by calculating the reconstruction residuals of the standardized dataset and the physical consistency residuals. The dual residual mechanism allows the model to simultaneously consider data fitting and physical consistency during the optimization process, thereby effectively reducing the risk of overfitting and improving prediction accuracy under different operating conditions.

[0034] In this embodiment, the momentum conservation equation is expressed as: ; in, This refers to the local density of liquefied oil. For along The velocity component in the direction, For along The velocity component in the direction, For time, For local pressure, and These are spatial coordinates. These are components of the viscous stress tensor. This is the interactive damping term caused by the interaction between the liquid and gas phases. The interactive damping coefficient, For liquid phase velocity vector, Gas phase velocity vector.

[0035] The application introduces the interaction damping term of liquid phase and gas phase interaction in the momentum conservation equation, enhances the momentum transfer ability of liquefied oil product in multiphase flow, improves the prediction accuracy of flow state under complex working conditions, and optimizes the quality monitoring effect of liquefied oil product.

[0036] In the embodiment, the energy conservation equation is expressed as: ; Wherein, Total energy per unit mass, defined as , Internal energy per unit mass, Velocity component in the direction of , Enthalpy, defined as , Effective thermal conductivity, Absolute temperature, Multiphase interface heat exchange correction term, Phase interface heat transfer coefficient, Unit volume phase interface area, And Liquid phase and gas phase temperature respectively.

[0037] The application can more accurately simulate the heat transfer process of liquefied oil product under different conditions by introducing the multiphase interface heat transfer correction term, improve the prediction accuracy of temperature change, provide scientific basis for temperature regulation and energy saving optimization, and enhance the prediction ability under extreme working conditions.

[0038] In the embodiment, the mass diffusion equation is expressed as: ; Wherein, Mass fraction of the first Component, Diffusion flux of the first Component in the direction of , defined as , Diffusion coefficient of the first Component, Non-equilibrium phase change rate correction term of the first Component, Phase change kinetics constant, Mass fraction of the first Component in equilibrium state.

[0039] This invention introduces a non-equilibrium phase change rate correction term, which effectively improves the mass transfer modeling of liquefied oils during the phase change process, enhances the prediction accuracy of the phase change behavior and diffusion characteristics of liquefied oils, and strengthens the ability to prevent mass loss and the adaptability of the system.

[0040] In this embodiment, S3 specifically includes: S31. Reconstructing Residuals Based on Data Physical consistency residuals Introducing a dynamic weighting factor related to the time step and Construct an adaptive comprehensive loss function: ; in, For the first The total loss value of the PINNs network. For the first Step data reconstruction residual weighting coefficients, For the first Step-by-step physical consistency residual weighting coefficient This is the gradient regularization weight factor. and They are respectively and Regarding PINNs network parameters The gradient vector; S32. Define a residual dynamic adjustment function that adaptively updates the weighting factors based on the trend of residual size changes in the current training step: ; in, For the first The ratio of the step data reconstruction residual to the physical consistency residual. To prevent small positive numbers with a denominator of zero from being converted to a fixed value, the value is set to 0. , and The updated weighting factor. To adjust the rate control constant for weighting, The equilibrium point for the expected residual ratio. This represents the natural exponential function with base e; S33. A dual-branch separation structure is introduced into the backbone of the PINNs network. The first branch outputs the data reconstruction prediction and the second branch outputs the physical consistency prediction, which correspond to the data reconstruction residual and the physical consistency residual, respectively. A differential consistency supervision unit is added after each hidden layer. The first and second spatial derivatives of the prediction function are forcibly constrained by the local derivative matching technique to enhance the modeling ability of the continuity and smoothness of the state variables of liquefied oil products. S34. Based on the updated weighting factors, jointly synthesize the loss function. The PINNs network training process employs an adaptive momentum optimizer to optimize network parameters. Iterative updates are performed, and the update rules are defined as follows: ; ; in, For the first Step network parameter vector, For learning rate, For the current step, the adaptive momentum direction, The momentum decay factor, This represents the gradient of the total loss function with respect to the parameters; S35. Continuously iteratively optimize until the comprehensive loss function is reached. Once convergence or the number of training rounds reaches the preset upper limit, the final output will be the parameter prediction results for liquefied oil products at each node of the entire process.

[0041] This invention introduces a dynamic weighting factor to adaptively adjust the weights of the dual residuals, ensuring that the balance between the data reconstruction residuals and the physical consistency residuals can be dynamically optimized according to the training situation at different training stages. This adaptive adjustment strategy greatly improves the flexibility of the training process, enabling the model to adapt to different optimization objectives at different stages, thereby further improving prediction accuracy and the model's generalization ability.

[0042] In this embodiment, S4 specifically includes: S41. Based on the parameter prediction results, set the initial conditions for the physical state of the liquefied oil. ,in, For liquefied oil products in time The predicted density, For liquefied oil products in time along Predicted values ​​of velocity components in the direction. For liquefied oil products in time The predicted temperature For liquefied oil products Components in time The predicted quality score; S42. Embedding a neural ordinary differential equation unit in the PINNs network, defining a continuous inference function for the evolution of the physical state of liquefied oil products over time, wherein the inference process satisfies the following neural ordinary differential equation: ; in, For time Predicted physical state vector of liquefied oil at any given time. denotes a dynamic evolution function in continuous time, is a set of trainable parameters in the neural ordinary differential equation unit, and the initial value condition is ; S43, to suppress the gradient explosion and gradient disappearance phenomenon, a gradient regularization mechanism is introduced to constrain the stability of the neural ordinary differential equation unit, and the gradient regularization term is defined as: ; wherein, is the gradient regularization loss of the neural ordinary differential equation unit, is a gradient regularization overall weighting factor, is the end point of the inference time interval, is a second derivative vector of the physical state of the liquefied oil product with respect to time, is a first derivative vector, is a first derivative regularization relative weight coefficient; S44, the gradient regularization loss is superimposed with the comprehensive loss function to form the total training loss after embedding the neural ordinary differential equation, and is used as the objective function to jointly optimize the PINNs network parameters and the neural ordinary differential equation unit parameters until the total training loss converges.

[0043] The present application embeds a neural ordinary differential equation unit in a PINNs network, uses continuous time reasoning to evolve and reason the physical state of liquefied oil products, so that the quality change of liquefied oil products in the whole process can be accurately simulated and predicted. The stability of the neural ordinary differential equation unit is constrained by the gradient regularization mechanism to avoid the gradient explosion and gradient disappearance phenomenon, thereby ensuring the stability and accuracy of the model in long time sequence prediction.

[0044] Embodiment 1: In order to verify the feasibility of the present application in implementation, the present application is applied to the quality monitoring of liquefied oil products in the storage and transportation process in a certain region, especially in the quality monitoring process of liquefied natural gas (LNG) storage tank. The process involves the flow of liquefied oil products from the storage tank to the transportation pipeline, and the key factors include temperature, pressure, flow rate and component concentration. Because the flow characteristics of liquefied oil products are very complex, involving multiphase flow, phase change process and heat conduction, the traditional quality monitoring method often has a large time delay, and it is difficult to reflect the quality change of the oil product in real time.

[0045] In this scenario, the present application builds a quality monitoring system based on deep learning, using multi-dimensional physical data of liquefied oil products for real-time prediction and analysis. First, the system collects real-time data such as temperature, pressure, flow rate, density, chemical composition concentration, etc. from the storage tank, transportation pipeline and related facilities of liquefied oil products. After preprocessing, these data form a standardized data set, which enters the PINNs network for data reconstruction and physical consistency optimization. By introducing a double residual mechanism, the model considers both the accuracy of the data and the physical constraints, ensuring the efficiency of liquefied oil product quality monitoring.

[0046] In this process, first of all, a system of differential equations based on the momentum conservation equation, the energy conservation equation and the mass diffusion equation is applied, which combines these physical models with deep learning models. Through this model, we can predict the state changes of liquefied oil products at different time points, ensuring the stability of their quality during the entire transportation and storage process. In each prediction step, a neural ordinary differential equation unit is embedded in the network for continuous time reasoning, and a gradient regularization mechanism is combined to ensure the stability of the training process.

[0047] To verify the effectiveness of this method, we selected a representative liquefied oil product monitoring process for a 30-day experiment. The experimental scenario includes the storage, transportation and distribution of liquefied oil products, focusing on monitoring the changes in physical parameters such as temperature, pressure and composition concentration of liquefied oil products at different time periods. In the experiment, the main quality indicators of liquefied oil products include gas density, temperature and composition concentration, which have an important impact on the quality of oil products during liquefaction and transportation.

[0048] Table 1 Comparison of experimental results of the present application method and traditional method ; As shown in Table 1, during the monitoring process, the main quality parameters of liquefied oil products, such as gas density, temperature and composition concentration, gradually change over time. By comparing with the traditional method, the prediction error of the present application is significantly reduced, especially in the prediction error of important parameters such as temperature and gas density in the storage tank, which is reduced by nearly 5 times. This shows that the deep learning-based liquefied oil product whole-process quality monitoring method can significantly improve the prediction accuracy and reduce the error, thereby more effectively ensuring the quality stability of liquefied oil products in the whole process.

[0049] In addition, the application can also discover potential quality change trend in real time in the monitoring process, and give early warning automatically through computer automation. When an important quality index (such as temperature or gas density) of the liquefied oil product appears abnormally, the system will automatically generate an early warning report and prompt the relevant personnel to take timely measures for adjustment. Through the implementation of the application, the quality fluctuation in the transportation and storage process of the liquefied oil product is effectively controlled, and the safety hazards and economic losses caused by the oil quality problem are effectively avoided.

[0050] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited to this. Any person skilled in the art can make equivalent replacement or change according to the technical scheme and the inventive concept of the application within the technical range disclosed by the application, which should be covered in the protection scope of the application.

Claims

1. A method for full-process quality monitoring of liquefied oil products based on deep learning, characterized in that, Includes the following steps: S1. Collect parameter data of liquefied oil products and preprocess them to form a standardized dataset; S2. Construct a system of differential equations based on physical constraints and design a PINNs network with a dual residual mechanism. The first residual mechanism performs data reconstruction residual calculation on the standardized dataset, and the second residual mechanism performs physical consistency residual calculation on the system of differential equations. S3. Introduce a dynamic weighting factor to adaptively adjust the weights of the dual residuals, define a loss function based on the dual residual mechanism, and train the PINNs network using the dynamically weighted loss function to output the parameter prediction results of liquefied oil products. S4. Embed the neural ordinary differential equation unit in the PINNs network, use the parameter prediction results as the initial conditions, continuously infer the evolution trajectory of the physical state of liquefied oil on the time axis, and introduce gradient regularization to constrain the stability of the neural ordinary differential equation unit to suppress gradient explosion and gradient vanishing phenomena. S5. Compare the evolution trajectory with the standardized dataset to calculate the physical state deviation and generate a liquefied oil quality monitoring report. S6. Based on the liquefied oil quality monitoring report, determine the quality change trend of liquefied oil in each stage of liquefaction, transportation, storage and distribution, and trigger process adjustment instructions and early warning information output.

2. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 1, characterized in that, The parameter data includes temperature, pressure, flow rate, density, and chemical composition concentration.

3. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 1, characterized in that, The preprocessing includes data cleaning, missing value imputation, outlier removal, and data standardization.

4. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 1, characterized in that, The differential equation set includes momentum conservation equation, energy conservation equation, and mass diffusion equation. The momentum conservation equation introduces an interactive damping term that considers the interaction between the liquid and gas phases. The energy conservation equation introduces a multiphase interface heat transfer correction term. The mass diffusion equation introduces a non-equilibrium phase change rate correction term. The three types of equations are constructed by sharing state variables and cross-coupling terms to form a unified differential equation set expression applicable to the entire process of liquefaction, transportation, storage, and distribution.

5. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Based on a standardized dataset, construct a set of differential equations applicable to the entire process of liquefaction, transportation, storage and distribution, and introduce interactive damping, multiphase interface heat transfer correction and non-equilibrium phase change rate correction. The set of differential equations consists of the momentum conservation equation, energy conservation equation and mass diffusion equation. S22. Based on the system of differential equations, design a PINNs network with a dual residual mechanism, wherein the dual residual mechanism includes data reconstruction residual and physical consistency residual: ； ； in, To reconstruct residuals from the first layer of data, To standardize the number of samples in the dataset, This refers to the number of different types of physical quantities in a single sample. For the first The first sample Network prediction values ​​for each physical quantity For the first The first sample The standardized true value of a physical quantity This is the second level of physical consistency residual. The number of sampling points for the physical constraints of the differential equation. , , The momentum conservation equation, energy conservation equation, and mass diffusion equation are respectively in the 1st, 2nd, 3rd, 4th, 5th, 6th, 7th, 8th, 9th, 10th, 11th, 12th, 13th, 14th, 15th, 16th, 17th, 18th, 19 ... sampling points The residual at the location.

6. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 5, characterized in that, The momentum conservation equation is expressed as: ； in, This refers to the local density of liquefied oil. For along The velocity component in the direction, For along The velocity component in the direction, For time, For local pressure, and These are spatial coordinates. These are components of the viscous stress tensor. This is the interactive damping term caused by the interaction between the liquid and gas phases. The interactive damping coefficient, For liquid phase velocity vector, This is the gas phase velocity vector.

7. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 5, characterized in that, The energy conservation equation is expressed as: ； in, The total energy per unit mass is defined as... , Internal energy per unit mass For along The velocity component in the direction, Enthalpy is defined as follows: , For effective thermal conductivity, Absolute temperature This is a correction term for heat exchange at multiphase interfaces. The heat transfer coefficient at the phase interface, The phase interface area per unit volume. and These are the liquid phase and gas phase temperatures, respectively.

8. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 5, characterized in that, The mass diffusion equation is expressed as: ； in, For the first Mass fraction of components For the first Components in The diffusion flux in the direction is defined as , For the first The diffusion coefficient of the component, For the first The non-equilibrium phase transition rate correction term for the component, The phase transition kinetic constant is The first in equilibrium state Component mass fraction.

9. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 1, characterized in that, S3 specifically includes: S31. Reconstructing Residuals Based on Data Physical consistency residuals Introducing a dynamic weighting factor related to the time step and Construct an adaptive comprehensive loss function: ； in, For the first The total loss value of the PINNs network. For the first Step data reconstruction residual weighting coefficients, For the first Step-by-step physical consistency residual weighting coefficient This is the gradient regularization weight factor. and They are respectively and Regarding PINNs network parameters The gradient vector; S32. Define a residual dynamic adjustment function that adaptively updates the weighting factors based on the trend of residual size changes in the current training step: ； in, For the first The ratio of the step data reconstruction residual to the physical consistency residual. To prevent small positive numbers with a denominator of zero from being converted to a fixed value, the value is set to 0. , and The updated weighting factor. To adjust the rate control constant for weighting, The equilibrium point for the expected residual ratio. This represents the natural exponential function with base e; S33. A dual-branch separation structure is introduced into the backbone of the PINNs network. The first branch outputs the data reconstruction prediction and the second branch outputs the physical consistency prediction, which correspond to the data reconstruction residual and the physical consistency residual, respectively. A differential consistency supervision unit is added after each hidden layer. The first and second spatial derivatives of the prediction function are forcibly constrained by the local derivative matching technique to enhance the modeling ability of the continuity and smoothness of the state variables of liquefied oil products. S34. Based on the updated weighting factors, jointly synthesize the loss function. The PINNs network training process employs an adaptive momentum optimizer to optimize network parameters. Iterative updates are performed, and the update rules are defined as follows: ； ； in, For the first Step network parameter vector, For learning rate, For the current step, the adaptive momentum direction, The momentum decay factor, This represents the gradient of the total loss function with respect to the parameters; S35. Continuously iteratively optimize until the comprehensive loss function is reached. Once convergence or the number of training rounds reaches the preset upper limit, the final output will be the parameter prediction results for liquefied oil products at each node of the entire process.

10. The method for full-process quality monitoring of liquefied oil products based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Based on the parameter prediction results, set the initial conditions for the physical state of the liquefied oil. ,in, For liquefied oil products in time The predicted density, For liquefied oil products in time along Predicted values ​​of velocity components in the direction. For liquefied oil products in time The predicted temperature For liquefied oil products Components in time The predicted quality score; S42. Embedding a neural ordinary differential equation unit in the PINNs network, defining a continuous inference function for the evolution of the physical state of liquefied oil products over time, wherein the inference process satisfies the following neural ordinary differential equation: ； in, For time Predicted physical state vector of liquefied oil at any given time. This represents the dynamic evolution function in continuous time. Let be the set of trainable parameters in the unit of the constant differential equation, with the initial conditions being: ; S43. To suppress gradient explosion and gradient vanishing phenomena, a gradient regularization mechanism is introduced to constrain the stability of the units of the neural frequent differential equation. The gradient regularization term is defined as follows: ； in, For the gradient regularization loss of the units of the God-like ordinary differential equations This is the global weighting factor for gradient regularization. The end of the reasoning time interval, Let be the vector of the second derivative of the physical state of liquefied oil with respect to time. It is a vector of first-order derivatives. The first derivative regularization of the relative weight coefficients; S44. Regularize the gradient loss. With the comprehensive loss function The summation forms the total training loss after embedding the frequent differential equations of the PINNs network, which is then used as the objective function to jointly optimize the PINNs network parameters. Unit parameters of the constant differential equation of God Until the total training loss converges.

Images