Petrochemical pipe gallery intelligent monitoring system based on internet of things

Abstract

This invention discloses an intelligent monitoring system for petrochemical pipeline corridors based on the Internet of Things (IoT), specifically relating to the field of petrochemical pipeline corridor control technology. The system involves deploying a sensor array in the pipeline corridor to collect data, which is then processed through noise suppression and spatiotemporal alignment to obtain a standardized time-series data stream. The system extracts the spatiotemporal P-matrix of the pressure field and performs singular value decomposition to obtain the dominant mode basis functions. A physical information neural network is trained using full-order flow field data generated through simulation and the standardized time-series data stream to obtain a transient pressure wave propagation prediction model. Based on the pressure wave prediction results, a state vector including pressure gradient, valve wear, and energy efficiency ratio is defined. A distributed strategy library is trained through a priority experience replay mechanism to generate a Pareto optimal control instruction set. A dual closed-loop control mechanism is used to achieve precise adjustment of valve actions and improve disturbance rejection capability, solving the prediction distortion problem caused by neglecting the transient pressure wave propagation characteristics in traditional methods.

Description

Technical Field

[0001] This invention relates to the field of petrochemical pipeline control technology, and more specifically, to an intelligent monitoring system for petrochemical pipelines based on the Internet of Things. Background Technology

[0002] In the current operation mode of petrochemical pipeline corridors, the pipelines are responsible for transporting high-temperature and high-pressure mixed media (such as oil and gas, chemicals, etc.). The petrochemical pipeline corridor monitoring system monitors various operating parameters in real time by deploying multiple sensors such as temperature, pressure, medium viscosity, and pipe wall vibration. The pipelines are usually designed with emergency shut-off valves to deal with sudden accidents, such as equipment failure or external damage. Once an abnormal signal is detected, the shut-off valve will close quickly to isolate the accident area and prevent the accident from spreading.

[0003] However, due to the high temperature and pressure of the fluid inside the pipeline, the physical properties of the medium and the elastic modulus of the pipe wall dynamically change due to temperature fluctuations, flow rate variations, and long-term aging. When the emergency shut-off valve closes rapidly, the fluid inside the pipe continues to move due to inertia, causing the fluid's kinetic energy to be instantly converted into pressure energy, thus forming a violent transient pressure wave (i.e., the Joukowsky shock). As this shock wave propagates in the pipeline, its velocity and waveform are affected by the viscosity of the medium and the elastic modulus of the pipe wall, which may lead to abnormally high pressure in local areas, threatening the integrity of the pipeline structure and potentially causing more serious safety accidents such as secondary leaks.

[0004] Currently, traditional technologies for addressing the transient hydraulic shock problem after emergency shut-off valve activation mainly employ a fixed pressure threshold to trigger a slow-release mechanism. This means that when the pressure inside the pipe exceeds a preset value, the slow-release device is activated to reduce the shock amplitude. However, this method has the following drawbacks:

[0005] Insufficient response sensitivity: Fixed thresholds cannot adapt to the dynamic changes in medium viscosity and pipe wall elastic modulus under operating conditions. They may be too sensitive in some cases, frequently triggering slow-release measures, while being insufficient in response to real dangers, increasing the risk of misjudgment and missed judgment.

[0006] Coupling effect neglected: Traditional methods fail to fully consider the complex coupling between fluid and structure, making it difficult to accurately capture the propagation characteristics of transient pressure waves, thus limiting the accuracy and real-time performance of the slow release strategy.

[0007] Lack of adaptive control: A single fixed threshold strategy cannot balance safety and economy under varying operating conditions, making it difficult to adjust local pressure anomalies in a timely manner, further exacerbating the risk of pipeline damage. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of existing technologies, this invention provides an IoT-based intelligent monitoring system for petrochemical pipeline corridors. The system collects data from a sensor array deployed in the pipeline corridor, which is then processed through noise suppression and spatiotemporal alignment to obtain a standardized time-series data stream. The P-matrix of the pressure field in the spatiotemporal region is extracted and singular value decomposition is performed to obtain the dominant mode basis functions. An LSTM-PINN physical information neural network is trained using simulated full-order flow field data and the standardized time-series data stream to obtain a transient pressure wave propagation prediction model. Based on the pressure wave prediction results, a state vector including pressure gradient, valve wear, and energy efficiency ratio is defined. The NSGA-II algorithm is used to optimize the weights of the multi-objective reward function, and a distributed policy library is trained through a priority experience replay mechanism to generate a Pareto optimal control instruction set for valve control in the petrochemical pipeline corridor, thereby solving the problems mentioned in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent monitoring system for petrochemical pipeline corridors based on the Internet of Things, comprising:

[0010] The petrochemical pipeline corridor data acquisition module, by deploying a sensor array in the petrochemical pipeline corridor, including at least pressure sensors, ultrasonic flow meters and temperature sensors, performs dynamic noise suppression and spatiotemporal alignment processing on multimodal sensor data to obtain a standardized time-series data stream;

[0011] In the implementation of the petrochemical pipeline corridor data acquisition module, a condition-adaptive Kalman filter is adopted to dynamically adjust the process noise covariance matrix according to the Reynolds number Re; distributed sensor node clock synchronization is achieved based on the IEEE1588PTP protocol, with a synchronization accuracy of ≤1μs, eliminating phase difference; outliers are filtered using the 3σ criterion to ensure data validity.

[0012] The pressure wave propagation prediction module is used to build a transient pressure wave propagation prediction model. It extracts the P matrix of the pressure field from the standardized time-series data stream, performs singular value decomposition on the P matrix using eigenorthogonal decomposition, and obtains the dominant mode basis functions. The LSTM-PINN physical information neural network is trained using full-order flow field data generated by simulation and the standardized time-series data stream to obtain the transient pressure wave propagation prediction model.

[0013] The valve control strategy generation module defines a state vector containing pressure gradient, valve wear, and energy efficiency ratio based on the pressure wave prediction results of the transient pressure wave propagation prediction model. It uses the NSGA-II algorithm to optimize the weights of the multi-objective reward function and trains a distributed Q-learning policy library through a priority experience replay mechanism to generate a Pareto optimal control instruction set.

[0014] In the implementation of the valve control strategy generation module, the synergy between NSGA-II and Q-learning is reflected in the following aspects: a hierarchical optimization framework is adopted, NSGA-II optimizes the reward function weights (Pareto front search), and Q-learning generates a strategy library based on the optimized weights; the strategy generation frequency is matched with the valve control delay, a sliding window is used to compress the size of the strategy library, and FPGA pipeline acceleration is combined.

[0015] The valve control and regulation module deploys the Pareto optimal control instruction set to FPGA hardware, and solidifies the dominant mode basis function and action value table query logic. Based on standardized time-series data flow, it achieves precise regulation of valve action and anti-disturbance capability through a dual closed-loop control mechanism including inner loop PID regulation and outer loop pressure feedback interpolation, and outputs valve action signal.

[0016] Preferably, the pressure gradient acquisition method is as follows: based on the predicted pressure field output by the transient pressure wave propagation prediction model, the prediction result includes the pressure distribution of the pipeline spatial coordinates at the prediction time step; the pressure gradient field is calculated using the three-dimensional central difference method, and the measured pressure gradient is collected in real time by a pressure sensor array deployed at key nodes of the pipe gallery (every 5 meters along the pipeline axis), and the predicted gradient is corrected for residual error based on the prediction error.

[0017] Preferably, the valve wear quantification method is as follows: valve opening changes are recorded in real time using a rotary encoder; the valve stem axial force is measured using a torque sensor; the cumulative wear is calculated using a modified Arcard wear formula, and the cumulative wear is normalized to a dimensionless wear parameter; the formula for calculating the cumulative wear L(t) is as follows:

[0018]

[0019] Where Ks is the wear coefficient of the valve material, H is the material hardness, v(τ) is the valve stem speed, and f(T(τ)) is the temperature compensation function (T is the real-time temperature), which is obtained by modeling the relationship between the material's thermal expansion coefficient and temperature-hardness; L(t) is normalized to a dimensionless wear parameter. Where L max The valve design life threshold is defined as F(τ); F(τ) represents the axial force on the valve stem (unit: N), which is measured in real time by a torque sensor installed on the valve stem drive mechanism.

[0020] Preferably, the method for obtaining the P matrix of the pressure field in spacetime includes:

[0021] The preprocessed, standardized time-series data stream is arranged in chronological order to form a matrix structure; rows represent the location of each sensor or grid point; columns represent the pressure value at each sampling time, including the pressure distribution throughout the pipeline.

[0022] The P matrix is ​​decomposed into three matrices using eigenorthogonal decomposition:

[0023] The left singular vector, each column represents a spatial fluctuation mode, i.e., a modal basis function;

[0024] A diagonal matrix, where the diagonal elements are singular values, representing the importance of each modal basis function;

[0025] The right singular vectors, each row representing the time evolution of the corresponding modal basis function.

[0026] Preferably, during the generation of the P matrix in the pressure field space-time, Kriging interpolation is used to dynamically compensate for the full-order flow field data and real-time sensing data to unify them to the same spatial resolution; the physical constraints ignore the gravity term; and a transfer learning mechanism is introduced into LSTM-PINN to adjust the distribution of the full-order flow field data and real-time data through the domain adaptation loss function.

[0027] Preferably, the method for obtaining the transient pressure wave propagation prediction model includes the following steps:

[0028] Step S11, CFD full-order flow field simulation and data matrix construction: Use CFD numerical simulation to generate full-order flow field data covering typical working conditions. The input includes pipe gallery geometric parameters and fluid property parameters; simulate typical working conditions by solving the Navier-Stokes equations to obtain simulation data of typical working conditions; construct a pressure field snapshot matrix by extracting the spatial pressure distribution of the full-order flow field data.

[0029] Step S12, POD mode basis function extraction and orthogonalization: Perform eigenorthogonal decomposition on the pressure field snapshot matrix, decompose it into several orthogonal mode basis functions, obtain the dominant mode basis functions, and verify the mode orthogonality;

[0030] Step S13: Construct an LSTM time-series encoder. By taking real-time collected pressure, flow, and temperature data as input, and combining them with the dominant mode basis function, an LSTM-PINN physical information neural network is constructed. The LSTM time-series encoder is used to process time-series data, and the physical constraint layer ensures that the output conforms to the Navier-Stokes equation. The transient pressure wave propagation prediction model is obtained by training based on the optimizer and training data.

[0031] Step S14, Online Dynamic Update and Accuracy Verification: At a fixed frequency, the LSTM-PINN network is incrementally learned by collecting new sensor data, and the network parameters are adjusted to adapt to real-time operating conditions. The transient pressure wave propagation prediction model is verified and optimized in real time by calculating the global average absolute error and the pressure wave peak position error.

[0032] Preferably, the physical constraint term L of the domain adaptation loss function of the LSTM-PINN hybrid network phy for:

[0033]

[0034] Where λ is the physical constraint weight coefficient, determined through cross-validation; P pred The predicted pressure field is represented by u, the velocity vector field is represented by ν, and the dynamic viscosity coefficient is represented by ν. For the Laplace operator, α i (t) represents the time-dependent modal weighting coefficients, characterizing the contribution of the i-th order modal basis function to the pressure field at time t, k represents the number of dominant modal basis functions, and i represents the index; φ i (t) represents the POD modal basis function, which is the dominant pressure wave spatial distribution mode extracted from the full-order flow field data.

[0035] Preferably, the domain adaptation loss function is configured as follows:

[0036] L total =||P pred -P preal ||+L phy

[0037] Among them, the data fitting term ||P pred -P preal || indicates that the prediction result is forced to approximate the measured data, and the physical constraint term L phy This is used to ensure that the prediction conforms to the laws of fluid dynamics conservation.

[0038] Preferably, the multi-objective reward function is a weighted summation of pressure fluctuation suppression, valve life optimization, and execution efficiency improvement. Pressure fluctuation suppression is negatively correlated with real-time pressure deviation, thus suppressing the risk of pressure overshoot. Valve life optimization is negatively correlated with cumulative mechanical wear of the valve, reducing equipment wear caused by frequent operations. Execution efficiency improvement is positively correlated with valve action torque efficiency, ensuring a balance between response speed and energy consumption. The weight coefficients of each sub-item are determined through a multi-objective optimization algorithm to generate a Pareto optimal control strategy set.

[0039] Preferably, in the physical constraint term L phy By adding a pipe wall elasticity correction term, the updated physical constraint term XL is obtained. phy :

[0040]

[0041] Where E(t) represents the real-time inverted elastic modulus of the pipe wall, obtained by deploying strain sensors at key locations on the pipe wall; ∈(t) represents the pipe wall strain tensor. The tensor product represents the coupling effect between the pressure gradient and the strain field. This represents the feedback effect of the elastic deformation of the pipe wall on the pressure wave, and ρ represents the fluid density in the petrochemical pipe gallery.

[0042] Preferably, the dual closed-loop anti-control mechanism includes:

[0043] The inner loop control layer is used for high-frequency regulation: it adopts integral separation PID, which dynamically amplifies the control signal according to the deviation between the target and actual opening degree of the valve; if the deviation between the target and actual opening degree is greater than the preset value, the proportional term quickly outputs a strong correction signal; the integral term is frozen, and only the proportional and derivative actions are retained to prevent the valve from getting stuck in the limit position.

[0044] The outer loop control layer is used for global planning: based on real-time pressure feedback, it generates a smooth valve opening curve to avoid mechanical shock caused by step movements and limit the pressure gradient to not exceed the preset threshold.

[0045] Preferably, during the inner-loop PID control process, the inner-loop PID proportional coefficient K p (t) is positively correlated with the viscosity of the medium. When the viscosity increases, the proportional coefficient of the inner loop PID is increased. This is the core parameter for controlling the response speed. The coefficient is increased to compensate for the delay and speed up the valve response speed, thus compensating for the delay caused by viscous resistance. When the viscosity decreases, the proportional coefficient of the inner loop PID is decreased to avoid overshoot oscillation.

[0046] The technical effects and advantages of this invention are as follows:

[0047] (1) The petrochemical pipeline intelligent monitoring system based on the Internet of Things provided by this invention realizes dynamic noise suppression and spatiotemporal alignment of multimodal sensor data through dynamic Kalman filtering and clock synchronization technology, generates a highly consistent standardized time-series data stream, and enables the standardized time-series data stream to dynamically adapt to the changes in the physical characteristics of fluids and pipelines, solving the problem of misjudgment / missed judgment caused by the inability of traditional fixed pressure threshold methods to adapt to the dynamic changes in medium viscosity and pipe wall elastic modulus; through transient pressure wave propagation prediction model and Kriging interpolation dynamic compensation technology, it realizes high-precision simulation of fluid-structure coupling effect, solves the prediction distortion problem caused by the traditional method ignoring the transient pressure wave propagation characteristics, and provides a physically interpretable decision basis for valve control.

[0048] (2) The petrochemical pipeline intelligent monitoring system based on the Internet of Things provided by the present invention achieves dynamic balance control of safety and economy through NSGA-II multi-objective optimization (balancing pressure fluctuation suppression, valve life, energy efficiency ratio weight) and dual closed-loop control mechanism, combined with FPGA hardware acceleration (end-to-end delay ≤ 5ms), and solves the problem that a single threshold strategy is difficult to adapt to complex working conditions. Attached Figure Description

[0049] Figure 1 This is a structural block diagram of the intelligent monitoring system for petrochemical pipeline corridors according to the present invention.

[0050] Figure 2 A flowchart illustrating the construction of the transient pressure wave propagation prediction model of this invention.

[0051] Figure 3 This is a block diagram of the intelligent monitoring system for petrochemical pipeline corridors based on the domain adaptation loss function of the present invention. Detailed Implementation

[0052] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0053] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0054] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0055] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0056] Example 1, see Figure 1 The present invention provides a structural block diagram of an intelligent monitoring system for petrochemical pipeline corridors. Figure 1 The illustrated intelligent monitoring system for petrochemical pipeline corridors based on the Internet of Things includes:

[0057] The petrochemical pipeline corridor data acquisition module is used to achieve full-dimensional capture of transient pressure wave characteristics and ensure data consistency. By deploying a sensor array (pressure sensor, ultrasonic flow meter and temperature sensor) in the petrochemical pipeline corridor, dynamic noise suppression and spatiotemporal alignment processing are performed on multimodal sensor data to obtain a standardized time-series data stream.

[0058] The pressure wave propagation prediction module is used to build a transient pressure wave propagation prediction model. It extracts the P matrix of the pressure field in space-time from the standardized time-series data stream, and performs singular value decomposition on the P matrix using intrinsic orthogonal decomposition (POD) to obtain the dominant mode basis functions. The LSTM-PINN physical information neural network is trained using full-order flow field data generated by simulation and the standardized time-series data stream to obtain the transient pressure wave propagation prediction model. LSTM is used to process the time-series data, and PINN is used to introduce physical constraints to ensure that the model conforms to the laws of fluid mechanics.

[0059] The valve control strategy generation module defines a state vector containing pressure gradient, valve wear, and energy efficiency ratio based on the pressure wave prediction results of the transient pressure wave propagation prediction model. It uses the NSGA-II algorithm to optimize the weights of the multi-objective reward function and trains a distributed Q-learning policy library through a priority experience replay mechanism to generate a Pareto optimal control instruction set.

[0060] Furthermore, the pressure gradient acquisition method is as follows: based on the predicted pressure field output by the transient pressure wave propagation prediction model, the prediction result includes the pressure distribution of the pipeline spatial coordinates at the prediction time step; the pressure gradient field is calculated using the three-dimensional central difference method, and the measured pressure gradient is collected in real time by a pressure sensor array deployed at key nodes of the pipe gallery (every 5 meters along the pipeline axis), and the predicted gradient is corrected for residual error based on the prediction error.

[0061] Furthermore, the valve wear quantification method is as follows: valve opening changes are recorded in real time using a rotary encoder; the axial force of the valve stem is measured based on a torque sensor; the cumulative wear amount is calculated using a modified Arcard wear formula, and the cumulative wear amount is normalized to a dimensionless wear parameter; the calculation formula for the cumulative wear amount L(t) is as follows:

[0062]

[0063] Where Ks is the wear coefficient of the valve material, H is the material hardness, v(τ) is the valve stem speed, and f(T(τ)) is the temperature compensation function (T is the real-time temperature), which is obtained by modeling the relationship between the material's thermal expansion coefficient and temperature-hardness; L(t) is normalized to a dimensionless wear parameter. Where L max The valve design life threshold; F(τ) represents the axial force on the valve stem (unit: N), which is measured in real time by a torque sensor installed on the valve stem drive mechanism;

[0064] The valve control and regulation module deploys the Pareto optimal control instruction set to FPGA hardware, and solidifies the dominant mode basis function and action value table lookup logic (which can be directly called through hardware logic after solidification, saving CPU calculation time; the action value table lookup logic maps the action value table Q-table obtained by strategy training to the FPGA lookup table to achieve nanosecond-level control instruction matching) to improve control efficiency. Based on the standardized timing data stream in step one, a dual closed-loop control mechanism (inner loop PID regulation, outer loop pressure feedback interpolation) is used to achieve precise regulation and disturbance rejection capability of valve action. Under the premise of ensuring pressure fluctuation safety constraints, the valve response speed and equipment life are optimized, and the valve action signal is output with an end-to-end delay of ≤5ms.

[0065] In this embodiment of the invention, it is necessary to further explain that the preprocessed standardized time-series data stream is arranged in chronological order to form a matrix structure: rows represent the location of each sensor or grid point (e.g., 1,000 pressure sensors deployed every 5 meters along the pipeline), and columns represent the pressure value at each sampling moment (e.g., 2,000 time points collected at 1ms intervals within 2 seconds), including the pressure distribution of the entire pipeline;

[0066] The P matrix is ​​decomposed into three matrices using eigenorthogonal decomposition:

[0067] The left singular vector, each column represents a spatial fluctuation mode, i.e., a modal basis function;

[0068] A diagonal matrix, where the diagonal elements are singular values, representing the importance of each modal basis function;

[0069] The right singular vectors, each row representing the time evolution of the corresponding modal basis function.

[0070] In this embodiment of the invention, it is necessary to further explain that during the generation of the spatiotemporal P matrix of the pressure field, Kriging interpolation dynamic compensation is performed on the full-order flow field data and real-time sensing data (5m interval sensors) to unify them to the same spatial resolution; the physical constraints ignore the gravity term; a transfer learning mechanism is introduced into LSTM-PINN, and the distribution of the full-order flow field data and real-time data is adjusted through a domain adaptation loss function. The domain adaptation loss function is used to align the distribution differences between CFD simulation data (including the elastic pipe wall assumption) and real-time sensing data (including strain measurements), thereby improving the generalization ability of the prediction model.

[0071] Further explanation is needed in the embodiments of the present invention, see the following: Figure 2 The flowchart for building a transient pressure wave propagation prediction model is shown below. The acquisition of the transient pressure wave propagation prediction model includes the following steps:

[0072] Step S11, CFD Full-Order Flow Field Simulation and Data Matrix Construction: Generate full-order flow field data covering typical working conditions through CFD numerical simulation. Inputs include pipe gallery geometric parameters (such as pipe diameter and pipe length) and fluid physical property parameters (such as viscosity and density). By extracting the spatial pressure distribution of each set of full-order flow field data, a pressure field snapshot matrix is ​​constructed. The pressure field snapshot matrix is ​​a subset of pressure parameters extracted from the full-order flow field data, which is structured and used for dimensionality reduction analysis.

[0073] Furthermore, in this embodiment of the invention, at least 1,000 sets of full-order flow field data are generated, using a grid number ≥ 1e 6 The time step Δt = 1 ms is obtained by solving the Navier-Stokes equations (the basic equations describing fluid flow) to simulate typical working conditions (laminar flow, turbulent flow, emergency valve closure);

[0074] Step S12, POD mode basis function extraction and orthogonalization: Perform intrinsic orthogonal decomposition (singular value decomposition is used in this embodiment of the invention) on the pressure field snapshot matrix, decompose it into several orthogonal mode basis functions, retain the first k singular values ​​with 99% energy, obtain the dominant mode basis functions, and verify the mode orthogonality to ensure that the residual energy ratio does not exceed 1%;

[0075] This step involves reducing the dimensionality of the flow field data, extracting the most representative modal features, reducing computational complexity, and maintaining the accuracy of fluid dynamics. Modal basis functions can be understood as "core pressure fluctuation patterns" extracted from a large amount of fluid simulation data, similar to using a few key waveform combinations to reconstruct complex water flow pressure changes. Modal basis functions can efficiently characterize the core laws of pressure wave propagation within a pipe.

[0076] Step S13: Construct an LSTM (Long Short-Term Memory) temporal encoder. By taking real-time collected pressure, flow, and temperature data as input, and combining them with the dominant mode basis function, an LSTM-PINN physical information neural network is constructed. The LSTM temporal encoder is used to process the temporal data, and the physical constraint layer ensures that the output conforms to the Navier-Stokes equation. The transient pressure wave propagation prediction model is obtained by training based on the optimizer and training data.

[0077] The explanation explains that this step combines fluid dynamics principles with real-time data using a deep learning model to achieve accurate prediction of transient pressure waves and output a transient pressure wave propagation prediction model. The training data is acquired as follows:

[0078] Data acquisition: Based on 1000 sets of full-order flow field data generated in step S11, covering classic working conditions (such as DN50-DN1200 pipe diameter, 0.1-500cP viscosity); standardized time-series data stream (pressure, flow rate, temperature, sampling rate 100kHz) from step one.

[0079] Data partitioning: The acquired data is divided into an offline training set and an online fine-tuning set. The offline training set is used for initial model training; the online fine-tuning set is used for dynamically updating model parameters.

[0080] Data augmentation: Improve model robustness by adding Gaussian noise (SNR=40dB) to simulate sensor errors;

[0081] Full-order flow field data refers to complete spatial resolution fluid field data obtained through CFD numerical simulation, including:

[0082] Spatial dimension: 3D grid points across the entire pipeline (grid number ≥ 1×10⁻⁶) 6 The transient pressure, flow rate, and temperature distribution of ( );

[0083] Time dimension: The simulation duration covers the complete pressure wave propagation cycle (e.g., 0-2s for an emergency valve closure process), with a time step Δt = 1ms;

[0084] Physical field completeness: It strictly satisfies the Navier-Stokes equations and includes characteristics of multiple flow regimes such as laminar flow, turbulent flow, and transient flow;

[0085] Step S14, Online Dynamic Update and Accuracy Verification: At a fixed frequency (e.g., every 5 minutes), the LSTM-PINN network is incrementally learned by collecting new sensor data, and the network parameters are adjusted to adapt to real-time operating conditions. The transient pressure wave propagation prediction model is verified and optimized in real time by calculating the global mean absolute error (MAE) and pressure wave peak position error (Δx) of the pressure wave, ensuring that the error is kept within 2% and the pressure wave peak position error is ≤0.5m.

[0086] In this embodiment of the invention, it is necessary to further explain the physical constraint term L of the domain adaptation loss function of the LSTM-PINN hybrid network. phy for:

[0087]

[0088] Wherein, λ is the physical constraint weight coefficient, determined through cross-validation (too high a λ, such as λ>0.5, will lead to model rigidity; too low a λ, such as λ<0.05, will not effectively constrain physical laws; in this embodiment of the invention, λ=0.08); P pred Predicted pressure field, u is the flow velocity vector field, ν is the dynamic viscosity coefficient. For the Laplace operator; α i (t) represents the time-dependent modal weighting coefficients, characterizing the contribution of the i-th order modal basis function to the pressure field at time t, k represents the number of dominant modal basis functions, and i represents the index; φ i (t) represents the POD modal basis function, which is the dominant pressure wave spatial distribution mode extracted from the full-order flow field data.

[0089] In this embodiment of the invention, it needs to be further explained that the domain adaptation loss function is constructed as follows:

[0090] L total =||P pred -P preal ||+L phy

[0091] Among them, the data fitting term ||P pred -P preal || indicates that the predicted result is forced to approximate the measured / simulation data, and the physical constraint term L phy This is used to ensure that predictions conform to the laws of fluid dynamics conservation, avoid "black box" models from outputting results that violate common sense physics, and ensure that prediction results conform to actual fluid dynamic laws, thereby improving the reliability of predictions.

[0092] In this embodiment of the invention, it is necessary to further explain that the present invention does not impose specific restrictions on the quantification form of the multi-objective reward function, as long as it reflects the constraints of pressure fluctuation suppression, valve life optimization, and execution efficiency improvement. Among them, pressure fluctuation suppression is negatively correlated with the real-time pressure deviation, suppressing the risk of pressure overshoot; valve life optimization is negatively correlated with the cumulative mechanical wear of the valve, reducing equipment losses caused by frequent actions; execution efficiency improvement is positively correlated with the valve action torque efficiency, ensuring a balance between response speed and energy consumption. The weight coefficients of each sub-item are determined by a multi-objective optimization algorithm (such as NSGA-II), generating a Pareto optimal control strategy set.

[0093] Furthermore, to facilitate understanding of the multi-objective reward function R, the following domain adaptation loss function is listed;

[0094]

[0095] Where w1, w2, and w3 are the weighting coefficients for each item, and w1 + w2 + w3 = 1.0, determined through optimization using the NSGA-II algorithm, with typical values ​​of w1 = 0.6, w2 = 0.3, and w3 = 0.1; ΔP is the absolute deviation between the real-time pressure and the target pressure; L is the cumulative wear of the valve, L max τ is the valve design life threshold. v (t) Valve operating torque, G rated This refers to the valve's rated torque.

[0096] In one possible embodiment, the weight coefficients w1, w2, and w3 of the multi-objective reward function are changed from static optimization to dynamic adjustment, including: when viscosity increases, increasing the pressure fluctuation suppression weight w1 to prioritize safety; when the elastic modulus decreases, increasing the valve life optimization weight w2 to reduce frequent actions.

[0097] In this embodiment of the invention, it needs to be further explained that the dual closed-loop anti-control mechanism includes:

[0098] The inner loop control layer is used for high-frequency regulation (e.g., 1kHz): it employs integral-separated PID control to dynamically amplify the control signal based on the valve opening deviation (target vs. actual). For example, if the target opening is 30% and the actual opening deviation reaches 5% (i.e., <25% or >35%), the proportional term quickly outputs a strong correction signal. When the opening error is ≥5%, the integral term is frozen (to avoid overshoot caused by continuous error accumulation), retaining only the proportional and derivative actions to prevent the valve from jamming at its limit positions (e.g., fully open / fully closed). This prevents damage to the actuator due to long-term operation at limit positions and extends the valve's lifespan.

[0099] Example scenario: When the pipeline pressure suddenly increases and the valve needs to be closed quickly, the inner loop finely adjusts the opening at a 1ms cycle. If the instantaneous error is too large (such as the target is 90% closed and the actual opening is 75%), the integral term is immediately frozen and the target value is quickly approximated only by the proportional term.

[0100] The outer loop control layer is used for global planning (100Hz): based on real-time pressure feedback (such as pressure peak position and gradient changes), it generates a smooth valve opening curve to avoid mechanical shock caused by step movements. For example, when adjusting from 30% opening to 50%, a continuously differentiable S-shaped curve is generated instead of an instantaneous jump; for example, a smooth trajectory is generated based on cubic spline interpolation to limit the pressure gradient from exceeding a preset threshold.

[0101] In one possible embodiment, the inner loop PID proportional coefficient K p (t) is positively correlated with the viscosity of the medium:

[0102]

[0103] Where μ(t) represents the real-time viscosity of the medium, μ ref Indicates the preset standard viscosity; K p0 This represents the preset inner loop PID proportional coefficient; obtained through the formula: when viscosity increases, the inner loop PID proportional coefficient is increased, which is the core parameter for controlling the response speed. It is increased to compensate for the delay and speed up the valve response speed (compensating for the delay caused by viscous resistance); when viscosity decreases, the inner loop PID proportional coefficient is decreased to avoid overshoot oscillation.

[0104] Implementation steps: The viscosity sensor collects μ(t) in real time, and noise is eliminated by moving average filtering; the proportional coefficient K of the inner loop PID controller is calculated. p (t), and update it to the FPGA's PID control register;

[0105] The outer loop pressure feedback interpolation layer synchronously adjusts the slope of the S-curve to ensure that the pressure gradient does not exceed a preset threshold, such as not exceeding a fixed threshold of 60% P_max / m, where P_max / m represents the maximum allowable pressure gradient per meter of pipe.

[0106] In one possible embodiment, the fixed threshold is changed to be linked to the physical properties of the medium (viscosity, density), and the threshold is automatically reduced when the viscosity decreases.

[0107] Example 2 differs from Example 1 in that it considers the influence of pipe wall elasticity on pressure. Based on this, in the physical constraint term L... phy By adding a pipe wall elasticity correction term, the updated physical constraint term XL is obtained. phy And set the domain adaptation loss function based on the updated physical constraints:

[0108]

[0109] Where E(t) represents the real-time inverted elastic modulus of the pipe wall (measured by a strain sensor), and ε(t) represents the pipe wall strain tensor (describing the degree of pipe wall deformation). The tensor product represents the coupling effect between the pressure gradient and the strain field. This represents the feedback effect of the elastic deformation of the pipe wall on the pressure wave, and ρ represents the fluid density in the petrochemical pipe gallery.

[0110] In one possible embodiment, after adding the pipe wall elasticity correction term, the following operation is added during Kriging interpolation dynamic compensation:

[0111] Data acquisition: Deploy strain sensors at key locations on the pipe wall (e.g., one group every 20 meters) to measure local strain ε(t) in real time and calculate the corresponding elastic modulus E(t) of the pipe wall;

[0112] Interpolation inversion: The sparse strain sensor data is interpolated into the elastic modulus distribution E(x,t) of the entire pipeline using the Kriging interpolation dynamic compensation algorithm (based on spatial correlation).

[0113] Dynamic updates: The real-time changing pipe wall elastic modulus is mapped to the physical constraint layer of LSTM-PINN to correct the wave velocity calculation model in real time.

[0114] In one possible embodiment, after adding the pipe wall elasticity correction term, during the training and optimization process in step S13, pipe wall strain sensor data is added to the training data, and fluid-structure coupling simulation datasets (such as the CFD case of pipe wall vibration and pressure wave interaction) are introduced; the domain adaptation loss function is used to align the distribution differences between CFD simulation data (including the elastic pipe wall assumption) and real-time sensing data (including strain measurement) to improve the generalization ability of the prediction model.

[0115] Example 3, see Figure 3 The structural block diagram of the intelligent monitoring system for petrochemical pipe gallery based on the domain adaptation loss function is presented. The transfer learning mechanism is introduced into LSTM-PINN. The domain adaptation loss function is used to analyze the full-order flow field data and real-time data distribution. The physical constraint terms in the domain adaptation loss function are: physical constraint terms based on dynamic viscosity (see Example 1), physical constraint terms with added pipe wall elasticity correction terms (see Example 2), and any one or a combination of physical constraint terms considering multiphase mixing.

[0116] The difference between this embodiment and embodiments 1 and 2 is that the system further includes:

[0117] The multiphase flow feature extraction module monitors the volume fraction of each phase (oil, gas, and water) in real time by deploying dielectric constant sensors and microwave moisture meters in the petrochemical pipeline corridor, and constructs a multiphase flow mixing parameter matrix M, which includes the volume fraction, mixing density, and mixing viscosity of each phase (oil, gas, and water).

[0118] In the pressure wave propagation prediction module of Example 1, multiphase flow is introduced into the physical constraint terms of LSTM-PINN to modify the Navier-Stokes equation, generating full-order flow field data containing phase distribution (such as gas phase volume fraction cloud map and oil-water interface tracking). Interphase interaction terms, such as surface tension and interphase momentum exchange terms, are introduced into the original fluid dynamics equation.

[0119] The physical constraint layer adds residual calculation of multiphase flow equations to ensure that the predicted pressure field conforms to the multiphase flow conservation law; the dynamics of the interphase interface are simulated by the VOF (Volume of Fluid) method, and the data dimension is expanded to four dimensions (space + time + phase fraction);

[0120] The stratification trend is obtained by calculating the ratio of the maximum spatial gradient of the volume fraction of oil, gas and water phases to the arithmetic mean of the volume fraction of each phase; if the stratification trend of the petrochemical pipeline is found to exceed the preset value, adaptive valve regulation is triggered to adjust the stratification trend by controlling the valve.

[0121] Stratification trend is an important indicator for dynamically quantifying stratification trends. It can accurately reflect the degree of volume change of fluid in local areas and the overall mixing uniformity.

[0122] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent monitoring system for a petrochemical pipeline based on the Internet of Things, characterized by, The method comprises the following steps: A petrochemical pipe gallery data acquisition module is arranged by deploying a sensor array in the petrochemical pipe gallery, at least including a pressure sensor, an ultrasonic flowmeter and a temperature sensor, dynamic noise suppression and space-time alignment processing are performed on multi-modal sensor data, and standardized time series data flow is obtained; A pressure wave propagation prediction module is used to build a transient pressure wave propagation prediction model, extract a P matrix of the pressure field space-time from the standardized time series data flow, perform singular value decomposition on the P matrix by using eigen-orthogonal decomposition, and obtain a dominant modal basis function; the LSTM-PINN physical information neural network is trained by using the full-order flow field data generated by simulation and the standardized time series data flow, and the transient pressure wave propagation prediction model is obtained; A valve control strategy generation module is used to define a state vector including a pressure gradient, valve wear and energy efficiency ratio based on the pressure wave prediction result of the transient pressure wave propagation prediction model; the NSGA-II algorithm is used to optimize the weight of the multi-objective reward function, the distributed Q-learning strategy library is trained by using the priority experience replay mechanism, and a Pareto optimal control instruction set is generated; A valve control adjustment module is used to deploy the Pareto optimal control instruction set to the FPGA hardware, solidify the dominant modal basis function and the action value table query logic; based on the standardized time series data flow, the precision adjustment and anti-disturbance ability of the valve action are realized by using the double closed-loop control mechanism including the inner loop PID adjustment and the outer loop pressure feedback interpolation, and the valve action signal is output. 2.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 1, characterized in that, The acquisition method of the P matrix of the pressure field space-time comprises the following steps: The preprocessed standardized time series data flow is arranged in time sequence to form a matrix structure; the rows represent the positions of each sensor or grid point; and the columns represent the pressure values at each sampling time, including the pressure distribution of the whole pipeline; The P matrix is decomposed into three matrices by using eigen-orthogonal decomposition, which are: The left singular vector, each column of which represents a spatial wave mode, i.e. a modal basis function; The diagonal matrix, the diagonal elements of which are singular values, indicating the importance of each modal basis function; The right singular vector, each row of which represents the time evolution law of the corresponding modal basis function. 3.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 2, characterized in that, In the generation process of the P matrix of the pressure field space-time, the Kriging interpolation dynamic compensation is performed on the full-order flow field data and the real-time sensor data, and they are unified to the same spatial resolution; the physical constraint ignores the gravity term; the transfer learning mechanism is introduced into the LSTM-PINN, and the domain adaptation loss function is used to align the full-order flow field data and the real-time data distribution.

4. The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 3, characterized in that, The acquisition method of the transient pressure wave propagation prediction model comprises the following steps: Step S11, CFD full-order flow field simulation and data matrix construction: using CFD numerical simulation to generate full-order flow field data covering typical working conditions, inputting including pipe gallery geometric parameters and fluid physical parameters; simulation data of typical working conditions are obtained by solving Navier-Stokes equation; a pressure field snapshot matrix is constructed by extracting the spatial pressure distribution of the full-order flow field data; Step S12, POD modal basis function extraction and orthogonalization: the pressure field snapshot matrix is subjected to eigen-orthogonal decomposition, is decomposed into a plurality of orthogonal modal basis functions, the dominant modal basis function is obtained, and the modal orthogonality is verified; Step S13, constructing an LSTM time series encoder, combining the real-time collected pressure, flow and temperature data as input with the dominant modal basis function, constructing an LSTM-PINN physical information neural network, wherein the LSTM time series encoder is used to process time series data, and the physical constraint layer ensures that the output conforms to the Navier-Stokes equation; training based on the optimizer and training data to obtain a transient pressure wave propagation prediction model; Step S14, online dynamic updating and accuracy verification: at a set fixed frequency, new sensor data is collected to perform incremental learning on the LSTM-PINN network, and the network parameters are adjusted to adapt to real-time working condition changes; the global average absolute error and pressure wave peak position error of the pressure wave are calculated to verify and optimize the transient pressure wave propagation prediction model in real time. 5.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 4, characterized in that, The physical constraint term L of the domain adaptation loss function of the LSTM-PINN hybrid network phy is: where λ is the physical constraint weight coefficient, determined by cross-validation; P pred represents the predicted pressure field, u is the flow velocity vector field, and v is the dynamic viscosity coefficient, is the Laplace operator, and α i (t) represents the time-dependent modal weight coefficient, representing the contribution of the i-th modal basis function to the pressure field at time t, k represents the number of dominant modal basis functions, and i represents the index; φ i (t) represents the POD modal basis function, which is the dominant pressure wave spatial distribution pattern extracted from the full-order flow field data. 6.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 5, characterized in that, The domain adaptation loss function is composed as follows: L total =‖P pred -P preal ‖+L phy where the data fitting term ||P pred -P preal ‖ indicates that the predicted results are forced to approximate the measured data, the physical constraint term L phy is used to ensure that the prediction complies with the conservation laws of fluid mechanics.

7. The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to any one of claims 1-4, characterized in that, The multi-objective reward function is a weighted sum of pressure wave suppression, valve life optimization, and execution efficiency improvement, wherein the pressure wave suppression is negatively correlated with the real-time pressure deviation, and the pressure overshoot risk is suppressed; the valve life optimization is negatively correlated with the cumulative mechanical wear of the valve, and the equipment loss caused by frequent action is reduced; the execution efficiency improvement is positively correlated with the valve action torque efficiency, ensuring the balance between response speed and energy consumption; the multi-objective optimization algorithm is used to determine the weight coefficients of each sub-item to generate a Pareto optimal control strategy set. 8.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 5, characterized in that, In the physical constraint term L phy , a tube wall elasticity correction term is added to obtain an updated physical constraint term XL phy : where E(t) represents the real-time inverted pipe wall elastic modulus, which is obtained by deploying strain sensors at key locations of the pipe wall; ∈(t) represents the pipe wall strain tensor, is the tensor product, which represents the coupling effect of the pressure gradient and the strain field, represents the feedback effect of the pipe wall elastic deformation on the pressure wave, and ρ represents the fluid density in the petrochemical pipe gallery. 9.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to any one of claims 1-4, characterized in that, The double closed-loop control mechanism includes: The inner loop control layer is used for high-frequency regulation: integral separation PID is used to dynamically amplify the control signal according to the target and actual opening deviation of the valve; if the target and actual opening deviation is greater than the preset value, the proportional term quickly outputs a strong correction signal; the integral term is frozen, and only the proportional and differential actions are retained to prevent the valve from being stuck at the limit position; The outer loop control layer is used for global planning: based on real-time pressure feedback, a smooth valve opening curve is generated to avoid mechanical impact caused by step action and limit the pressure gradient to be less than a preset threshold. 10.The intelligent monitoring system for petrochemical pipeline based on Internet of Things according to claim 8, characterized in that, In the inner loop PID regulation process, the inner loop PID proportional coefficient K p (t) is positively correlated with the medium viscosity. When the viscosity increases, the inner loop PID proportional coefficient is increased, which is the core parameter of controlling the response speed. When the viscosity increases, it is increased to compensate for the delay and speed up the valve response speed to compensate for the delay caused by the viscous resistance. When the viscosity decreases, the inner loop PID proportional coefficient is reduced to avoid overshoot oscillation.

Images