Reverse flow valve based on early warning control of digital twin technology

Abstract

The present invention discloses a reverse flow valve based on digital twin technology early warning control, which is suitable for precise control and safety assurance of unidirectional flow of fluid; the reverse flow valve comprises a valve body and a valve disc assembly, and utilizes multi-source sensors to collect key data such as fluid flow direction and water hammer impact characteristics in real time; the edge control unit processes the sensor data, monitors the risks of reverse flow and water hammer, and performs rapid valve closing operations in emergency situations, while communicating with the central processing unit; the central processing unit establishes and updates the dynamic characteristic model of the reverse flow valve in real time through the digital twin engine, including valve disc opening characteristics, fluid pressure and water hammer characteristics; the intelligent analysis and early warning engine identifies risk levels and development trends based on deep learning algorithms, and generates early warning signals; the adaptive control engine combines risk signals and dynamic data to optimize the valve disc closing parameters in the control strategy library; this technical solution effectively improves the intelligence level and operational reliability of the reverse flow valve, and can be widely used in the field of industrial fluid control.

Description

Technical Field

[0001] The present invention relates to the technical field of reverse flow valves, and in particular to a reverse flow valve based on early warning control of digital twin technology. Background Art

[0002] In existing technology, reverse flow valves are widely used in various industrial piping systems to ensure unidirectional fluid flow and prevent reverse flow. Traditional reverse flow valves typically achieve automatic fluid isolation through mechanical structures. Some improved reverse flow valves incorporate sensors and automatic control devices to enhance performance. These reverse flow valves monitor basic parameters such as fluid pressure and flow rate and make appropriate adjustments when anomalies are detected, thereby mitigating the risk of water hammer and reverse flow in pipelines.

[0003] However, existing reverse flow valve technology has the following problems: First, the single data collected by the sensor is often insufficient to fully reflect the complex fluid environment, resulting in inaccurate predictions of reverse flow and water hammer; second, the control logic is mostly based on fixed preset rules, which makes it difficult to dynamically adapt to changes in operating conditions, which may lead to a mismatch between the control strategy and the actual situation; finally, due to the lack of intelligent data analysis and real-time optimization methods, the existing system is less efficient when dealing with complex working conditions, which may cause excessive wear of the valve body and pipeline or operational delays.

[0004] Based on this, it is necessary to develop a reverse flow valve based on early warning control of digital twin technology. Summary of the Invention

[0005] The present application provides a reverse flow valve based on early warning control of digital twin technology to improve the intelligence level and operational reliability of the reverse flow valve.

[0006] This application provides a reverse flow valve based on early warning control of digital twin technology, including:

[0007] A valve body, wherein a valve flap assembly is provided in the valve body for controlling the unidirectional flow of fluid, and the valve flap assembly includes a valve flap and a valve seat;

[0008] A multi-source sensor is provided on the valve body, and the multi-source sensor includes a differential pressure sensor provided before and after the valve disc, for detecting the flow direction and pressure difference of the fluid; an angle sensor provided on the valve disc, for detecting the opening angle of the valve disc; a stress sensor provided on the valve seat, for detecting the stress state of the valve seat sealing surface; and a vibration sensor provided on the valve body, for detecting water hammer impact characteristics;

[0009] an electric actuator, provided on the valve body, for quickly closing the valve disc when reverse flow is detected;

[0010] an edge control unit, mounted on the valve body, for receiving the reverse flow valve operating data collected by the multi-source sensor; monitoring the reverse flow state and water hammer risk through the reverse flow valve operating data, and performing rapid valve closing control in an emergency; transmitting the reverse flow valve operating data collected by the multi-source sensor to a central processing unit; receiving a control instruction sent by the central processing unit, and adjusting the reverse flow valve through the electric actuator according to the control instruction;

[0011] A central processing unit, connected to one or more edge control units of the reverse flow valve via a communication network, comprising:

[0012] a digital twin engine configured to receive operating data transmitted by the edge control unit; establish a digital twin model of the reverse flow valve based on the operating data; continuously update the digital twin model based on real-time operating data to generate dynamic characteristic data including valve disc opening characteristics, fluid pressure distribution, and water hammer characteristics; and transmit the dynamic characteristic data to an intelligent analysis and early warning engine;

[0013] an intelligent analysis and warning engine, configured to receive the dynamic characteristic data, analyze the valve disc motion characteristics and fluid state characteristics using a deep learning algorithm, identify backflow risks and water hammer risks, generate a warning signal including the risk type and level based on the identified risk characteristics, and transmit the warning signal to the adaptive control engine;

[0014] An adaptive control engine is used to receive the dynamic characteristic data and early warning signal; based on a preset control strategy library, a basic control strategy is selected according to the risk type and level in the early warning signal, wherein the control strategy library stores corresponding valve flap closing parameter templates for different risk levels; in combination with the fluid pressure distribution and water hammer characteristics in the dynamic characteristic data, the closing parameters in the basic control strategy are dynamically optimized to generate an optimal valve flap closing curve, wherein the closing curve defines the segmented speed and timing of the valve flap from starting to close to fully close; the closing curve is converted into a control instruction of the actuator, wherein the control instruction includes a speed instruction, a position instruction and a timing instruction of the actuator; and the control instruction is sent to the corresponding edge control unit.

[0015] This application has the following beneficial technical effects:

[0016] (1) By collecting the differential pressure, valve disc angle, sealing surface stress and vibration characteristics of the fluid through multi-source sensors, and combining the digital twin model to update the dynamic characteristic data in real time, high-precision monitoring and prediction of backflow and water hammer risks can be achieved, significantly improving the operational reliability of the backflow valve. (2) Based on the warning signal and dynamic characteristic data, the adaptive control engine can dynamically adjust the valve disc closing parameters and generate the optimal closing curve, thereby providing accurate and fast response under different risk levels, reducing the wear of the valve body and pipeline, and extending the service life of the equipment. (3) The intelligent analysis and warning engine analyzes the valve disc action and fluid state characteristics through deep learning algorithms, matches them with the fault mode library, identifies the fault type and risk level in real time, and provides early warning signals to effectively avoid safety hazards caused by sudden failures. (4) Through the collaborative work of the edge control unit and the central processing unit, combined with the communication network, distributed management of multiple backflow valves is achieved, enhancing the flexibility and scalability of the system and adapting to complex industrial application environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic diagram of a reverse flow valve based on digital twin technology early warning control provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0018] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0019] The first embodiment of this application provides a reverse flow valve based on early warning control of digital twin technology. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1 The first embodiment of the present application provides a detailed description of a reverse flow valve based on early warning control of digital twin technology.

[0020] The ergonomic keyboard with automatic height and angle adjustment includes a valve body 101 , a multi-source sensor 102 , an electric actuator 103 , an edge control unit 104 and a central processing unit 105 .

[0021] The valve body 101 is provided with a valve flap assembly for controlling the unidirectional flow of the fluid. The valve flap assembly includes a valve flap and a valve seat.

[0022] The valve body 101 is the core structural component of a reverse flow valve using digital twin technology for early warning control. It contains fluid and provides a working environment for internal mechanical components (such as the disc assembly). It also serves as a platform for the installation and operation of the multi-source sensor 102, electric actuator 103, edge control unit 104, and their supporting systems. The valve body's outer shell is typically made of high-strength, corrosion-resistant materials such as stainless steel or special alloys to ensure long-term stability and reliability in high-pressure, high-temperature, and corrosive fluid environments.

[0023] The valve body's internal cavity features a precisely designed fluid channel whose geometry supports unidirectional flow and optimizes fluid dynamics. The channel's inlet and outlet are located at opposite ends of the valve body, and the channel's cross-sectional design is optimized based on the fluid's flow range and system pressure characteristics to minimize fluid resistance and turbulence. The valve body's mounting interfaces typically include flanged, threaded, or welded connections to meet the connection requirements of various piping systems.

[0024] A disc assembly, consisting of a disc and a seat, is installed within the valve body, providing one-way control of the fluid. The seat, typically fixed to the downstream port of the valve body, provides a sealing surface that forms an effective seal with the disc to prevent reverse flow. The seat's sealing surface is typically made of wear-resistant materials, such as hardened stainless steel or coated with a special ceramic layer, to extend its service life and improve erosion resistance. The disc can rotate around a fixed axis within the valve body or move along a preset linear trajectory, switching between open and closed to regulate the fluid's flow.

[0025] The outer wall of the valve body is designed with mounting holes for multiple sensors, including dedicated mounting sockets for differential pressure sensors, angle sensors, stress sensors, and vibration sensors. The locations of the mounting holes are precisely designed to ensure that each type of sensor can accurately collect data on the fluid flow state and the operating status of the mechanical structure. For example, a differential pressure sensor is usually installed on the channel wall near the inlet and outlet to capture the upstream and downstream fluid pressure difference; an angle sensor is fastened near the valve disc bearing to monitor the rotation angle of the valve disc in real time; a stress sensor is embedded in the valve seat area to detect the stress state of the sealing surface; and a vibration sensor is placed in a key position on the valve body shell to capture the impact characteristics of water hammer.

[0026] The top or side of the valve body also features an interface for installing an electric actuator, typically including a mechanical mount and a cable outlet. This interface allows the electric actuator to directly drive the disc assembly, ensuring rapid closure in abnormal situations and effective fluid control.

[0027] All connections and components inside and outside the valve body must be designed to comply with relevant industrial standards and undergo precise calibration and sealing to ensure that the overall performance of the valve body meets the expected functional requirements and can withstand the operating environment under actual working conditions.

[0028] The multi-source sensor 102 is arranged on the valve body. The multi-source sensor includes a differential pressure sensor arranged before and after the valve disc, which is used to detect the flow direction and pressure difference of the fluid; an angle sensor arranged on the valve disc, which is used to detect the opening angle of the valve disc; a stress sensor arranged at the valve seat, which is used to detect the stress state of the valve seat sealing surface; and a vibration sensor arranged on the valve body, which is used to detect the water hammer impact characteristics.

[0029] The multi-source sensor 102 is a key component in this embodiment for collecting real-time data on the reverse flow valve's operating status. Its purpose is to comprehensively monitor the dynamic characteristics of the fluid and the mechanical behavior of the valve body through the coordinated operation of multiple sensor types, providing basic data support for subsequent analysis and control. The placement, function, and operating characteristics of each sensor are precisely designed to ensure comprehensive, accurate, and real-time data collection.

[0030] Differential pressure sensors are installed on the upstream and downstream channel walls of the valve disc assembly, measuring the fluid pressure at the inlet and outlet, respectively, to calculate the pressure difference and flow direction. Differential pressure sensors typically use highly sensitive, high-pressure capacitive or piezoresistive sensors. They are secured to the valve body via a sealing flange or threaded interface and protected from direct fluid contact by a corrosion-resistant dielectric diaphragm. The collected pressure signal is output in real time to the edge control unit via an internal electrical signal conversion module, which monitors the flow direction and abnormal pressure fluctuations.

[0031] An angle sensor, mounted securely near the valve disc's rotating shaft or linear motion mechanism, monitors the disc's opening angle in real time. This sensor typically employs a non-contact Hall-effect sensor or optical encoder, and its mounting angle is precisely calibrated to ensure the data collected accurately reflects the disc's dynamic position. The sensor's output angle signal is used to analyze the disc's operating status, such as its opening amplitude and closing speed, thereby supporting precise control of the disc's movement.

[0032] Stress sensors are embedded in the sealing surface of the valve seat or in the nearby supporting structure to monitor the stress state of the sealing surface in real time. These sensors typically utilize thin-film resistance strain gauges or piezoelectric stress sensors, and require an embedded structure to minimize the impact of fluid impact or mechanical interference on the data. These stress sensors can capture the stress distribution on the sealing surface under different operating conditions, enabling timely detection of any degradation in sealing performance due to fluid anomalies or structural wear.

[0033] Vibration sensors are placed at key locations on the valve body, such as near the disc support bearing or at the fluid impact point, to detect water hammer signatures. These sensors typically use high-frequency piezoelectric or MEMS vibration sensors. By sampling the valve body's vibration spectrum, they can identify potential water hammer events. The sensors must be in close contact with the valve body to avoid attenuation or interference with the vibration signal.

[0034] Multi-source sensors transmit collected signals to the edge control unit in real time via pre-defined communication interfaces. The sensor installation location, orientation, and connection method must adhere to standardized design principles to ensure the stability and consistency of sensor data.

[0035] The electric actuator 103 is provided on the valve body and is used to quickly close the valve disc when reverse flow is detected.

[0036] Electric actuator 103 is a key component in the reverse flow valve, enabling rapid disc actuation. Its primary function is to rapidly close the disc when reverse flow or water hammer is detected, effectively protecting the pipeline system from damage. This mechanism must be designed to meet the requirements of high response speed, high control accuracy, and long-term operational reliability.

[0037] An electric actuator typically consists of a motor, a speed reducer, a drive rod, a position feedback device, and its control circuitry. The motor can be a DC servo motor or a stepper motor to ensure sufficient torque and high-precision position control under complex operating conditions. The motor is connected to the drive rod via a speed reducer, which converts the motor's high-speed rotational motion into low-speed, high-torque motion suitable for driving the valve disc, ensuring the actuator can quickly complete the transition from fully open to fully closed or vice versa.

[0038] The drive rod is a high-strength metal rod, one end of which is connected to the motor output shaft. The other end directly applies torque to the valve disc or transmits torque to the disc through a fulcrum structure. The length and stroke of the drive rod are precisely designed based on the valve body size and the opening and closing angle of the valve disc to ensure displacement accuracy and reliability with each movement.

[0039] The position feedback device uses a sensor (such as an encoder or Hall effect sensor) to monitor the displacement or angle of the actuator rod in real time, providing the control unit with information on the current position of the valve disc. This feedback data is then compared with the control instructions in a closed-loop manner to ensure that the actuator's movements always occur according to the predetermined speed, position, and timing. The accuracy of the position feedback device directly affects the precision of the actuator's movements, so its selection must meet the requirements of high resolution and low latency.

[0040] The control circuitry of an electric actuator consists of a driver module and a signal processing module. The driver module receives speed, position, and timing commands from the edge control unit, precisely controlling the motor's speed, direction, and stop position. The signal processing module integrates signals from the position feedback device and transmits them to the edge control unit to update the valve disc's real-time status. The control circuit typically integrates overload protection and abnormality self-detection functions to enhance the actuator's operational safety.

[0041] The actuator is typically mounted externally to the valve body and secured to it via mechanical interfaces (such as flanges, bolts, or locking mechanisms). During installation, ensure that the direction of motion of the drive rod is perfectly aligned with the direction of movement of the valve disc to prevent mechanical jamming or error accumulation during operation. To protect the actuator from environmental influences during operation, its housing must be waterproof, dustproof, and corrosion-resistant, and meet relevant industrial safety standards (such as IP rating).

[0042] The above design and installation ensures efficient and reliable operation of the electric actuator under complex operating conditions. Its coordinated operation with the edge control unit provides a solid guarantee for rapid valve closure and safe operation of the pipeline system. Based on actual application requirements, motor power, drive rod stroke, and control parameters can be flexibly adjusted to suit the operating conditions of different pipeline systems.

[0043] The edge control unit 104 is installed on the valve body and is used to receive the reverse flow valve operation data collected by the multi-source sensor; monitor the reverse flow status and water hammer risk through the reverse flow valve operation data, and perform rapid valve closing control in an emergency; transmit the reverse flow valve operation data collected by the multi-source sensor to the central processing unit; receive the control instructions sent by the central processing unit, and adjust the reverse flow valve through the electric actuator according to the control instructions.

[0044] The edge control unit 104 is the core component of the reverse flow valve, responsible for real-time data processing and rapid response control. Its primary function is to receive operational data collected by the multi-source sensors 102, analyze this data to determine reverse flow conditions and water hammer risks, and directly trigger rapid valve closing in emergency situations. The edge control unit also interacts with the central processing unit 105, uploading locally collected and processed data and receiving control commands from the central processing unit to implement higher-level optimization and regulation.

[0045] Edge control units typically include a high-performance embedded processor, a data storage module, a communication module, and an interface circuit. The embedded processor possesses real-time multitasking capabilities and can quickly complete logical operations from sensor data acquisition to valve closing. The raw signals transmitted by multiple sensors are filtered and signal conditioned by the interface circuit before entering the processor for data fusion and analysis. The processor uses built-in analysis algorithms to determine whether there is a risk of backflow or water hammer. For example, the processor calculates whether the fluid flow direction is abnormal based on differential pressure sensor data, identifies water hammer impact characteristics based on vibration sensor data, and evaluates the working status of the valve disc and sealing surface based on data from angle sensors and stress sensors.

[0046] In emergency situations, such as when significant water hammer or a serious backflow risk is detected, the edge control unit can independently execute a rapid valve closing operation without waiting for instructions from the central processing unit. This rapid response capability is achieved by the processor directly generating a valve closing command and issuing it to the electric actuator 103. The valve closing command includes parameter settings for the valve disc closing speed and action timing, ensuring accurate and timely operation.

[0047] The edge control unit's communication module supports a variety of industrial communication protocols, such as Modbus, CAN bus, and Industrial Ethernet, establishing a reliable communication link with the central processing unit. Through this communication module, the edge control unit can transmit collected operational data (including sensor data such as differential pressure, vibration, and stress) to the central processing unit for subsequent digital twin model updates and global optimization analysis. The edge control unit is also responsible for receiving control commands from the central processing unit. These commands may include optimized valve disc closing speed curves or adjustment strategies, and converting them into specific action parameters for the electric actuator.

[0048] To enhance the reliability of the edge control unit, its hardware design takes into account factors such as high temperature, high humidity, and strong electromagnetic interference found in industrial environments. The hardware circuitry features a reinforced design, and the housing is waterproof and dustproof, meeting IP65 or higher protection requirements. Furthermore, the edge control unit supports software fault tolerance and hardware redundancy to ensure normal operation even in the event of a single point of failure.

[0049] Through the above design, the edge control unit realizes real-time data processing, local rapid response and remote collaborative control functions.

[0050] The central processing unit 105 is connected to one or more edge control units of the reverse flow valve via a communication network, and includes:

[0051] a digital twin engine configured to receive operating data transmitted by the edge control unit; establish a digital twin model of the reverse flow valve based on the operating data; continuously update the digital twin model based on real-time operating data to generate dynamic characteristic data including valve disc opening characteristics, fluid pressure distribution, and water hammer characteristics; and transmit the dynamic characteristic data to an intelligent analysis and early warning engine;

[0052] An intelligent analysis and warning engine is configured to receive the dynamic characteristic data, analyze the valve disc motion characteristics and fluid state characteristics through a deep learning algorithm, identify the risk of reverse flow and water hammer, match the identified risk characteristics with a preset fault mode library, determine the risk level and development trend, generate a warning signal including the risk type and level, and transmit the warning signal to the adaptive control engine;

[0053] An adaptive control engine is used to receive the dynamic characteristic data and early warning signal; based on a preset control strategy library, a basic control strategy is selected according to the risk type and level in the early warning signal, wherein the control strategy library stores corresponding valve flap closing parameter templates for different risk levels; in combination with the fluid pressure distribution and water hammer characteristics in the dynamic characteristic data, the closing parameters in the basic control strategy are dynamically optimized to generate an optimal valve flap closing curve, wherein the closing curve defines the segmented speed and timing of the valve flap from starting to close to fully close; the closing curve is converted into a control instruction of the actuator, wherein the control instruction includes a speed instruction, a position instruction and a timing instruction of the actuator; and the control instruction is sent to the corresponding edge control unit.

[0054] The central processing unit 105 is the core component responsible for global optimization analysis and control coordination in this embodiment. Its main functions are to receive data from multiple edge control units, build and update the digital twin model in real time, conduct in-depth analysis of the operating status of the reverse flow valve, and generate optimized control instructions and send them to the corresponding edge control units to achieve dynamic optimization control on a global scale.

[0055] The central processing unit (CPU) typically consists of a high-performance computing platform, including a multi-core processor, a large-capacity storage module, and a high-bandwidth communication interface. This unit connects to multiple edge control units via a communication network, enabling high-speed, bidirectional data transmission. The CPU receives operational data from the edge control units, including fluid differential pressure, vibration characteristics, valve disc angle, and stress distribution. After entering the CPU through the communication interface, this data is automatically categorized and stored in a database, providing data support for subsequent analysis and modeling.

[0056] The digital twin engine, a core functional module of the central processing unit, is used to construct a digital twin model of the reverse flow valve based on real-time operating data. This model, based on the actual valve body and fluid environment, recreates the dynamic characteristics of valve operation through mathematical modeling and physical simulation, including disc opening variations, fluid pressure distribution, and water hammer characteristics. The digital twin model is continuously updated during operation, with new data from the edge control unit automatically incorporated into the model to maintain a highly accurate representation of actual operating conditions.

[0057] The digital twin model's inputs can include a variety of sensor data transmitted in real time from edge control units, such as upstream and downstream pressure differentials from differential pressure sensors, valve disc opening from angle sensors, sealing surface force distribution from stress sensors, and water hammer characteristics from vibration sensors. This data forms the foundation for building and updating the model, reflecting the critical characteristics of the reverse flow valve under different operating conditions.

[0058] To build a digital twin model, a combination of physics-based modeling and data-driven approaches can be employed. First, a physical simulation model is established based on the design parameters of the reverse flow valve (such as valve body geometry, material properties, and fluid properties). For example, fluid dynamics software (such as ANSYS Fluent or COMSOL) can be used to simulate the pressure distribution and flow state of the fluid flowing through the valve, calculating the forces acting on the valve disc at different openings and the transient pressure fluctuations caused by water hammer.

[0059] Based on the physical model, sensor data is used for calibration and optimization. Specifically, real-time data can be fed into the physical model to adjust the model's boundary conditions or material parameters, making the simulation results more closely resemble actual operating conditions. For example, if sensors detect an increase in upstream pressure and a rapid decrease in downstream pressure, the model can dynamically adjust the fluid density and flow rate to simulate possible water hammer.

[0060] The output of a digital twin model typically includes dynamic property data, such as:

[0061] The real-time opening of the valve disc and the speed curve of its closing process;

[0062] The spatiotemporal distribution of fluid pressure, including upstream and downstream pressures and local pressure on the sealing surface;

[0063] Vibration characteristics, such as frequency and amplitude of water hammer impacts;

[0064] Predicted failure characteristics, such as abnormal sealing surface forces or hysteresis in valve disc movement.

[0065] As an auxiliary module of the digital twin engine, the intelligent analysis and warning engine receives and analyzes the dynamic characteristic data generated by the digital twin model. Using deep learning algorithms, this engine comprehensively analyzes the valve disc's operating characteristics, fluid flow conditions, and other key parameters to identify potential risk characteristics, such as backflow risk, water hammer impact, and abnormal sealing surface stress. These identified risk characteristics are matched against a pre-defined fault pattern library to determine their type, risk level, and potential development trend. The analysis results are output as warning signals, which provide input to the adaptive control engine.

[0066] To implement the deep learning algorithms in the intelligent analysis and early warning engine, it is necessary to extract key parameters from the dynamic characteristic data generated by the digital twin model and use this data to train a model capable of identifying risk characteristics. The following is a simple and feasible implementation.

[0067] Deep learning algorithms are fed with dynamic characteristic data, typically including the real-time valve disc opening, upstream and downstream pressure differentials, vibration frequency, and sealing surface stress. This data is fed into the algorithm in a time series format to capture dynamic changes and patterns. For example, multiple data points within a specific time window can be used to reflect current fluid conditions and mechanical behavior.

[0068] The algorithm outputs an analysis of risk characteristics, including risk type (e.g., reverse flow, water hammer, or seal failure), risk level (e.g., low or high), and possible development trends. These outputs are used directly to generate early warning signals for use by the adaptive control engine.

[0069] To implement this algorithm, a deep learning model suitable for time series data analysis can be selected, such as a long short-term memory network (LSTM) or a convolutional neural network (CNN). These models can effectively capture key patterns in time series and are suitable for analyzing the dynamic characteristics of valve operation.

[0070] The first step in building a model is data preprocessing. Dynamic feature data is organized into time series, with each time step containing a set of feature values. Next, the data needs to be normalized to ensure a consistent range for model training. Next, the historical data is divided into training and validation sets for model learning and optimization.

[0071] The model training process requires data labeling. This involves labeling each set of input data with its corresponding risk characteristics. For example, within known operational data, the occurrence of reverse flow risk or water hammer could be marked within a certain period. Through these annotations, the model can learn the relationship between input data and risk.

[0072] Once the model is trained, it can be deployed in an intelligent analysis and early warning engine, which receives input data generated by the digital twin model in real time and outputs risk analysis results. For example, if it detects rapid changes in valve disc opening and sharp fluctuations in upstream and downstream pressure differences, the model may identify a high water hammer risk and issue a corresponding early warning signal.

[0073] Furthermore, the appropriate model size and training data volume can be selected based on the complexity of the reverse flow valve system. During actual operation, the model can be continuously optimized, and the accuracy and reliability of risk identification can be further enhanced through the addition of new operational data. This allows the intelligent analysis and early warning engine to achieve efficient and accurate risk assessment, providing reliable support for the adaptive control engine.

[0074] Furthermore, the deep learning algorithm of the intelligent analysis and early warning engine is implemented using a pre-trained deep learning model, which includes a multi-feature decoupling module, a cross-feature correlation module, a spatiotemporal sequence prediction module, and a risk grading and trend prediction module;

[0075] Among them, the multi-characteristic decoupling module is used to receive multi-source dynamic characteristic data from the digital twin model, and the multi-source dynamic characteristic data includes the valve flap opening change curve, the upstream and downstream pressure difference of the fluid, the sealing surface stress distribution and the vibration frequency characteristics; the output of the multi-characteristic decoupling module is an independent high-dimensional characteristic vector for each characteristic, specifically including outputting the pressure fluctuation characteristics for the pressure difference data and the vibration spectrum characteristics after Fourier transformation for the vibration frequency; outputting multi-scale time domain characteristics for the valve flap opening and acceleration data; the multi-characteristic decoupling module uses a convolutional network to extract the local fluctuation characteristics of the pressure difference signal, uses a Fourier domain network to extract the spectrum characteristics of the vibration frequency, and uses multi-scale analysis to extract the acceleration and angle change patterns of the valve flap action;

[0076] The input of the cross-characteristic association module is the pressure fluctuation characteristics, vibration spectrum characteristics, and multi-scale time domain characteristics; the output of the cross-characteristic association module is a characteristic association vector and a dynamic evolution vector; the cross-characteristic association module is implemented using a multi-characteristic interaction model based on a graph neural network, performs correlation analysis on the decoupled characteristics, and dynamically models the coupling relationship between the characteristics;

[0077] The input of the spatiotemporal sequence prediction module is the characteristic association vector and the dynamic evolution vector. The output of the spatiotemporal sequence prediction module is the predicted characteristic trend in the time series, which includes the prediction of the change of water hammer impact intensity, the prediction of the offset of sealing surface stress, and the possibility of valve disc action hysteresis. The spatiotemporal sequence prediction module is implemented using a recursive network.

[0078] The input of the risk grading and trend prediction module is the predicted characteristic trend; the risk grading and trend prediction module uses a classification regression network to perform risk grading and trend prediction on the input data, wherein the classification branch generates risk levels including backflow risk, water hammer impact and sealing abnormality according to characteristic weights and development trends; the regression branch predicts the risk intensity and possible change trends, including increased water hammer impact frequency or sealing stress offset.

[0079] The intelligent analysis and early warning engine is implemented through a pre-trained model based on a deep learning algorithm. The model consists of multiple modules that work together to analyze complex dynamic characteristics and predict risks. First, the multi-characteristic decoupling module receives multi-source dynamic characteristic data from the digital twin model, such as the valve flap opening change curve, the upstream and downstream pressure difference of the fluid, the sealing surface stress distribution, and the vibration frequency characteristics. This module extracts the pressure fluctuation characteristics through a convolutional network, the vibration spectrum characteristics through a Fourier domain network, and the acceleration and angle change patterns during the valve flap movement through multi-scale analysis, thereby outputting independent high-dimensional characteristic vectors for each characteristic.

[0080] The cross-feature correlation module then performs correlation analysis on these decoupled feature vectors. Using a multi-feature interaction model based on a graph neural network, this module dynamically models the coupling relationships between features, generating feature correlation vectors and dynamic evolution vectors. This module reveals complex interaction patterns between different features, such as the correlation between pressure fluctuations and vibration characteristics.

[0081] Next, the spatiotemporal series prediction module receives these characteristic correlation vectors and dynamic evolution vectors and uses a recursive network to analyze characteristic trends within the time series, including changes in water hammer intensity, the offset direction of sealing surface stress, and potential hysteresis in valve disc actuation. These predicted trends provide an important foundation for further risk assessment.

[0082] Finally, the risk grading and trend prediction module analyzes the predicted characteristic trends and uses a classification and regression network to perform risk grading and trend prediction on the input data. The classification branch generates risk levels for backflow, water hammer, and seal anomalies based on characteristic weights and trends. The regression branch predicts the intensity of the risk and possible development trends, such as an increase in water hammer frequency or further shift in seal stress.

[0083] Through the intelligent analysis and early warning engine, the present invention can comprehensively and efficiently identify and predict the operating risks of the reverse flow valve, and provide accurate risk signals and adjustment suggestions for subsequent control modules.

[0084] Furthermore, the spatiotemporal sequence prediction module is specifically used to:

[0085] Receive the feature correlation vector and dynamic evolution vector from the cross-feature correlation module, group these input data into multi-dimensional time series, and dynamically adjust the grouping method according to the time correlation in the feature correlation vector;

[0086] A recursive network is used to process the grouped time series data. In the first stage of the recursive network, a unidirectional recursive unit is used to extract short-term dynamic characteristic patterns, including pressure change trends, vibration spectrum changes, and stress offset rates.

[0087] In the second stage, the short-term and long-term dynamic characteristics are combined through a bidirectional recursive unit to capture the potential causal relationship between pressure and vibration characteristics;

[0088] Use the temporal attention mechanism to assign weights to different time steps and dynamically enhance attention to key time points, such as abnormal enhancement of vibration spectrum or rapid fluctuation of pressure difference;

[0089] The output is the predicted characteristic trend in the time series, including the changing trend of water hammer impact intensity, the offset direction of sealing surface stress and the time range of possible lag of valve disc action.

[0090] The spatiotemporal series prediction module of the present invention is designed to perform efficient time series analysis and trend forecasting on complex dynamic characteristic data. This module receives characteristic correlation vectors and dynamic evolution vectors from the cross-characteristic correlation module and groups these input data into multidimensional time series based on temporal correlation. This grouping can be dynamically adjusted to accommodate changes in the temporal dependencies between different characteristics in the characteristic correlation vectors, thereby ensuring accurate analysis.

[0091] The module processes the grouped time series data using a recurrent network. In the first stage, a unidirectional recurrent unit extracts short-term dynamic patterns, identifying pressure trends, local variations in the vibration spectrum, and the rate of stress excursion. This stage focuses on capturing rapid changes and localized characteristics.

[0092] In the second stage, the bidirectional recursive unit combines short-term and long-term dynamic characteristics, further capturing potential causal relationships between characteristics by analyzing the contextual information in the time series. For example, the synchronization of pressure changes and vibration spectra can be analyzed to assess the causes of water hammer risk.

[0093] Furthermore, the module integrates a temporal attention mechanism that assigns dynamic weights to different time steps, enhancing focus on key time points. This mechanism can identify important characteristic changes, such as abnormal increases in the vibration spectrum or rapid fluctuations in pressure differential, and highlight the importance of this key information for subsequent analysis.

[0094] Ultimately, the spatiotemporal series prediction module outputs the predicted characteristic trends within the time series, including the changing trend of water hammer intensity, the offset direction of sealing surface stress, and the time range of possible valve disc actuation delay. These outputs provide accurate data support for risk assessment and control decisions.

[0095] Furthermore, the risk grading and trend prediction module includes functions specifically for:

[0096] Receive the predicted characteristic trend data from the spatiotemporal series prediction module and first assign importance weights to each characteristic through the characteristic weight adjustment module. Specifically, the weights of vibration spectrum changes and pressure fluctuations are increased to highlight the analysis of water hammer risk.

[0097] The classification branch of the classification regression network is used to perform risk classification on the adjusted input data. The predicted trend data and characteristic weights are combined to generate risk types and their levels, including backflow risk, water hammer impact, and sealing anomaly.

[0098] The regression branch of the classification regression network is used to predict the trend of the input data. For each risk type, a numerical prediction of the risk intensity and its development trend are output, such as the gradual increase in water hammer impact frequency or the expansion of the seal stress excursion range.

[0099] The output includes comprehensive prediction results of risk type, risk level and development trend, and generates early warning signals for adaptive control engine.

[0100] The Risk Grading and Trend Prediction module conducts in-depth analysis of predicted characteristic trend data and provides detailed risk assessment and trend prediction results. The module first receives predicted characteristic trend data from the spatiotemporal series prediction module and uses the characteristic weighting module to assign importance weights to each characteristic. For example, when analyzing water hammer risk, the module appropriately increases the weights for vibration spectrum changes and pressure fluctuations based on actual operating conditions to more accurately capture relevant risk characteristics.

[0101] The adjusted data is then fed into the classification branch of the classification regression network for risk stratification. The classification branch combines the predicted trend data with characteristic weights to generate risk categories, including reverse flow risk, water hammer impact, and seal anomaly. Each risk category is assigned a risk level (e.g., low, medium, or high). This process ensures clear identification of different risk types and provides a clear priority for subsequent control decisions.

[0102] The module then uses the regression branch of a classification regression network to predict trends in the input data. This branch generates specific numerical predictions for each risk type, such as the frequency and amplitude of water hammer impacts, and the trend of increasing seal stress excursions. These numerical predictions further refine the risk assessment and provide valuable input data for precise control.

[0103] The module's final output includes comprehensive predictions of risk type, risk level, and development trend. These results are further converted into early warning signals for the adaptive control engine to guide subsequent control strategy adjustments.

[0104] Furthermore, the multi-characteristic decoupling module extracts the local fluctuation characteristics of the pressure difference signal through the following formula 1:

[0105] ;

[0106] in, Represents the total number of convolution kernels used to extract local fluctuation characteristics;

[0107] In time The pressure fluctuation characteristic vector, which represents the pressure fluctuation at time The pressure fluctuation characteristic vector at is the final extracted characteristic output, used to describe the local fluctuation pattern of the pressure difference signal. Each element of this vector corresponds to an extracted characteristic value, which serves as the output of the multi-characteristic decoupling module and is used for analysis in subsequent modules.

[0108] It's in time The first one in the pressure difference signal sequence input at The specific value comes from the upstream and downstream pressure difference signal collected by the sensor. , As input signal, by time window Divide and process.

[0109] is the time window length, which is used to capture the short-term fluctuation characteristics of the signal; the recommended value is determined by the signal sampling frequency. For example, when the sampling frequency is 100 Hz, it is recommended to take =0.1 seconds.

[0110] is the Gaussian weighted kernel function; is the convolution kernel weight; is the periodic weight function;

[0111] is the variance of the pressure difference signal within the time window, which is used to enhance the sensitivity to local abnormal changes;

[0112] Variance gain weight, used to control the contribution of the variance term. The recommended value is 0.5 to 2.0, which should be adjusted based on the noise level of the signal and the requirement for abnormal sensitivity.

[0113] It is a bias term used to balance the output. It is recommended to initialize it to zero or set it according to the data distribution.

[0114] The Gaussian weighted kernel function is implemented using the following formula 2:

[0115] ;

[0116] in, It's time The signal value at is taken out from the sequence.

[0117] is the convolution kernel weight, which indicates the sensitivity to specific signal patterns. A Gaussian random distribution is recommended as the initial value, which can be optimized through training.

[0118] The width of the Gaussian kernel determines the kernel function's sensitivity to signal deviations. The recommended value range is 0.1 to 1.0, adjusted based on the signal noise level.

[0119] The periodic weight function is implemented using the following formula 3:

[0120] ;

[0121] in, is the frequency parameter that controls the oscillation frequency of the period weight. The recommended value is a multiple of the main frequency component of the signal, for example Indicates a frequency of 50 Hz.

[0122] is the phase parameter, used to adjust the phase offset of the weight function. It is recommended to initialize it to a random distribution.

[0123] Furthermore, the cross-feature association module models the dynamic coupling relationship between features through the following formula 4:

[0124] ;

[0125] in, Representation characteristics and features In time The correlation weight is used to describe the coupling strength between two features at a specific point in time. This value is directly calculated by the formula and is the output of the cross-feature correlation module. It is used for feature fusion and dynamic evolution analysis in subsequent modules.

[0126] The number of frequency layers representing the coupling between features is the number of frequency components in the formula. Each frequency layer corresponds to a specific coupling analysis range. For example, a low-frequency layer can capture slow changes between features, while a high-frequency layer can capture rapid fluctuations. The recommended value is 3 to 5, depending on system complexity.

[0127] For time Lower characteristics and No. Layer coupling strength represents the weights of different frequency layers. Its initial value is usually determined by training data. For example, it can be initialized to a normally distributed random value based on the frequency distribution of historical characteristic coupling relationships, and then optimized through training.

[0128] Representation characteristics and No. Layer-specific frequency components are used to capture the periodic relationship between characteristics. For example, if the system signal is mainly distributed in the low frequency range, ,in is the fundamental frequency (recommended value 10 Hz).

[0129] For characteristics and In time The time offset is defined as:

[0130] ;

[0131] in, It is a feature signal delay or peak occurrence time. It is a feature The signal delay or peak occurrence time of the signal can be obtained by real-time signal processing methods such as peak detection.

[0132] is a small positive number used to avoid the denominator being zero;

[0133] For characteristics In time characteristic value of For characteristics In time These values ​​are directly output from the previous module (such as the multi-feature decoupling module) and represent the real-time data of the feature at a specific point in time.

[0134] The feature difference weight is used to amplify nonlinear relationships between features. Its value indicates the sensitivity to feature differences. The recommended range is 0.1 to 1.0, adjusted based on the complexity of the system's features. The initial value can be set to 0.5 and updated during training.

[0135] The following is the reference implementation code of the deep learning model:

[0136] import torch

[0137] import torch.nn as nn

[0138] import torch.optim as optim

[0139] import numpy as np

[0140] # Define multi-feature decoupling modules

[0141] class MultiFeatureDecomposition(nn.Module):

[0142] def __init__(self, input_dim, num_kernels, kernel_width):

[0143] super(MultiFeatureDecomposition, self).__init__()

[0144] self.num_kernels = num_kernels

[0145] self.kernel_width = kernel_width

[0146] self.weights = nn.Parameter(torch.randn(num_kernels, input_dim))

[0147] self.bias = nn.Parameter(torch.zeros(num_kernels))

[0148] self.sigma = nn.Parameter(torch.ones(num_kernels) kernel_width)

[0149] self.frequency_params = nn.Parameter(torch.ones(num_kernels) 2 np.pi)

[0150] self.phase_params = nn.Parameter(torch.zeros(num_kernels))

[0151] def forward(self, x):

[0152] # Gaussian weighted kernel function

[0153] gauss_kernel = torch.exp(-torch.square(x.unsqueeze(1) -self.weights) / (2 torch.square(self.sigma)))

[0154] # Periodic weight function

[0155] time_indices = torch.arange(x.size(1), device=x.device).float().unsqueeze(0)

[0156] periodic_weights = torch.cos(self.frequency_params.unsqueeze(1) time_indices + self.phase_params.unsqueeze(1))

[0157] # Weighting of pressure fluctuation characteristics

[0158] weighted_output = torch.sum(gauss_kernel periodic_weights,dim=2) + self.bias.unsqueeze(0)

[0159] # Variance enhancement

[0160] variance = torch.var(x, dim=1, unbiased=False).unsqueeze(1)

[0161] enhanced_output = weighted_output + 0.5 variance

[0162] return enhanced_output

[0163] # Define the cross - feature association module

[0164] class CrossFeatureAssociation(nn.Module):

[0165] def __init__(self, num_features, num_frequency_layers):

[0166] super(CrossFeatureAssociation, self).__init__()

[0167] self.num_frequency_layers = num_frequency_layers

[0168] self.alpha = nn.Parameter(torch.randn(num_features, num_features, num_frequency_layers))

[0169] self.frequency = nn.Parameter(torch.linspace(0.1, 1.0, num_frequency_layers))

[0170] self.lambda_weights = nn.Parameter(torch.ones(num_features,num_features, num_frequency_layers))

[0171] def forward(self, y_i, y_j, time_deltas):

[0172] time_deltas = time_deltas.unsqueeze(-1).expand(-1, -1,self.num_frequency_layers)

[0173] sine_term = torch.sin(np.pi self.frequency time_deltas) / (np.pi self.frequency time_deltas + 1e-6)

[0174] nonlinear_difference = torch.log(1 + torch.abs(y_i.unsqueeze(2) - y_j.unsqueeze(2)))

[0175] association = torch.sum(self.alpha sine_term (1 +self.lambda_weights nonlinear_difference), dim=2)

[0176] return association

[0177] # Define the spatio-temporal sequence prediction module

[0178] class SpatioTemporalPrediction(nn.Module):

[0179] def __init__(self, input_dim, hidden_dim):

[0180] super(SpatioTemporalPrediction, self).__init__()

[0181] self.rnn1 = nn.LSTM(input_dim, hidden_dim, batch_first=True)

[0182] self.rnn2 = nn.LSTM(hidden_dim, hidden_dim, batch_first=True,bidirectional=True)

[0183] self.attention_weights = nn.Linear(hidden_dim 2, 1)

[0184] def forward(self, x):

[0185] rnn1_output, _ = self.rnn1(x)

[0186] rnn2_output, _ = self.rnn2(rnn1_output)

[0187] attention_scores = torch.softmax(self.attention_weights(rnn2_output), dim=1)

[0188] weighted_output = torch.sum(rnn2_output attention_scores,dim=1)

[0189] return weighted_output

[0190] # Define the risk classification and trend prediction module

[0191] class RiskClassificationAndTrendPrediction(nn.Module):

[0192] def __init__(self, input_dim, num_classes):

[0193] super(RiskClassificationAndTrendPrediction, self).__init__()

[0194] self.feature_weight_adjustment = nn.Linear(input_dim, input_dim)

[0195] self.classification_branch = nn.Linear(input_dim, num_classes)

[0196] self.regression_branch = nn.Linear(input_dim, 1)

[0197] def forward(self, x):

[0198] adjusted_features = torch.relu(self.feature_weight_adjustment(x))

[0199] classifications = torch.softmax(self.classification_branch(adjusted_features), dim=1)

[0200] trends = self.regression_branch(adjusted_features)

[0201] return classifications, trends

[0202] # Integrated model implementation

[0203] class IntelligentAnalysisEngine(nn.Module):

[0204] def __init__(self, input_dim, num_kernels, kernel_width, num_frequency_layers, hidden_dim, num_classes):

[0205] super(IntelligentAnalysisEngine, self).__init__()

[0206] self.decomposition = MultiFeatureDecomposition(input_dim, num_kernels, kernel_width)

[0207] self.association = CrossFeatureAssociation(num_kernels, num_frequency_layers)

[0208] self.prediction = SpatioTemporalPrediction(num_kernels,hidden_dim)

[0209] self.risk_module = RiskClassificationAndTrendPrediction(hidden_dim, num_classes)

[0210] def forward(self, x, time_deltas):

[0211] decomposed_features = self.decomposition(x)

[0212] associations = self.association(decomposed_features,decomposed_features, time_deltas)

[0213] temporal_predictions = self.prediction(associations)

[0214] classifications, trends = self.risk_module(temporal_predictions)

[0215] return classifications, trends

[0216] To fully train and deploy the above deep learning model, you can follow the following process:

[0217] First, collect and process multi-source dynamic characteristic data, including valve disc opening change curves, upstream and downstream pressure differences, sealing surface stress distribution, and vibration frequency characteristics. Convert the raw data into a format suitable for input to the model, such as a time series tensor, and ensure that each characteristic data has the same time alignment and sampling rate. Standardize the data and scale the characteristic values ​​to a range with a mean of 0 and a standard deviation of 1 to improve the convergence speed and stability of the model. Next, divide the data into training, validation, and test sets. It is recommended to divide them in a ratio of 80%, 10%, and 10% to ensure that each part of the data evenly covers different operating states.

[0218] During the training phase, loss functions are defined, including the cross-entropy loss for the classification task and the mean squared error loss for the trend prediction task. A weight balancing strategy is also designed to ensure that the classification and regression tasks contribute appropriately to the total loss. The Adam optimizer is used, with an initial learning rate of 0.001. A learning rate scheduler is also implemented to dynamically adjust the learning rate to suit the model's convergence. Model training is performed in batches, with a recommended batch size of 32 to ensure a balance between training efficiency and video memory usage. After each training epoch, model performance is evaluated using a validation set, including classification accuracy, mean squared error for trend prediction, and overall loss change. Training can be stopped when validation set performance no longer significantly improves.

[0219] After training is complete, the model's generalization performance is evaluated using the test set to ensure that it performs as expected on unseen data. To deploy the model, the trained model parameters are first saved as a file. Next, the model is loaded into the deployment environment and the data preprocessing and inference processes are implemented. Real-time data is fed into the model to obtain risk classification and trend prediction results.

[0220] Finally, through the integrated deployment system, such as edge devices or servers, the model output is passed to the adaptive control module to achieve closed-loop control, and the model performance is continuously monitored and optimized during actual operation.

[0221] The adaptive control engine performs global optimization calculations based on early warning signals and dynamic characteristic data to select the optimal valve flap closing strategy. Specifically, the adaptive control engine selects a basic strategy from a preset control strategy library based on the fault type and risk level, and optimizes the closing parameters in combination with real-time data. For example, in response to severe water hammer impact, the control engine may generate a nonlinear rapid closing curve; in the case of mild backflow, it may choose to close slowly to reduce system impact. The optimized closing curve is converted into specific speed, position, and timing instructions, which are ultimately sent to the corresponding edge control unit through the communication network, which drives the electric actuator to perform the corresponding operation.

[0222] Furthermore, the adaptive control engine is specifically used to:

[0223] Receive warning signals from the intelligent analysis and warning engine, analyze risk types and risk levels, and identify priority control targets based on risk types. This includes prioritizing adjustment of valve disc closing speed for water hammer risks and adjustment of valve seat stress for sealing anomalies.

[0224] Query the preset control strategy library and select the corresponding basic control strategy according to the risk level, including slow shutdown strategy for low risk and fast shutdown strategy for high risk;

[0225] The basic control strategy is input as the initial solution into the next step of dynamic optimization process.

[0226] The adaptive control engine receives warning signals generated by the intelligent analysis and warning engine and analyzes them based on the risk type and level. First, the adaptive control engine categorizes the risk type, such as water hammer risk or sealing anomaly, and prioritizes control objectives based on this classification. For example, for water hammer risk, the control objective is to adjust the valve disc closing speed to minimize the impact of water hammer; for sealing anomaly, the control objective is to optimize the valve seat stress state to ensure sealing performance.

[0227] The adaptive control engine then retrieves a basic control strategy from a pre-defined control strategy library that is appropriate for the current risk level. For example, in low-risk situations, a slow closing strategy might be selected to reduce wear on the valve body and sealing components; in high-risk situations, a fast closing strategy might be chosen to minimize the impact of water hammer or reverse flow in emergencies. Each strategy in the control strategy library is pre-defined with specific closing speeds and action sequences, ensuring that appropriate basic control parameters are provided for different risk levels.

[0228] Finally, the adaptive control engine uses the selected basic control strategy as an initial solution and inputs it into the dynamic optimization process, where it is further optimized based on real-time data. Through these steps, the adaptive control engine can efficiently respond to early warning signals and achieve precise control of the valve.

[0229] Furthermore, the adaptive control engine optimizes the parameters in the basic control strategy according to the dynamic characteristic data, specifically including:

[0230] Analyze the fluid pressure distribution, water hammer characteristics and valve disc opening change trend in the dynamic characteristic data, and dynamically adjust the valve disc closing speed according to the pressure change rate;

[0231] Real-time adjustment of the closing curve based on the frequency and amplitude of the water hammer impact to ensure rapid closing during the low-amplitude phase of the water hammer impact and avoid pressure surges during the high-amplitude phase;

[0232] Optimize the closing sequence according to the stress state of the sealing surface to avoid excessive or uneven stress on the sealing surface caused by closing too quickly.

[0233] The adaptive control engine utilizes dynamic characteristic data to optimize the parameters of the basic control strategy for more precise valve control. First, the dynamic characteristic data, including fluid pressure distribution, water hammer characteristics, and valve flap opening trends, is used to analyze the system's current operating status. By monitoring the rate of pressure change, the adaptive control engine adjusts the valve flap's closing speed in real time, synchronizing closing with pressure fluctuations to minimize additional impact caused by pressure fluctuations.

[0234] When water hammer is detected, the adaptive control engine further analyzes the frequency and amplitude of the water hammer impact and adjusts the valve closing curve based on this information. During periods of low water hammer amplitude, the adaptive control engine prioritizes a rapid closing strategy to reduce the potential for sustained impact caused by the water hammer. During periods of high amplitude, the adaptive control engine controls the closing speed and timing to avoid further pressure surges in the system.

[0235] Furthermore, the adaptive control engine optimizes the closing sequence based on the stresses on the sealing surfaces to balance the force distribution on the sealing surfaces during closing. By preventing excessive or uneven localized force on the sealing surfaces caused by premature closing, the adaptive control engine extends the life of the sealing surfaces and improves seal reliability.

[0236] In this way, the adaptive control engine can dynamically adjust the closing parameters to make the valve operation more in line with the needs of real-time working conditions.

[0237] Furthermore, the adaptive control engine generates an optimal valve flap closing curve through a multi-objective optimization method, specifically including:

[0238] Define multi-objective optimization goals, including minimizing valve disc closing time, minimizing sealing surface forces during closing, and reducing water hammer impact energy;

[0239] Use a dynamic weight adjustment mechanism to assign weights to different optimization objectives, including prioritizing reducing shutdown time when high risk levels are detected;

[0240] The optimal closing curve is generated by an iterative optimization method, which contains the segmented speed and timing of the valve disc from the start of closing to full closing.

[0241] The adaptive control engine uses a multi-objective optimization method to generate the optimal valve flap closing curve to achieve precise control of the reverse flow valve under different risk conditions. The optimization objectives include minimizing the valve flap closing time, reducing the force on the sealing surface during closing, and reducing the impact energy of water hammer. To ensure the dynamic adaptability of the optimization objectives, the adaptive control engine uses a dynamic weight adjustment mechanism to adjust the priority of the optimization objectives according to the real-time operating conditions. For example, when a high risk level is detected, more weight will be assigned to the goal of reducing the closing time to curb the potential risk as soon as possible; in low-risk situations, more weight will be assigned to the goal of reducing the force on the sealing surface to extend the service life of the component.

[0242] During the optimization process, the adaptive control engine employs an iterative approach. By continuously adjusting parameters and validating outputs, it generates a closing curve that encompasses segmented speeds and timings from the disc's initial closing to full closure. Each segment is optimized to achieve the optimal balance of objectives within each timeframe. This iterative approach not only adapts to real-time dynamic characteristic data but also generates highly accurate control parameters for complex operating conditions.

[0243] Through the above optimization method, the adaptive control engine can generate the optimal closing curve that combines efficiency and safety.

[0244] Furthermore, the adaptive control engine converts the optimized closing curve into control instructions for the electric actuator, specifically including:

[0245] Analyze the segmented speed and timing parameters in the closing curve into speed instructions and position instructions for the electric actuator;

[0246] Add redundant check parameters to each control instruction to ensure that instruction errors caused by noise or data loss during transmission can be detected and corrected;

[0247] The parsed control instructions are sent to the edge control unit, which drives the electric actuator to perform the valve disc closing operation.

[0248] After generating the optimized closing curve, the adaptive control engine converts it into control commands that can be directly executed by the electric actuator to achieve precise closing of the valve disc. The segmented speed and timing parameters in the closing curve are key inputs, which the adaptive control engine analyzes to generate specific speed and position commands. For example, the segmented speed within a specific time period is used to set the operating speed of the electric actuator, while the timing parameters determine the position change of the valve disc at different stages.

[0249] To improve the reliability of control commands, the adaptive control engine appends redundant checksum parameters to each generated command. These parameters are used to detect potential errors during command transmission, such as those caused by communication noise or data loss. Through built-in error checking and correction mechanisms, the adaptive control engine ensures the accuracy of commands transmitted to the edge control unit.

[0250] Finally, the adaptive control engine sends the parsed control instructions to the edge control unit. Based on the received speed and position commands, the edge control unit precisely drives the electric actuator to close the valve disc. The entire process, from parsing the closing curve to completing the action, is designed to optimize efficiency and ensure safety.

[0251] The hardware and software design of the central processing unit (CPU) must consider the high reliability and efficiency requirements of industrial environments. The hardware utilizes an anti-interference design and supports redundant backups to prevent single points of failure. The software utilizes fault-tolerant algorithms and real-time task scheduling mechanisms to ensure stable system operation under complex operating conditions.

[0252] Although the present application is disclosed as above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.

Claims

1. A reverse flow valve based on early warning control of digital twin technology, characterized in that: include: A valve body, wherein a valve flap assembly is disposed in the valve body for controlling the unidirectional flow of a fluid, wherein the valve flap assembly comprises a valve flap and a valve seat; A multi-source sensor is arranged on the valve body, and the multi-source sensor includes a differential pressure sensor arranged before and after the valve disc, which is used to detect the flow direction and pressure difference of the fluid; an angle sensor arranged on the valve disc, which is used to detect the opening angle of the valve disc; a stress sensor arranged at the valve seat, which is used to detect the stress state of the valve seat sealing surface; and a vibration sensor arranged on the valve body, which is used to detect the water hammer impact characteristics; An electric actuator, disposed on the valve body, for quickly closing the valve disc when reverse flow is detected; an edge control unit, mounted on the valve body, for receiving the reverse flow valve operation data collected by the multi-source sensor; monitoring the reverse flow state and water hammer risk through the reverse flow valve operation data, and performing rapid valve closing control in an emergency; transmitting the reverse flow valve operation data collected by the multi-source sensor to the central processing unit; receiving the control command sent by the central processing unit, and adjusting the reverse flow valve through the electric actuator according to the control command; A central processing unit, the central processing unit is connected to one or more edge control units of the reverse flow valve via a communication network, comprising: A digital twin engine, configured to receive the operating data transmitted by the edge control unit; establish a digital twin model of the reverse flow valve according to the operating data; continuously update the digital twin model according to the real-time operating data to generate dynamic characteristic data including valve flap opening characteristics, fluid pressure distribution and water hammer characteristics; and transmit the dynamic characteristic data to an intelligent analysis and early warning engine; An intelligent analysis and warning engine, for receiving the dynamic characteristic data, analyzing the valve disc action characteristics and fluid state characteristics through a deep learning algorithm, identifying the risk of backflow and water hammer; generating a warning signal including the risk type and level according to the identified risk characteristics; and transmitting the warning signal to the adaptive control engine; An adaptive control engine is used to receive the dynamic characteristic data and the early warning signal; based on a preset control strategy library, a basic control strategy is selected according to the risk type and level in the early warning signal, wherein the control strategy library stores corresponding valve flap closing parameter templates for different risk levels; in combination with the fluid pressure distribution and water hammer characteristics in the dynamic characteristic data, the closing parameters in the basic control strategy are dynamically optimized to generate an optimal valve flap closing curve, wherein the closing curve defines the segmented speed and timing of the valve flap from the start of closing to the complete closing process; the closing curve is converted into a control instruction of an actuator, wherein the control instruction includes a speed instruction, a position instruction and a timing instruction of the actuator; and the control instruction is sent to a corresponding edge control unit; The deep learning algorithm of the intelligent analysis and early warning engine is implemented using a pre-trained deep learning model, which includes a multi-feature decoupling module, a cross-feature association module, a spatiotemporal sequence prediction module, and a risk grading and trend prediction module; Among them, the multi-characteristic decoupling module is used to receive multi-source dynamic characteristic data from the digital twin model, and the multi-source dynamic characteristic data includes the valve flap opening change curve, the upstream and downstream pressure difference of the fluid, the sealing surface stress distribution and the vibration frequency characteristics; the output of the multi-characteristic decoupling module is an independent high-dimensional characteristic vector for each characteristic, specifically including the output of pressure fluctuation characteristics for pressure difference data, the output of vibration spectrum characteristics after Fourier transformation for vibration frequency; and the output of multi-scale time domain characteristics for valve flap opening and acceleration data; the multi-characteristic decoupling module uses a convolutional network to extract local fluctuation characteristics for the pressure difference signal, uses a Fourier domain network to extract spectrum characteristics for the vibration frequency, and uses multi-scale analysis to extract acceleration and angle change patterns for valve flap action; The input of the cross-characteristic association module is the pressure fluctuation characteristic, the vibration spectrum characteristic and the multi-scale time domain characteristic; the output of the cross-characteristic association module is the characteristic association vector and the dynamic evolution vector; the cross-characteristic association module is implemented using a multi-characteristic interaction model based on a graph neural network, performs association analysis on the decoupled characteristics, and dynamically models the coupling relationship between the characteristics; The input of the spatiotemporal sequence prediction module is the characteristic association vector and the dynamic evolution vector, and the output of the spatiotemporal sequence prediction module is the predicted characteristic trend in the time series, which includes the change prediction of water hammer impact strength, the offset prediction of sealing surface stress, and the possibility of valve flap action lag; the spatiotemporal sequence prediction module is implemented by a recursive network; The input of the risk grading and trend prediction module is the predicted characteristic trend; the risk grading and trend prediction module uses a classification regression network to perform risk grading and trend prediction on the input data, wherein the classification branch generates risk levels including backflow risk, water hammer impact and sealing abnormality according to characteristic weights and development trends; the regression branch predicts the risk intensity and possible change trends, including increased water hammer impact frequency or sealing stress offset; The multi-characteristic decoupling module extracts the local fluctuation characteristics of the pressure difference signal through the following formula 1: ； in, Represents the total number of convolution kernels used to extract local fluctuation characteristics; In time The pressure fluctuation characteristic vector of It's in time The first one in the pressure difference signal sequence input at values; is the time window length, which is used to capture the short-term fluctuation characteristics of the signal; is the Gaussian weighted kernel function; is the convolution kernel weight; is the periodic weight function; is the variance of the pressure difference signal within the time window, which is used to enhance the sensitivity to local abnormal changes; is the variance gain weight; is a bias term used to balance the output; the Gaussian weighted kernel function is implemented using the following formula 2: ； in, is the convolution kernel weight, is the Gaussian kernel width; the periodic weight function is implemented using the following formula 3: ； in, is the frequency parameter; is the phase parameter.

2. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The adaptive control engine is specifically used for: Receive warning signals from the intelligent analysis and warning engine, analyze risk types and risk levels, and identify priority control targets based on risk types, including adjusting the valve disc closing speed for water hammer risks and adjusting the valve seat stress state for sealing anomalies; Query the preset control strategy library and select the corresponding basic control strategy according to the risk level, including slow shutdown strategy for low risk and fast shutdown strategy for high risk; The basic control strategy is used as the initial solution and input into the next step of dynamic optimization process.

3. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The adaptive control engine optimizes the parameters in the basic control strategy according to the dynamic characteristic data, specifically including: Analyze the fluid pressure distribution, water hammer characteristics and valve disc opening change trend in the dynamic characteristic data, and dynamically adjust the valve disc closing speed according to the pressure change rate; Real-time adjustment of the closing curve based on the frequency and amplitude of water hammer impact, ensuring fast closing at the low amplitude stage of water hammer impact and avoiding pressure shock at the high amplitude stage; Optimize the closing sequence according to the stress state of the sealing surface to avoid excessive or uneven stress on the sealing surface due to closing too quickly.

4. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The adaptive control engine generates an optimal valve flap closing curve through a multi-objective optimization method, specifically including: Define multi-objective optimization goals, including minimizing the valve disc closing time, minimizing the sealing surface force during closing, and reducing the water hammer impact energy; Use a dynamic weight adjustment mechanism to assign weights to different optimization goals, including prioritizing reducing shutdown time when high risk levels are detected; The optimal closing curve is generated by an iterative optimization method, which contains the segmented speed and timing of the valve disc from starting to closing to full closing.

5. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The adaptive control engine converts the optimized closing curve into control instructions for the electric actuator, specifically including: Analyze the segmented speed and timing parameters in the closing curve into speed instructions and position instructions of the electric actuator; Add redundant check parameters to each control command to ensure that command errors caused by noise or data loss during transmission can be detected and corrected; The parsed control command is sent to the edge control unit, which drives the electric actuator to perform the valve flap closing operation.

6. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The spatiotemporal sequence prediction module is specifically used for: receiving the feature association vector and the dynamic evolution vector from the cross feature association module, and grouping the input data into multi-dimensional time series, wherein the grouping method is dynamically adjusted according to the time correlation in the feature association vector; The grouped time series data are processed using a recursive network. In the first stage of the recursive network, short-term dynamic characteristic patterns are extracted through unidirectional recursive units, including pressure change trends, vibration spectrum changes, and stress displacement rates. In the second stage, short-term and long-term dynamic characteristics are combined through a bidirectional recursive unit to capture the potential causal relationship between pressure and vibration characteristics; Use the temporal attention mechanism to assign weights to different time steps and dynamically enhance attention to key time points, including abnormal enhancement of vibration spectrum or rapid fluctuation of pressure difference; The output is the predicted characteristic trend in the time series, including the changing trend of water hammer impact intensity, the offset direction of sealing surface stress and the time range of possible delay of valve disc action.

7. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The risk grading and trend prediction module is specifically used for: Receive the predicted characteristic trend data from the spatiotemporal series prediction module, and first assign importance weights to each characteristic through the characteristic weight adjustment module, specifically including increasing the weights of vibration spectrum changes and pressure fluctuations to highlight the analysis of water hammer risks; Use the classification branch of the classification regression network to classify the adjusted input data, combine the predicted trend data with the characteristic weights, and generate risk types and their levels including backflow risk, water hammer impact and sealing abnormality; Use the regression branch of the classification regression network to predict the trend of the input data, and output the numerical prediction of the risk intensity and the development trend for each risk type, including the gradual increase of water hammer impact frequency or the expansion of the sealing stress excursion range; The output includes comprehensive prediction results of risk type, risk level and development trend, and generates early warning signals for adaptive control engine.

8. The reverse flow valve based on early warning control of digital twin technology according to claim 1 is characterized in that: The cross-feature association module models the dynamic coupling relationship between features through the following formula 4: ； in, Representation characteristics and Features In time The association weight of The number of frequency layers representing coupling between features; For time Lower characteristics and No. Layer coupling strength; Representation characteristics and No. layer-specific frequency content; For Features and In time The time offset of is a small positive number used to avoid the denominator being zero; For Features In time The characteristic value of For Features In time The characteristic value of is the feature difference weight, which is used to amplify the nonlinear relationship between features.

Images