An Adaptive Predictive Control Method and System for Eccentric Hemispherical Valves Based on Deep Learning

Abstract

This invention discloses an adaptive predictive control method and system for an eccentric hemispherical valve based on deep learning, belonging to the field of industrial valve control technology. The method includes: real-time acquisition of the eccentric hemispherical valve's operating status data; inputting the operating status data into a long short-term memory network prediction model to predict the dynamic characteristics of the eccentric hemispherical valve over multiple future control cycles; determining the initial opening position of the sealing and tightening stage based on the predicted future sealing contact state; constructing and solving a model predictive control optimization problem based on the predicted dynamic characteristics and the initial opening position to obtain the optimal control sequence; and outputting the first control quantity in the optimal control sequence to the actuator. This invention achieves early prediction of torque abrupt changes by predicting the sealing contact state, suppresses the impact of torque step abrupt changes through torque smoothing constraints, and adaptively compensates for the dynamic drift of the initial angle caused by factors such as sealing surface wear through an online learning update mechanism, thereby improving the control accuracy and stability of the eccentric hemispherical valve.

Description

Technical Field

[0001] This invention belongs to the field of eccentric hemispherical valve technology, specifically relating to an adaptive predictive control method and system for eccentric hemispherical valves based on deep learning. Background Technology

[0002] The eccentric hemispherical valve is a fluid control device widely used in petrochemical, power energy, and urban water systems. Its structural feature is a radial offset between the valve body's central axis and the hemispherical central axis, causing the hemispherical to gradually separate from and compress the valve seat during rotation. This structure ensures that the hemispherical is completely disengaged from the valve seat for most of the opening and closing process, with the sealing pair in an undamped, free-spinning state. Only in the final few degrees of the opening and closing motion does the hemispherical contact and press against the valve seat to achieve a self-locking seal. Unlike conventional ball valves where the sealing pair maintains contact and friction throughout the entire stroke, this characteristic of the eccentric hemispherical valve results in lower opening and closing torque, less wear on the sealing surface, and a longer service life.

[0003] However, the non-contact to contact transition characteristics of the eccentric hemispherical valve also present challenges to its precise control. During the closing process, as the valve enters the final few degrees of its stroke, the hemisphere needs to overcome the sealing clamping force to complete the wedging seal. The required torque suddenly increases from near zero to a peak value, exhibiting a step-like abrupt change, and the magnitude of this change is highly dependent on the medium pressure: the higher the medium pressure, the greater the medium clamping torque on the back of the hemisphere, and the torque required for sealing clamping increases non-linearly. Similarly, during the opening process, the torque required at the moment the hemisphere breaks free from the clamping seal is also much greater than the torque requirement during the subsequent idling phase, resulting in a sharp abrupt change from high to low. In addition, the sealing surface will undergo progressive wear during long-term operation, the valve seat compensation spring is at risk of fatigue failure, and mechanical clearance will accumulate at the connection between the actuator and the valve stem. These factors will all cause dynamic drift in the sealing contact initiation angle, making it impossible for the factory-preset fixed parameters to remain effective throughout the entire lifespan. In the application of eccentric ball valves, traditional PID control methods cannot predict the timing of the sealing contact transition. This leads to controller lag during the final few degrees of the opening and closing stroke when torque surges occur, resulting in excessive impact during closing and jamming or overshoot during opening. Therefore, there is an urgent need to improve existing technologies to solve the problems of torque step abrupt changes during the opening and closing process caused by the non-contact-to-contact transition characteristics of eccentric ball valves, as well as the control mismatch caused by the dynamic drift of the sealing contact initiation angle. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an adaptive predictive control method and system for eccentric hemispherical valves based on deep learning, so as to solve the problem of torque abrupt change during the opening and closing process caused by the non-contact-contact switching characteristics of eccentric hemispherical valves and the control mismatch problem caused by the dynamic drift of the sealing contact initiation angle.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an adaptive predictive control method for an eccentric hemispherical valve based on deep learning, comprising the following steps: S100: Real-time acquisition of operating status data of the eccentric hemispherical valve, the operating status data including at least the hemispherical opening position, the medium pressure before the valve, the medium pressure after the valve, the medium flow rate through the eccentric hemispherical valve, and the torque output by the actuator of the eccentric hemispherical valve. S200: Input the operating status data into a pre-built long short-term memory network prediction model to predict the dynamic characteristics of the eccentric hemispherical valve in multiple future control cycles. The dynamic characteristics include at least the future opening position, future medium pressure, future medium flow rate, and future sealing contact state. S300: Based on the future sealing contact state, determine the initial opening position of the eccentric hemispherical valve as it is about to enter the sealing and pressing stage; S400: Based on the future dynamic characteristics and the initial opening position, construct and solve the model predictive control optimization problem to obtain the optimal control sequence of the eccentric hemispherical valve in the future control cycle, wherein the objective function of the model predictive control optimization problem includes a tracking error term, a control increment smoothing term, and a torque smoothing constraint term in the sealing and pressing stage. S500: Output the first control quantity in the optimal control sequence to the actuator of the eccentric hemispherical valve to drive the actuator to adjust the opening position of the hemispherical valve.

[0006] Preferably, step S100 specifically includes: acquiring the torque output by the actuator through a torque sensor installed on the output shaft of the actuator; acquiring the medium pressure before and after the valve through pressure transmitters installed in the pipeline before and after the valve, respectively; acquiring the medium flow rate through the eccentric hemispherical valve through a flow meter installed on the valve outlet side; and acquiring the hemispherical opening position through an encoder linked to the actuator.

[0007] Preferably, the long short-term memory network prediction model in step S200 includes a main prediction model and a sealing contact state prediction sub-model; step S200 further includes: constructing the sealing contact state prediction sub-model, which adopts a long short-term memory network structure, and its input layer receives the hemispherical opening position, valve inlet medium pressure, historical torque sequence, and sealing surface wear coefficient of the current cycle, wherein the historical torque sequence is taken from the torque values ​​of several consecutive control cycles before the current cycle; the output layer of the sealing contact state prediction sub-model outputs two prediction values, the first prediction value is a binary classification result of whether the hemispherical and valve seat are in a sealing contact state in each future control cycle, and the second prediction value is the expected sealing compression starting opening position; the operating status data is simultaneously input into the main prediction model and the sealing contact state prediction sub-model, the main prediction model outputs the future opening position, future medium pressure, and future medium flow rate, and the sealing contact state prediction sub-model outputs the future sealing contact state and the expected sealing compression starting opening position.

[0008] Preferably, the specific method for determining the initial opening position of the eccentric hemispherical valve about to enter the sealing and pressing stage in step S300 is as follows: read the expected initial opening position of sealing and pressing from the sealing contact state prediction sub-model, and record this position as the first initial boundary value; compare the hemispherical opening position of the current cycle with the first initial boundary value; when the hemispherical opening position of the current cycle is less than the first initial boundary value and the difference between the two is greater than a preset angle threshold, it is determined that the valve is in the idling stage and has not yet entered the sealing and pressing stage, and the torque smoothing constraint term is set to an inactive state; when the hemispherical opening position of the current cycle is greater than or equal to the first initial boundary value, it is determined that the valve has entered the sealing and pressing stage, the torque smoothing constraint term is set to an active state, and the first initial boundary value is used as the starting boundary of the sealing and pressing stage.

[0009] Preferably, after the torque smoothing constraint term is activated, the following constraints are applied during the solution of the model predictive control optimization problem: the absolute value of the torque change rate of the actuator is limited to not exceeding a preset maximum torque change rate limit, which is pre-calibrated based on the rated torque of the eccentric hemispherical valve and the maximum allowable impact coefficient; when the torque value estimated by the control quantity corresponding to the sealing and clamping stage in the optimal control sequence obtained by solving the model predictive control optimization problem is lower than the preset minimum sealing and holding torque threshold, a penalty term is added to force the solver to increase the torque output so that the actual torque is not lower than the minimum sealing and holding torque threshold.

[0010] Preferably, the model predictive control optimization problem construction in step S400 further includes a segmented control weight adjustment step: dividing the motion range of the hemisphere into an idling range and a sealing and pressing range according to the initial opening position, wherein the idling range is the range between the hemisphere opening position decreasing from the fully open position to the initial opening position, and the sealing and pressing range is the range between the hemisphere opening position decreasing from the initial opening position to the fully closed position; when the predicted hemisphere opening position is in the idling range, the weight matrix of the tracking error term in the objective function is set as the first weight matrix; when the predicted hemisphere opening position is in the sealing and pressing range, the weight matrix of the tracking error term in the objective function is set as the second weight matrix, and at the same time, the weight of the torque smoothing constraint term is set to a non-zero value, wherein the penalty weight for the opening position tracking error in the second weight matrix is ​​higher than the corresponding weight in the first weight matrix.

[0011] Preferably, the segmented control weight adjustment step further includes interval boundary smoothing transition processing: during the transition control cycle when the predicted hemispherical opening position crosses from the idling interval to the sealing and pressing interval, the weight matrix of the tracking error term is transitioned from the first weight matrix to the second weight matrix periodically using linear interpolation, and the weight of the torque smoothing constraint term is increased from zero to a preset target value periodically.

[0012] Preferably, the method further includes an online learning and update step: during the operation of the eccentric hemispherical valve, the operating status data and actual torque feedback value of each control cycle are continuously collected and stored in a circular buffer; when the amount of new data accumulated in the circular buffer reaches a preset update threshold, the data from the most recent continuous time period is extracted from the circular buffer to form an incremental training set; the current long short-term memory network prediction model is trained using the incremental training set with mini-batch gradient descent, and a learning rate lower than the initial training learning rate is set during the training process, and the incremental training set is mixed with some historical samples stored in the circular buffer for training using the experience replay technique; the prediction error of the model after incremental training is evaluated on an independent validation set, and if the prediction error is reduced by more than a preset relative improvement threshold compared with the model before the update, the updated model is deployed to the online prediction module to replace the original model.

[0013] Preferably, the real-time iterative method is used to solve the model predictive control optimization problem in step S400, specifically including: in the preparation stage of each control cycle, that is, in the previous control cycle, the following operations are performed: the long short-term memory network prediction model is linearized based on the current operating point, a coefficient matrix of the quadratic programming form of the model predictive control optimization problem is constructed, and the coefficient matrix is ​​pre-decomposed, and the decomposition result is stored in memory; in the feedback stage of the current control cycle, the following operations are performed: the current state measurement value is received, the constraint boundary of the quadratic programming form is updated according to the measurement value, the quadratic programming problem is solved using the pre-decomposition result stored in memory, the optimal control increment is obtained, and then the optimal control quantity is calculated.

[0014] Secondly, the present invention also provides an adaptive predictive control system for an eccentric hemispherical valve based on deep learning, comprising: The sensing module is used to collect the operating status data of the eccentric hemispherical valve in real time. The operating status data includes at least the opening position of the hemispherical valve, the medium pressure before the valve, the medium pressure after the valve, the medium flow rate through the eccentric hemispherical valve, and the torque output by the actuator of the eccentric hemispherical valve. A long short-term memory network prediction module, connected to the sensing module, is used to receive the operating status data and predict the dynamic characteristics of the eccentric hemispherical valve in multiple future control cycles. The dynamic characteristics include at least the future opening position, future medium pressure, future medium flow rate, and future sealing contact state. A sealing state determination module, connected to the long short-term memory network prediction module, is used to determine the initial opening position of the eccentric hemispherical valve as it is about to enter the sealing and pressing stage based on the predicted future sealing contact state. The model predictive control module, connected to the long short-term memory network prediction module and the sealing state determination module, is used to construct and solve the model predictive control optimization problem based on the predicted future dynamic characteristics and the determined initial opening position to obtain the optimal control sequence. The objective function of the model predictive control optimization problem includes a tracking error term, a control increment smoothing term, and a torque smoothing constraint term for the sealing and pressing stage. The execution drive module, connected to the model prediction control module, is used to output the first control quantity in the optimal control sequence to the actuator of the eccentric hemispherical valve, driving the actuator to adjust the opening position of the hemispherical body.

[0015] This invention discloses a deep learning-based adaptive predictive control method and system for eccentric hemispherical valves. Its core lies in: predicting the future sealing contact state of the eccentric hemispherical valve using a long short-term memory network prediction model to determine the initial opening position during the sealing and tightening stage; and then introducing a torque smoothing constraint term into the model predictive control optimization problem to actively constrain torque changes during the sealing and tightening stage. Compared to existing technologies, this invention has significant advantages: predicting the sealing contact state allows for early prediction of torque abrupt changes, avoiding the impact or jamming problems caused by response lag in traditional feedback control; the torque smoothing constraint term limits the rate of torque change during the sealing and tightening stage, effectively suppressing the impact of torque step abrupt changes; segmented control weight adjustment allows for different control strategies in the idling and sealing and tightening stages, balancing rapid response and stable sealing; and an online learning update mechanism adaptively compensates for the dynamic drift of the sealing contact initiation angle caused by factors such as sealing surface wear and accumulated mechanical clearance, ensuring that control parameters remain effective throughout the entire lifecycle. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the deep learning-based adaptive predictive control method for an eccentric hemispherical valve provided in an embodiment of the present invention. Figure 2 A structural block diagram of an eccentric hemispherical valve adaptive predictive control system based on deep learning provided in an embodiment of the present invention; Figure 3 A schematic diagram of the overall structure of the deep learning-based adaptive predictive control system for an eccentric hemispherical valve provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of the Long Short-Term Memory Network prediction model provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the solution process for the model predictive control optimization problem provided in an embodiment of the present invention. Figure 6 A bar chart comparing the key performance indicators of the three control methods provided in the embodiments of the present invention; Figure 7 The diagram shows the opening tracking, control output, and disturbance comparison curves of three control methods under the high pressure differential large drift condition provided in the embodiments of the present invention. Figure 8The graph shows the opening tracking, control output, and disturbance comparison curves of three control methods under high friction conditions with small opening provided in the embodiments of the present invention.

[0018] Explanation of reference numerals in the attached figures: 10. Sensing module; 20. Long Short-Term Memory Network Prediction Module; 30. Sealing State Determination Module; 40. Model Predictive Control Module; 50. Execution Drive Module; 60. Eccentric Hemispherical Valve; 61. Actuator; 62. Hemispherical body. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] In traditional eccentric hemispherical valves, the non-contact to contact transition between the hemisphere and valve seat during opening and closing causes a step-like change in the required torque, and the magnitude of this change is non-linearly affected by the medium pressure. Traditional control methods, unable to predict the timing of the sealing contact transition, result in a delayed controller response, leading to problems such as excessive closing impact, opening jamming, or overshoot.

[0021] To address this, this application proposes a deep learning-based adaptive predictive control method for eccentric hemispherical valves, such as... Figure 1 As shown, the method includes the following steps: S100: Real-time acquisition of operating status data of the eccentric hemispherical valve 60, including at least the hemispherical opening position, the medium pressure before the valve, the medium pressure after the valve, the medium flow rate through the eccentric hemispherical valve, and the torque output by the actuator 61 of the eccentric hemispherical valve 60. S200: Input the operating status data into a pre-built long short-term memory network prediction model to predict the dynamic characteristics of the eccentric hemispherical valve 60 in multiple future control cycles. The dynamic characteristics include at least the future opening position, future medium pressure, future medium flow rate, and future sealing contact state. S300: Based on the future sealing contact state, determine the initial opening position of the eccentric hemispherical valve 60 as it is about to enter the sealing and tightening stage; S400: Based on the future dynamic characteristics and the initial opening position, construct and solve the model predictive control optimization problem to obtain the optimal control sequence of the eccentric hemispherical valve 60 in the future control cycle. The objective function of the model predictive control optimization problem includes the tracking error term, the control increment smoothing term, and the torque smoothing constraint term in the sealing and pressing stage. S500: Output the first control quantity of the optimal control sequence to the actuator 61 of the eccentric hemispherical valve 60, and drive the actuator 61 to adjust the opening position of the hemispherical body 62.

[0022] The eccentric ball valve 60 is a valve with a special structure. Its hemisphere 62 and valve seat switch between non-contact and contact during opening and closing. For most of the stroke, the hemisphere 62 does not contact the valve seat and is in a free-running state; only in the last few strokes of the closing stroke does the hemisphere 62 contact and press against the valve seat to achieve a seal. The Long Short-Term Memory (LSTM) network prediction model is a deep learning model suitable for processing and predicting time-series data. This model can learn long-term dependencies in the data, thereby predicting the future dynamic behavior of the eccentric ball valve 60 based on historical and current operating data. The sealing and pressing stage refers to the specific stroke range during the closing process of the eccentric ball valve 60, where the hemisphere 62 begins to contact the valve seat and gradually presses against it until a complete seal is achieved. During this stage, the required torque of the valve increases significantly. The initial opening position refers to the critical opening value at which the eccentric ball valve 60 transitions from the free-running stage to the sealing and pressing stage. Accurate determination of this position is crucial for achieving smooth, shock-free sealing and pressing control. Model predictive control optimization (MMC) is a control strategy based on model prediction. It determines the optimal control sequence over a future period by solving an optimization problem within each control cycle, based on the current state and future predictions. The optimal control sequence refers to a series of future control variables obtained by solving the MMC, which aim to enable the system to achieve the expected control objective while satisfying constraints, such as minimizing tracking error or smoothing control actions. The actuator 61 is a device that drives the movement of the hemisphere 62 of the eccentric ball valve 60. It typically consists of a motor, gearbox, etc. It receives control signals and converts them into mechanical motion to adjust the opening position of the hemisphere 62.

[0023] Specifically, in step S100, the operating status data of the eccentric hemispherical valve 60 is collected in real time by sensors, including the hemispherical opening degree, the pressure before and after the valve, the medium flow rate, and the actuator torque, and transmitted to the control system. In step S200, the data is input into the LSTM predictive model, which learns the valve's dynamic behavior and predicts the opening degree, pressure, flow rate, and sealing contact state over future multiple cycles, providing forward-looking information for control. In step S300, based on the predicted sealing contact state, the initial opening degree position for entering the sealing and tightening stage is determined to accurately identify the first contact point between the hemispherical valve and the valve seat. In step S400, based on the future dynamic characteristics and the initial opening degree position, a model predictive control optimization problem is constructed, whose objective function includes a tracking error term, a control increment smoothing term, and a torque smoothing constraint term, and the optimal control sequence is obtained by solving the problem. In step S500, the first control quantity of the sequence is output to the actuator 61, driving the hemispherical valve 62 to adjust its opening degree, and the above process is repeated to form a closed-loop control. The above technical solution effectively solves the control problems caused by the sudden change in torque and the dynamic drift of the initial sealing contact angle during the opening and closing of the eccentric hemispherical valve 60. By combining deep learning prediction and model predictive control, precise adaptive adjustment of the valve opening position is achieved, thereby effectively avoiding problems such as excessive impact when closing, jamming or overshoot when opening, and ensuring stable operation of the valve under different working conditions and wear conditions.

[0024] This application further proposes that step S100 specifically includes: acquiring the torque output by the actuator 61 through a torque sensor installed on the output shaft of the actuator 61; acquiring the upstream and downstream medium pressures through pressure transmitters installed in the upstream and downstream pipelines respectively; acquiring the medium flow rate through the eccentric hemispherical valve 60 through a flow meter installed on the valve outlet side; and acquiring the opening position of the hemispherical valve 62 through an encoder linked to the actuator 61. Figure 3 As shown.

[0025] Specifically, torque is acquired by a torque sensor mounted on the output shaft of actuator 61. A strain gauge sensor, after calibration, provides a precise torque value for monitoring the stress and wear of the sealing surface during the sealing and tightening phase. Medium pressure is acquired through pressure transmitters in the upstream and downstream pipelines, such as diffused silicon pressure transmitters, to monitor the pressure difference and flow resistance characteristics across the valve in real time. Medium flow rate is acquired through flow meters on the valve outlet side, such as electromagnetic flow meters or vortex flow meters, providing a basis for process stability and control optimization. The opening position of the hemisphere 62 is acquired through an encoder linked to actuator 61, such as an absolute rotary encoder, directly outputting a digital signal corresponding to the angle of the hemisphere. The position information is maintained even after power failure, ensuring strict synchronization between the angle and the opening position, achieving high-precision position feedback.

[0026] This application further proposes that the long short-term memory network prediction model in step S200 includes a main prediction model and a sealing contact state prediction sub-model. This modular design allows the prediction task to be decomposed. The main prediction model can focus on predicting conventional dynamic characteristics such as future opening position, future medium pressure, and future medium flow rate, while the sealing contact state prediction sub-model is specifically used to handle complex dynamics related to the seal.

[0027] Specifically, the sealing contact state prediction sub-model employs a long short-term memory network structure. Its input layer receives the current cycle's hemispherical 62 opening position, upstream medium pressure, historical torque sequence, and sealing surface wear coefficient. The hemispherical 62 opening position and upstream medium pressure provide direct information about the valve's current operating state; the historical torque sequence is taken from torque values ​​of several consecutive control cycles prior to the current cycle, allowing the model to learn the trend of torque changes over time, thus more accurately determining the occurrence and degree of sealing contact; the sealing surface wear coefficient, as a key parameter reflecting the valve's physical state, directly affects sealing performance and the required clamping force. The combination of these input features provides comprehensive information for accurate prediction of the sealing contact state. The sealing contact state prediction sub-model outputs two predicted values. The first predicted value is a binary classification result indicating whether the hemispherical 62 and valve seat are in a sealing contact state in each future control cycle, directly indicating whether the valve has entered or is about to enter a sealing state. The second predicted value is the expected sealing clamping starting opening position, providing a precise starting point for the sealing clamping stage, which is crucial for adjusting subsequent control strategies.

[0028] As a specific implementation method, this embodiment uses a standardized Long Short-Term Memory (LSTM) network structure to achieve engineering implementation of the sealed contact state prediction sub-model. Its core computational logic revolves around the gating mechanism of the LSTM unit. The LSTM unit achieves long-term memory and updating of temporal information through the collaborative work of forget gates, input gates, and output gates. The core computation follows the following set of formulas. in The output of the forget gate is used to determine which information from the previous cell state is discarded. The input gate output determines which new information to update to the cell state. The output gate determines which cell state information is output to the hidden layer. This represents the current cell state. This is the output of the hidden layer at the current moment. This is the weight matrix corresponding to the gating. For the bias term corresponding to the gating, for The input feature vector at time t, It is the sigmoid activation function. Using the Hadamard product, this gating mechanism allows the model to effectively capture long-term dependencies in timing data such as torque and opening degree, and accurately identify the timing characteristics of seal contact state transitions.

[0029] For the binary classification output of the sealed contact state, this embodiment uses the sigmoid activation function to achieve the probability output, and the calculation formula is as follows: ,in for The predicted probability that hemisphere 62 is in sealed contact with the valve seat at time [0,1] is given. If the binary classification result indicates a contact state, then it is determined to be a non-contact state. These are the weights and biases for the classification layer.

[0030] Regarding the wear coefficient of the sealing surface, which reflects the long-term aging state of the valve, this embodiment uses a multi-parameter linear fitting method to achieve real-time updates. The fitting formula is as follows: in The current wear coefficient of the sealing surface, initial value. , This represents the cumulative number of valve opening and closing cycles. The drift angle is the position of the fully closed position. This represents the cumulative change in the peak torque during the sealing and tightening phase. The fitting coefficients, calibrated through valve type testing, require no additional sensors; the wear coefficients are updated in real time using only the collected operational data, providing crucial information for model prediction of valve aging conditions. Figure 4 As shown, in this embodiment, the sealing contact state prediction sub-model adopts a network structure of 2 LSTM layers + 1 fully connected layer, with 64 hidden layer neurons, an input length of 50 control cycles for the historical torque sequence, and a prediction time domain of 20 control cycles. The main prediction model adopts a network structure isomorphic to the sub-model, only adjusting the output layer dimension to match the regression prediction requirements of opening, pressure, and flow. Through this modular dual-branch network structure, accurate prediction of both the macroscopic dynamic characteristics of the valve and the microscopic characteristics of the sealing contact can be achieved simultaneously, avoiding the problem of insufficient prediction accuracy of core events caused by multi-task prediction of a single model.

[0031] This application further proposes a specific method for determining the initial opening position of the eccentric hemispherical valve 60 when it is about to enter the sealing and pressing stage. The method includes: reading the expected initial opening position of the sealing and pressing stage from the sealing contact state prediction sub-model and recording this position as the first initial boundary value; comparing the opening position of the hemispherical valve 62 in the current cycle with the first initial boundary value; when the opening position of the hemispherical valve 62 in the current cycle is less than the first initial boundary value and the difference between the two is greater than a preset angle threshold, it is determined that the valve is in the idling stage and has not yet entered the sealing and pressing stage, and the torque smoothing constraint term is set to an inactive state; when the opening position of the hemispherical valve 62 in the current cycle is greater than or equal to the first initial boundary value, it is determined that the valve has entered the sealing and pressing stage, the torque smoothing constraint term is set to an active state, and the first initial boundary value is used as the starting boundary of the sealing and pressing stage.

[0032] Specifically, the predicted initial opening position for sealing and tightening is read from the sealing contact state prediction sub-model and recorded as the first initial boundary value. This value is predicted in real time by a deep learning model based on historical torque sequences, current opening, and pressure data, and can dynamically reflect the initial angle drift caused by sealing surface wear, mechanical clearance, etc. Subsequently, the opening position of the hemisphere 62 in the current cycle is compared with the first initial boundary value: if the current opening position is less than the first initial boundary value and the difference between the two is greater than a preset angle threshold, the valve is determined to be in the idling stage, and the torque smoothing constraint is set to the inactive state. In the idling stage, the hemisphere 62 has no contact with the valve seat, the torque requirement is low, and the control target is rapid positioning, so releasing the torque constraint can improve the response speed of the actuator 61. If the current opening position is greater than or equal to the first initial boundary value, the valve is determined to have entered the sealing and tightening stage, and the torque smoothing constraint is immediately switched to the active state, with the first initial boundary value as the starting boundary for sealing and tightening. During this stage, the hemisphere 62 begins to contact the valve seat and gradually presses against it, causing a sharp increase in torque. Activating the torque smoothing constraint limits the rate of torque change and ensures a minimum sealing torque, thereby suppressing impact and protecting the sealing surface. A preset angle threshold is used to avoid misjudgments caused by prediction errors or measurement noise, ensuring that the control mode is switched only when the system is clearly approaching or entering the sealing and pressing stage. Through the above stage determination, the system achieves differentiated control between the idling and sealing and pressing stages: the idling stage prioritizes rapid response, while the sealing and pressing stage prioritizes stability and reliability, effectively solving the impact or jamming problems caused by the inability to predict the contact moment in traditional control.

[0033] As a specific implementation method, this embodiment adopts a quantitative engineering implementation method for the angle threshold and boundary judgment logic of the stage judgment. First, the preset angle threshold is completed. The calibration is performed using the following formula: ,in This refers to the theoretical sealing contact initiation angle specified at the valve's factory. This is the threshold coefficient, ranging from 0.2 to 0.5. It can be adjusted according to the valve control accuracy requirements, with 0.3 recommended for normal operating conditions. This calibration method can adaptively match the threshold size based on the inherent structural characteristics of the valve, avoiding the problem of insufficient adaptability of a fixed threshold. After completing the threshold calibration, a quantitative stage judgment logic is established, defining the current cycle opening degree as... The first initial boundary value is When satisfied and When the system determines that the valve is in the idling stage, the weighting coefficient of the torque smoothing constraint term is adjusted accordingly. That is, the constraint term is in an inactive state; when the condition is met... At this point, the system determines that the valve has entered the sealing and tightening stage, and the weighting coefficient of the torque smoothing constraint term is adjusted accordingly. That is, the constraint term is in an active state, where The target weight value is preset for the sealing and pressing range; through this quantitative judgment logic, the frequent switching of stages caused by prediction noise can be effectively avoided, while ensuring the timeliness and accuracy of control strategy switching, and adapting to the application needs of valves with different diameters and different working conditions.

[0034] This application further proposes that after the torque smoothing constraint term is activated, the following constraints are applied during the solution of the model predictive control optimization problem: the absolute value of the torque change rate of the actuator 61 is limited to not exceeding the preset maximum torque change rate limit, which is pre-calibrated based on the rated torque of the eccentric hemispherical valve 60 and the maximum allowable impact coefficient; when the torque value estimated by the control quantity corresponding to the sealing and clamping stage in the optimal control sequence obtained by solving the model predictive control optimization problem is lower than the preset minimum sealing and holding torque threshold, a penalty term is added to force the solver to increase the torque output so that the actual torque is not lower than the minimum sealing and holding torque threshold.

[0035] As a specific implementation method, this embodiment adopts a quantitative engineering approach for torque smoothing constraints and penalty terms. First, a hard constraint on the torque change rate is established, and the constraint formula is as follows: ,in For the first Torque change per control cycle , The first , The actuator 61 outputs torque in each control cycle. The maximum torque change rate limit is a hard constraint directly added to the inequality constraints of the quadratic programming problem to ensure that the solved control sequence meets the torque stability requirement; where the maximum torque change rate limit is... The calibration follows the formula In the formula The rated torque of valve actuator 61, The maximum permissible impact coefficient is typically set between 0.1 and 0.3 under normal operating conditions, with lower values ​​for high-pressure conditions and higher values ​​for low-pressure conditions. This calibration method, which determines the number of control cycles expected during the sealing and tightening phase, allows for adaptive matching of a reasonable torque change rate limit based on the valve's rated parameters and operating characteristics, balancing control response speed and stability. For the minimum sealing holding torque threshold, this embodiment employs a dynamic calibration method related to differential pressure, with the calibration formula being: ,in To achieve the minimum seal holding torque threshold, The minimum torque required for sealing under zero differential pressure is determined by the valve's factory type test. The pressure difference-torque fitting coefficient is obtained through valve hydrostatic testing. This dynamic calibration method uses the real-time pressure difference between the medium before and after the valve to adaptively adjust the sealing torque threshold based on the real-time pressure difference in the pipeline, avoiding the problems of insufficient sealing under high-pressure conditions and excessive torque under low-pressure conditions when using a fixed threshold. For the penalty term for insufficient torque, this embodiment adopts a quadratic penalty function, calculated using the following formula: ,in The penalty term value is directly added to the objective function of the model's predictive control to participate in the optimization solution. The weighting coefficient for the penalty term. For the first The predicted torque value for each control cycle, through this secondary penalty term, can force the optimization result to meet the minimum sealing torque requirement while ensuring the convergence of the solver, thus avoiding internal leakage of the valve.

[0036] This application further proposes a segmented control weight adjustment step when constructing the model predictive control optimization problem. Based on the initial opening position, the motion range of the hemisphere is divided into an idling range and a sealing / pressing range. The idling range is the interval between the hemisphere's opening position decreasing from the fully open position to the initial opening position, and the sealing / pressing range is the interval between the hemisphere's opening position decreasing from the initial opening position to the fully closed position. When the predicted hemisphere opening position is in the idling range, the weight matrix of the tracking error term in the objective function is set as the first weight matrix. When the predicted hemisphere opening position is in the sealing / pressing range, the weight matrix of the tracking error term in the objective function is set as the second weight matrix, and simultaneously, the weight of the torque smoothing constraint term is set to a non-zero value. The penalty weight for the opening position tracking error in the second weight matrix is ​​higher than the corresponding weight in the first weight matrix.

[0037] The segmented control weight adjustment step divides the movement range of the hemisphere 62 into an idling range and a sealing and tightening range based on the initial opening position. The idling range is between the opening degree decreasing from fully open to the initial opening position, where the hemisphere has not yet contacted the valve seat; the control focus is on rapid and accurate positioning. The sealing and tightening range is between the opening degree decreasing from the initial opening to fully closed, where the hemisphere gradually tightens; the control focus shifts to smoothly applying the tightening force, avoiding impact, and ensuring a seal. When the predicted opening degree is in the idling range, the weight matrix of the tracking error term in the objective function is set to the first weight matrix, giving a higher weight to the opening tracking error to achieve rapid and efficient positioning. When the predicted opening degree enters the sealing and tightening range, the weight matrix of the tracking error term is switched to the second weight matrix, where the penalty weight for the opening tracking error is higher than that of the first weight matrix, to ensure the smoothness of the tightening process. Simultaneously, the weight of the torque smoothing constraint term is set to a non-zero value, making the controller pay more attention to the smoothness of the output torque of the actuator 61, avoiding drastic torque fluctuations, thereby preventing impact and protecting the valve sealing surface and the actuator 61.

[0038] As a specific implementation method, this embodiment adopts a standardized engineering construction method for the objective function and piecewise weight matrix of model predictive control. First, the overall form of the objective function is established as follows: In the formula For time-domain prediction, the standard setting is 20 control cycles. To control the time domain, it is typically set to 5 control cycles. The predicted output vector includes system states such as aperture, pressure, and flow rate. As a reference setpoint vector, To control the increment, This represents the change in torque. For the tracking error weight matrix, To control the incremental weight matrix, The torque smoothing constraint weight matrix is... The activation weights for the torque constraint term. As a penalty term for insufficient torque, the objective function simultaneously considers tracking accuracy, control stability, torque constraint, and sealing reliability. It can be adapted to the control requirements of different stages by adjusting the weight matrix. For the differentiated control requirements of the idling and sealing tightening intervals, this embodiment establishes a segmented weight matrix configuration rule, with the first weight matrix corresponding to the idling interval... In the middle, traffic tracking weight Pressure tracking weights Opening tracking weight This configuration prioritizes the flow and pressure tracking response speed during the idling phase, adapting to the core requirements of process adjustment; the second weight matrix corresponds to the sealing and tightening zone. In the middle, traffic tracking weight Pressure tracking weights Opening tracking weight This configuration prioritizes precise control of the opening position during the sealing and tightening stage to avoid overshoot impact; the weight matrix... It is a diagonal matrix, with the diagonal elements being the tracking weight values ​​for the corresponding states, and the remaining elements being 0. Through this segmented weight configuration, differentiated control objectives can be achieved for different operating ranges, taking into account both the speed of valve adjustment and the stability of the sealing process.

[0039] This application further proposes that the segmented control weight adjustment step also includes interval boundary smoothing transition processing: during the transition control cycle from the idling interval to the sealing and pressing interval at the predicted hemispherical 62-degree opening position, the weight matrix of the tracking error term is transitioned from the first weight matrix to the second weight matrix cycle by cycle using linear interpolation, and the weight of the torque smoothing constraint term is increased from zero to the preset target value cycle by cycle.

[0040] The smooth transition processing at the boundary of the valve's operating range is used to avoid abrupt changes in control parameters at the boundary of the operating range, ensuring the smoothness of valve state switching and suppressing system oscillations and shocks. The transition control cycle refers to the set of continuous control cycles in which weight parameters are gradually adjusted as the hemispherical opening transitions from the idling range to the sealing and tightening range. Its length can be preset and optimized according to the valve's dynamic characteristics. Using linear interpolation, the weight matrix and torque smoothing constraint weights are calculated proportionally based on the relative position within the transition cycle. Specifically, as the valve transitions from idling to sealing and tightening, the focus on the opening tracking error needs to be gradually increased. Through linear interpolation, the weight matrix of the tracking error term smoothly transitions from the first weight matrix in the idling range to the second weight matrix in the sealing and tightening range, avoiding oscillations or overshoot caused by abrupt weight changes. Simultaneously, the weight of the torque smoothing constraint term increases linearly from zero to the preset target value, allowing the torque constraint to gradually take effect during the sealing and tightening stage, effectively controlling the output torque of the actuator 61, avoiding excessive shocks, and protecting the valve sealing surface. This smooth transition mechanism improves the stability and reliability of valve operation.

[0041] As a specific implementation method, this embodiment uses a linear interpolation engineering approach for smooth transition processing at interval boundaries. First, the number of transition control cycles is set to... The first control cycle ensures that the weighted parameters smoothly transition before the valve enters the sealing and tightening zone, avoiding system jitter caused by sudden parameter changes; for the first... A transition cycle, in which The interpolation calculation of the tracking error weight matrix follows the formula In the formula For the first The tracking error weight matrix for each transition period. This is the first weight matrix for the idle interval. The second weight matrix for the sealing and tightening zone can achieve a smooth, cycle-by-cycle transition from the idling zone configuration to the sealing and tightening zone configuration using this linear interpolation formula, without parameter jumps. The interpolation calculation for the weights of the torque smoothing constraint term follows the formula... In the formula For the first Torque constraint weight for each transition cycle The preset target weight value for the sealing and pressing zone is used to achieve a linear increase in the torque constraint weight from zero to the target value through this formula, ensuring that the torque constraint gradually takes effect during the transition phase. In this embodiment, the triggering condition for interpolation calculation is: when the predicted opening is less than or equal to the first initial boundary value for the first time in the prediction time domain, the interpolation calculation of the transition period is immediately started to ensure that the weight is smoothly transitioned before the valve actually enters the sealing and pressing zone, further improving the stability of the control process and adapting to the operating conditions of actuators 61 and valves with different response speeds.

[0042] This application further proposes an online learning update step. During the operation of the eccentric hemispherical valve, the operating status data and actual torque feedback value of each control cycle are continuously collected and stored in a circular buffer. When the amount of new data accumulated in the circular buffer reaches a preset update threshold, data from the most recent continuous time period is extracted from the circular buffer to form an incremental training set. The current long short-term memory network prediction model is trained using mini-batch gradient descent with the incremental training set. During training, a learning rate lower than the initial training learning rate is set, and an experience replay technique is used to mix the incremental training set with some historical samples stored in the circular buffer for training. The prediction error of the model after incremental training is evaluated on an independent validation set. If the prediction error is reduced by more than a preset relative improvement threshold compared to the model before the update, the updated model is deployed to the online prediction module to replace the original model.

[0043] As a specific implementation method, this embodiment adopts an engineered incremental training and model update approach for the online learning update step. First, a combined loss function for incremental training is constructed, taking into account the accuracy requirements of both regression and classification tasks. The loss function formula is as follows: In the formula This represents the number of samples in the incremental training set. For regression tasks such as opening degree, pressure, flow rate, and initial opening position, the predicted values ​​are... To regress the true measurement value of the task, This is the actual label indicating the sealed contact state, with a value of 0 or 1. The loss function, which predicts the sealed contact state probability, consists of a mean squared error loss in the first half to optimize regression prediction accuracy and a cross-entropy loss in the second half to optimize classification accuracy. Combining these loss functions ensures the accuracy of multi-task predictions. For the experience playback sampling strategy, this embodiment sets the mixing ratio of incremental training set to historical samples to 7:3. Historical samples prioritize data from the sealed contact transition phase to ensure the model retains its core characteristics and avoids catastrophic forgetting issues caused by incremental training. Regarding the hyperparameter configuration for incremental training, the initial training base learning rate is... The learning rate for incremental training is set to The batch size for mini-batch training is set to 32, and the number of training epochs is set to 10 to avoid model overfitting caused by incremental training.

[0044] This application further proposes a real-time iterative method for solving the model predictive control optimization problem in step S400, specifically including: in the preparation phase of each control cycle, i.e., in the previous control cycle, the following operations are performed: the long short-term memory network prediction model is linearized based on the current operating point, a coefficient matrix in the quadratic programming form of the model predictive control optimization problem is constructed, and the coefficient matrix is ​​pre-decomposed, with the decomposition results stored in memory; in the feedback phase of the current control cycle, the following operations are performed: the current state measurement value is received, the constraint boundary in the quadratic programming form is updated according to the measurement value, the quadratic programming problem is solved using the pre-decomposition results stored in memory, the optimal control increment is obtained, and then the optimal control quantity is calculated.

[0045] Specifically, in the preparation phase of the previous control cycle, the LSTM model is first linearized based on the current operating point, transforming the nonlinear model into a linear incremental model. Then, based on the linearized system model, objective function, and constraints, a quadratic programming (QP) coefficient matrix for the model predictive control optimization problem is constructed. Finally, this matrix is ​​pre-decomposed (e.g., Cholesky decomposition), and the decomposition results are stored in memory, thus completing time-consuming calculations such as matrix inversion in advance and avoiding repeated calculations in each cycle. In the feedback phase of the current control cycle, the system obtains the latest state measurement value of the eccentric hemispherical valve 60 through the sensing module 10 and dynamically updates the constraint boundary of the QP problem. Then, the QP problem is quickly solved using the pre-decomposition results in memory to obtain the optimal control increment. Since the pre-decomposition has been completed in advance, the feedback phase only requires simple operations such as forward and backward substitution, resulting in extremely fast solution. Finally, the increment is superimposed with the control quantity of the previous cycle to obtain the optimal control quantity for the current cycle, which is then output to the actuator 61 to drive the valve to adjust its opening. This method moves most of the calculations to the preparation phase, with only a small amount of real-time calculation in the feedback phase, significantly reducing control latency and meeting the requirements of rapid response and continuous adjustment.

[0046] As a specific implementation method, such as Figure 5 As shown, this embodiment, for the real-time iterative method, adopts a phased engineering implementation approach. First, in the preparation phase, the nonlinear prediction model is linearized by approximating it using a first-order Taylor series expansion at the current operating point, resulting in a linearized state-space model. In the formula Let be the system state vector. For the system output vector, To control the increment, For the state matrix, For the input matrix, The output matrices are all obtained by linearizing the current working point. The linearization error compensation term is used to reduce the model error caused by linearization approximation. Based on the linearized state-space model, the standard quadratic programming form of the model predictive control optimization problem is constructed, with the expression being: The constraints are In the formula To control the incremental sequence, The Hessian matrix is ​​a symmetric positive definite matrix. The gradient vector, For the constraint matrix, , These represent the lower and upper bounds of the constraint, respectively; for the pre-decomposition of the Hessian matrix, this embodiment uses the Cholesky decomposition method, and the decomposition formula is as follows: In the formula The matrix is ​​a lower triangular matrix. The decomposition result is directly stored in the controller's memory and called directly during the feedback phase. This pre-decomposition operation advances the most time-consuming matrix inversion operation in the quadratic programming solution, allowing the single solution time in the feedback phase to be controlled within 5ms. During the feedback phase of the current control cycle, the system first receives the operating status data collected in the current cycle as status measurement values, and dynamically updates the upper and lower bounds of the constraints of the quadratic programming problem based on the measurement values. and Then utilize the lower triangular matrix already stored in memory. By rapidly solving the quadratic programming problem through forward and backward substitution, the optimal control increment is obtained. Finally, the optimal control increment is superimposed with the control quantity of the previous cycle to obtain the optimal control quantity of the current cycle and output to the actuator 61. Through this phased real-time iteration method, more than 80% of the computation can be transferred to the non-real-time preparation stage, which greatly reduces the computational delay of real-time control and ensures the control response speed and accuracy of the valve when the operating conditions change rapidly.

[0047] This application proposes a deep learning-based adaptive predictive control system for an eccentric hemispherical valve. For example... Figure 2 and Figure 3 As shown, the system includes a sensing module 10, used to collect real-time operating status data of the eccentric hemispherical valve 60. This operating status data includes at least the opening position of the hemispherical valve 62, the upstream medium pressure, the downstream medium pressure, the medium flow rate through the eccentric hemispherical valve 60, and the torque output by the actuator 61 of the eccentric hemispherical valve 60. A long short-term memory network prediction module 20, connected to the sensing module 10, is used to receive the operating status data and predict the dynamic characteristics of the eccentric hemispherical valve 60 over multiple future control cycles. These dynamic characteristics include at least the future opening position, future medium pressure, future medium flow rate, and future sealing contact state. A sealing state determination module 30, connected to the long short-term memory network prediction module 20, is used to determine the sealing state based on the predicted future sealing contact. The state determines the initial opening position of the eccentric hemispherical valve 60 as it is about to enter the sealing and tightening stage. The model predictive control module 40, connected to the long short-term memory network prediction module 20 and the sealing state determination module 30, is used to construct and solve the model predictive control optimization problem based on the predicted future dynamic characteristics and the determined initial opening position to obtain the optimal control sequence. The objective function of the model predictive control optimization problem includes a tracking error term, a control increment smoothing term, and a torque smoothing constraint term for the sealing and tightening stage. The execution drive module 50, connected to the model predictive control module 40, is used to output the first control quantity in the optimal control sequence to the actuator 61 of the eccentric hemispherical valve 60, driving the actuator 61 to adjust the opening position of the hemisphere 62. The long short-term memory network prediction module 20 predicts the future sealing contact state in real time, and the sealing state determination module 30 dynamically determines the starting opening position of the sealing and pressing stage. The model prediction control module 40 can activate the torque smoothing constraint term in the objective function in advance, thereby effectively suppressing torque step change, adapting to the dynamic drift of the sealing contact starting angle, ensuring that the valve runs smoothly during the opening and closing process, and avoiding impact and jamming.

[0048] To verify the actual control effect of the deep learning-based adaptive predictive control method for eccentric hemispherical valves (hereinafter referred to as the DLAPC method) described in this invention, this embodiment builds a dynamic model of an industrial-grade eccentric hemispherical valve with a nominal diameter of DN150 and a rated torque of 120 N·m on the MATLAB / Simulink simulation platform. The model simulates the nonlinear torque characteristics, media pressure difference disturbances, sealing surface wear, and drift of the initial sealing contact angle during the valve's idling-sealing tightening stage. Three sets of control schemes are set up for comparative experiments: the first set uses traditional PID control, the second set uses conventional linear compensation control (LinearComp), and the third set uses the DLAPC method of this invention. All three sets of experiments use a unified simulation model, unified operating parameters, and a unified 20 ms sampling period to ensure the fairness and effectiveness of the comparison.

[0049] The experiment was set up with two typical operating conditions: the first type was the high voltage difference large drift condition (corresponding to...). Figure 7 The results shown are as follows: The medium pressure difference is set to 8 MPa, the initial angle drift of the seal contact is 1.5°, the wear coefficient of the seal surface is 1.2, the target opening is 80%, and step disturbance and pulse disturbance are superimposed; the second type is the small opening high friction condition (corresponding to...). Figure 8 The results are shown below. The medium pressure difference was set to 2 MPa, the initial sealing contact angle drift to 1.0°, the sealing surface wear coefficient to 1.3, and the target opening to 20%. Gradual and pulse disturbances were superimposed. During the experiment, key indicators such as the opening tracking curve, control output curve, settling time, overshoot, steady-state error, integral absolute error (IAE), and disturbance recovery time were simultaneously collected for the three control methods. The results are as follows: Figure 6 As shown.

[0050] Depend on Figure 6 A comparison of performance indicators shows that the DLAPC method of this invention significantly outperforms traditional PID control and linear compensation control in all core indicators: In terms of settling time, the DLAPC method is comparable to PID control, but approximately 25% shorter than linear compensation control; regarding overshoot, the DLAPC method achieves an overshoot of 78%, far exceeding the 20% of PID control and the 35% of linear compensation control. This result stems from the torque smoothing constraint and piecewise weight adjustment strategy employed by the DLAPC method during the sealing and clamping stage, which actively sacrifices some speed to suppress torque impact; and in terms of steady-state error, the DLAPC method has a steady-state error of 44%. The steady-state errors of both PID control and linear compensation control are below 5%. This difference is due to the adaptive compensation characteristics of the online learning mechanism of the DLAPC method for sealing surface wear, which achieves long-term sealing reliability by dynamically adjusting control parameters. In terms of integral absolute error (IAE), the linear compensation control has the lowest IAE value, followed by PID control, while the DLAPC method has the highest IAE value, further verifying the control logic of the DLAPC method that trades tracking error for torque stability. In terms of disturbance recovery time, the DLAPC method is comparable to PID control, but about 40% longer than linear compensation control, which meets its design goal of smooth control.

[0051] Combination Figure 7 As can be seen from the high pressure differential and large drift operating condition curve, traditional PID control and linear compensation control exhibit significant overshoot and oscillation in the initial stage during the tracking process of the target opening degree of 80%. The overshoot of PID control is about 100%, and the overshoot of linear compensation control is about 95%, and continuous oscillation occurs under disturbance. The opening degree tracking curve of the DLAPC method of this invention shows a steady upward trend with no significant overshoot. Under the action of step disturbance and pulse disturbance, the opening degree fluctuation amplitude is less than 5%, and the control output curve is smooth without drastic changes, effectively suppressing the torque step impact during the sealing and tightening stage.

[0052] Combination Figure 8 As can be seen from the curve of the small opening and high friction condition, the traditional PID control and linear compensation control have an initial overshoot of about 30% during the tracking process of the target opening of 20%, and jamming and oscillation occur in the small opening region, and the control output is frequently adjusted. The opening tracking curve of the DLAPC method of this invention is stable, the overshoot is less than 5%, and the opening remains stable under the action of gradual disturbance and pulse disturbance, and the control output is smooth, effectively avoiding the jamming and impact problems under the small opening and high friction condition.

[0053] In summary, the experimental results show that the DLAPC method of this invention achieves early prediction of the sealing contact state through a long short-term memory network prediction model, suppresses torque step abrupt changes through torque smoothing constraints, and adaptively compensates for the dynamic drift of the sealing contact initiation angle through an online learning update mechanism. Under complex industrial conditions, it can achieve stable control of the eccentric hemispherical valve while ensuring sealing reliability. Compared with traditional PID control and linear compensation control, it has significant advantages in torque impact suppression, long-term operational stability, and adaptability to operating conditions, meeting the long-term stable operation requirements of eccentric hemispherical valves in petrochemical, power energy, and urban water affairs.

[0054] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A deep learning-based adaptive predictive control method for an eccentric hemispherical valve, characterized in that, Includes the following steps: S100: Real-time acquisition of operating status data of the eccentric hemispherical valve, the operating status data including at least the hemispherical opening position, the medium pressure before the valve, the medium pressure after the valve, the medium flow rate through the eccentric hemispherical valve, and the torque output by the actuator of the eccentric hemispherical valve. S200: Input the operating status data into a pre-built long short-term memory network prediction model to predict the dynamic characteristics of the eccentric hemispherical valve in multiple future control cycles. The dynamic characteristics include at least the future opening position, future medium pressure, future medium flow rate, and future sealing contact state. S3 00: Based on the future sealing contact state, determine the initial opening position of the eccentric hemispherical valve as it is about to enter the sealing and tightening stage; S400: Based on the future dynamic characteristics and the initial opening position, construct and solve the model predictive control optimization problem to obtain the optimal control sequence of the eccentric hemispherical valve in the future control cycle, wherein the objective function of the model predictive control optimization problem includes a tracking error term, a control increment smoothing term, and a torque smoothing constraint term in the sealing and pressing stage. S500: Output the first control quantity in the optimal control sequence to the actuator of the eccentric hemispherical valve to drive the actuator to adjust the opening position of the hemispherical valve.

2. The deep learning-based adaptive predictive control method for eccentric hemispherical valves according to claim 1, characterized in that, Step S100 specifically includes: acquiring the torque output by the actuator through a torque sensor installed on the output shaft of the actuator; acquiring the medium pressure before and after the valve through pressure transmitters installed in the pipeline before and after the valve, respectively; acquiring the medium flow rate through the eccentric hemispherical valve through a flow meter installed on the valve outlet side; and acquiring the hemispherical opening position through an encoder linked to the actuator.

3. The deep learning-based adaptive predictive control method for eccentric hemispherical valves according to claim 1, characterized in that, The long short-term memory network prediction model in step S200 includes a main prediction model and a sealed contact state prediction sub-model; step S200 further includes: The sealing contact state prediction sub-model is constructed. The sealing contact state prediction sub-model adopts a long short-term memory network structure. Its input layer receives the hemispherical opening position, valve inlet medium pressure, historical torque sequence and sealing surface wear coefficient of the current cycle. The historical torque sequence is taken from the torque values ​​of several consecutive control cycles before the current cycle. The output layer of the sealing contact state prediction sub-model outputs two prediction values. The first prediction value is a binary classification result of whether the hemisphere and the valve seat are in a sealing contact state in each future control cycle. The second prediction value is the expected sealing compression start opening position. The operating status data is simultaneously input into the main prediction model and the sealing contact state prediction sub-model. The main prediction model outputs the future opening position, future medium pressure, and future medium flow rate. The sealing contact state prediction sub-model outputs the future sealing contact state and the expected sealing compression starting opening position.

4. The deep learning-based adaptive predictive control method for an eccentric hemispherical valve according to claim 3, characterized in that, The specific method for determining the initial opening position of the eccentric hemispherical valve as it is about to enter the sealing and tightening stage in step S300 is as follows: The predicted sealing clamping start position is read from the sealing contact state prediction sub-model, and this position is recorded as the first starting boundary value; Compare the current hemispherical opening position with the first initial boundary value; When the hemispherical opening position of the current cycle is less than the first starting boundary value and the difference between the two is greater than the preset angle threshold, it is determined that the valve is in the idling stage and has not yet entered the sealing and pressing stage, and the torque smoothing constraint item is set to the inactive state. When the hemispherical opening position of the current cycle is greater than or equal to the first starting boundary value, it is determined that the valve has entered the sealing and pressing stage, the torque smoothing constraint term is set to the active state, and the first starting boundary value is used as the starting boundary of the sealing and pressing stage.

5. The deep learning-based adaptive predictive control method for an eccentric hemispherical valve according to claim 4, characterized in that, Once the torque smoothing constraint term is activated, the following constraints are applied during the solution of the model predictive control optimization problem: The absolute value of the torque change rate of the actuator is limited to not exceeding a preset maximum torque change rate limit, which is pre-calibrated based on the rated torque of the eccentric hemispherical valve and the maximum allowable impact coefficient. When the estimated torque value of the control quantity corresponding to the sealing and clamping stage in the optimal control sequence obtained by solving the model predictive control optimization problem is lower than the preset minimum sealing and clamping torque threshold, a penalty term is added to force the solver to increase the torque output so that the actual torque is not lower than the minimum sealing and clamping torque threshold.

6. The deep learning-based adaptive predictive control method for an eccentric hemispherical valve according to claim 1, characterized in that, The step S400 in constructing the model predictive control optimization problem also includes a piecewise control weight adjustment step: The movement range of the hemisphere is divided into an idle range and a sealing and pressing range based on the initial opening position. The idle range is the range between the hemisphere's opening position and the initial opening position, and the sealing and pressing range is the range between the hemisphere's opening position and the initial opening position. When the predicted hemispherical opening position is within the idle interval, the weight matrix of the tracking error term in the objective function is set as the first weight matrix; When the predicted hemispherical opening position is within the sealing and pressing range, the weight matrix of the tracking error term in the objective function is set as the second weight matrix, and the weight of the torque smoothing constraint term is set to a non-zero value. The penalty weight for the opening position tracking error in the second weight matrix is ​​higher than the corresponding weight in the first weight matrix.

7. The deep learning-based adaptive predictive control method for an eccentric hemispherical valve according to claim 6, characterized in that, The segmented control weight adjustment step also includes interval boundary smoothing transition processing: during the transition control period when the predicted hemispherical opening position crosses from the idling interval to the sealing and pressing interval, the weight matrix of the tracking error term is transitioned from the first weight matrix to the second weight matrix periodically using linear interpolation, and the weight of the torque smoothing constraint term is increased from zero to the preset target value periodically.

8. The deep learning-based adaptive predictive control method for an eccentric hemispherical valve according to claim 1, characterized in that, It also includes online learning update steps: During the operation of the eccentric hemispherical valve, the operating status data and actual torque feedback value of each control cycle are continuously collected and stored in the loop buffer area. When the amount of new data accumulated in the circular buffer reaches a preset update threshold, data from the most recent continuous time period is extracted from the circular buffer to form an incremental training set. The current Long Short-Term Memory Network prediction model is trained using the incremental training set with mini-batch gradient descent. During training, a learning rate lower than the initial training learning rate is set, and the incremental training set is mixed with some historical samples stored in the circular buffer for training using the experience replay technique. The prediction error of the incrementally trained model is evaluated on an independent validation set. If the prediction error is reduced by more than a preset relative improvement threshold compared to the model before the update, the updated model is deployed to the online prediction module to replace the original model.

9. The deep learning-based adaptive predictive control method for an eccentric hemispherical valve according to claim 1, characterized in that, The solution to the model predictive control optimization problem in step S400 employs a real-time iterative method, specifically including: In the preparation phase of each control cycle, i.e. in the previous control cycle, the following operations are performed: the long short-term memory network prediction model is linearized based on the current operating point, a coefficient matrix in the form of a quadratic programming problem of the model prediction control optimization problem is constructed, the coefficient matrix is ​​pre-decomposed, and the decomposition result is stored in memory. During the feedback phase of the current control cycle, the following operations are performed: receive the current state measurement value, update the constraint boundary of the quadratic programming form according to the measurement value, solve the quadratic programming problem using the pre-decomposition results stored in memory, obtain the optimal control increment, and then calculate the optimal control quantity.

10. An adaptive predictive control system for an eccentric hemispherical valve based on deep learning, characterized in that, include: The sensing module is used to collect the operating status data of the eccentric hemispherical valve in real time. The operating status data includes at least the opening position of the hemispherical valve, the medium pressure before the valve, the medium pressure after the valve, the medium flow rate through the eccentric hemispherical valve, and the torque output by the actuator of the eccentric hemispherical valve. A long short-term memory network prediction module, connected to the sensing module, is used to receive the operating status data and predict the dynamic characteristics of the eccentric hemispherical valve in multiple future control cycles. The dynamic characteristics include at least the future opening position, future medium pressure, future medium flow rate, and future sealing contact state. A sealing state determination module, connected to the long short-term memory network prediction module, is used to determine the initial opening position of the eccentric hemispherical valve as it is about to enter the sealing and pressing stage based on the predicted future sealing contact state. The model predictive control module, connected to the long short-term memory network prediction module and the sealing state determination module, is used to construct and solve the model predictive control optimization problem based on the predicted future dynamic characteristics and the determined initial opening position to obtain the optimal control sequence. The objective function of the model predictive control optimization problem includes a tracking error term, a control increment smoothing term, and a torque smoothing constraint term for the sealing and pressing stage. The execution drive module, connected to the model prediction control module, is used to output the first control quantity in the optimal control sequence to the actuator of the eccentric hemispherical valve, driving the actuator to adjust the opening position of the hemispherical body.

Images