Intelligent vibration prediction and active control method and system for flow conveying pipeline

Abstract

The invention discloses an intelligent vibration prediction and active control method and system for a fluid conveying pipeline, and relates to the technical field of fluid conveying pipeline dynamic control. The method comprises the steps of 1, data acquisition and preprocessing, 2, dynamic response prediction based on a hybrid deep learning model, 3, self-adaptive active control based on reinforcement learning, and 4, model online updating and maintenance, a feature data set is obtained through acquisition of multi-dimensional parameters and preprocessing, and the dynamic response prediction based on the hybrid deep learning model is obtained. A hybrid prediction model of a physical information neural network (PINN) and a long-short-term memory network is constructed to accurately capture a vibration response rule, a control instruction is dynamically generated through a deep reinforcement learning agent, model suitability is guaranteed through online updating, the problems that traditional vibration prediction is large in deviation and control lags are solved, and the accuracy of vibration prediction is improved. And the vibration prediction precision and reliability of the fluid conveying pipeline are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of dynamic control technology for transmission pipelines, specifically to an intelligent vibration prediction and active control method and system for transmission pipelines. Background Technology

[0002] Dynamic response control of transmission pipelines is one of the key technologies to ensure the safe operation of industrial systems. With the expansion of industrial production scale, transmission pipelines are gradually developing towards long distances, high pressures, and large flow rates, which significantly increases the complexity of their operating conditions. Unsteady flow of internal fluids can cause pressure pulsations on the pipeline wall, and vibrations of external equipment can transmit excitations through the pipeline supports. Extreme operating conditions can also exacerbate the structural coupling effect of the fluid. All these factors can induce complex dynamic responses in transmission pipelines. However, in recent years, deep learning has shown significant advantages in industrial modeling and control due to its powerful nonlinear fitting and time series prediction capabilities. Therefore, it is necessary to study an intelligent vibration prediction and active control method and system for transmission pipelines.

[0003] Existing technologies, such as the invention patent application with announcement number CN115293054B, disclose a method for evaluating the failure of dented pipes based on neural networks. This method involves acquiring training samples of the dented pipe, dividing the dataset, constructing and training a BP neural network model, and substituting the model into the predicted strain data for failure evaluation. This solves the problems of difficulty in obtaining stress and strain and low accuracy in existing evaluation methods, achieving rapid and accurate assessment and improving evaluation efficiency. Another existing technology, such as the invention patent application with announcement number CN119862762B, discloses a deep learning method combining physical information neural networks and finite element analysis. This method includes embedding partial differential equations into a neural network, followed by mesh generation, data collection, loss function construction, training, and optimization convergence steps. By integrating physical information neural networks and finite element analysis, it demonstrates significant advantages and beneficial effects in multiple dimensions.

[0004] As can be seen from the above solutions, current traditional technologies for predicting dynamic response mostly rely on finite element modeling combined with experimental verification. By establishing a geometric model of the conveying pipeline, defining material properties and boundary conditions, the response characteristics under different operating conditions are simulated. However, this method requires simplification of a large number of nonlinear factors, resulting in a large deviation between the predicted vibration results and the actual operating conditions. Active control often adopts a scheme of sensor acquisition combined with PID control. The parameters of the PID controller need to be preset manually and cannot be dynamically adjusted. When the conveying pipeline experiences a sudden vibration response, there is a significant lag, which can easily miss the best control opportunity. In addition, the application of existing neural network technology in this field only focuses on single aspects such as energy consumption prediction and does not fully explore the value of historical operating data, resulting in poor model generalization ability and difficulty in adapting to the needs of conveying pipeline systems of different specifications and operating conditions. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to provide an intelligent vibration prediction and active control method and system for transmission pipelines.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The first aspect of the present invention provides an intelligent vibration prediction and active control method for a conveying pipeline, including step 1. Data acquisition and preprocessing: acquiring internal fluid parameters, pipeline structural parameters and external environmental parameters that affect the dynamic response of the conveying pipeline, and preprocessing the time series data of each acquired parameter.

[0007] Preferably, the preprocessing of the time-series data of each parameter is specifically performed as follows: abnormal data points that do not conform to the 3σ criterion are removed from the time-series data of each parameter collected by each sensor, and the time-series data of each parameter is denoised using wavelet threshold denoising technology. Feature engineering is performed on the processed time-series data of each parameter, and a feature dataset for model training is constructed by extracting its time-domain features and frequency-domain features. The feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio.

[0008] Step 2. Dynamic response prediction based on a hybrid deep learning model: Construct a hybrid prediction model of a physical information neural network (PINN) and a long short-term memory network (LSTM). Train and optimize the model based on preprocessed time series data, and predict the dynamic response parameters of the conveying pipeline in future time steps.

[0009] Preferably, the method for constructing the hybrid prediction model of the Physical Information Neural Network (PINN) and the Long Short-Term Memory Network (LSTM) is as follows: the hybrid prediction model is constructed by the synergistic effect of the PINN and the LSTM. The input layer receives the preprocessed feature dataset, and the input dimension is set according to the number of indicators of the time-domain and frequency-domain features extracted by feature engineering. The LSTM serves as a hidden layer, capturing the temporal dependency of the pipeline dynamic response through an internal control mechanism. The fully connected layer is used to fuse the features output by the LSTM. The PINN constraint layer serves as the top layer, physically constraining the pipeline dynamics by using the control equations of the pipeline dynamics as the regularization term of the loss function. The output layer outputs the predicted values ​​of the pipeline dynamic response parameters for future time steps.

[0010] Preferably, the specific method for training and optimizing the model is as follows: adding a mean squared error formula and a physical constraint formula to the hybrid prediction model to form a composite loss function. ,in These are the weighting coefficients for the physical constraint formula; Mean square error formula This represents the deviation between the predicted values ​​of the hybrid prediction model and the actual data, where... Indicates the first The true values ​​of the dynamic response parameters of the pipeline corresponding to each data point. Indicates the mixed prediction model for the first The predicted values ​​of the dynamic response parameters of the pipeline corresponding to each data point. , It is a positive integer greater than 2. This represents the total number of data points in the training set.

[0011] Physical constraint formula This represents the deviation between the output of the hybrid prediction model and the physical equations, where The governing equations that govern the dynamics of the transport pipeline are represented. Indicates the hybrid prediction model in the 1st... The predicted physical field quantities and their first and second derivatives for each physical constraint point. , It is a positive integer greater than 2. This represents the total number of sample points randomly selected within the computational domain of the transport pipeline to apply physical constraints.

[0012] During the training of the hybrid prediction model, the Adam optimizer is used to set the initial learning rate, and the learning rate is dynamically adjusted through the cosine annealing strategy. If the composite loss function value on the validation set does not decrease within the preset number of training iterations, the training of the hybrid prediction model is stopped, which constitutes the optimization of the model parameters.

[0013] Preferably, the method for predicting the dynamic response parameters of the pipeline in future time steps is as follows: inputting the time series data of each parameter after preprocessing in the historical and current time steps into the trained hybrid prediction model, capturing the time series dependency of the pipeline dynamic response through the Long Short-Term Memory (LSTM) network in the hidden layer and learning its dynamic evolution law, and ensuring that each prediction step in the prediction process satisfies the constraints of the pipeline dynamic control equation in the PINN constraint layer of the physical information neural network, and finally directly generating the dynamic response parameters of the pipeline in future time steps through the output layer, including the peak value of vibration acceleration and the maximum strain value of the wall.

[0014] Step 3. Reinforcement learning-based adaptive active control: Set the core safety control target based on the predicted dynamic response parameters of the pipeline, and construct a prediction-decision-feedback deep reinforcement learning control framework. Through real-time interaction between the deep learning agent and the pipeline system, generate the optimal control command and drive the actuator to respond.

[0015] Preferably, the specific method for setting the core safety control target is as follows: based on the predicted future time step dynamic response parameters of the conveying pipeline, if the peak value of its vibration acceleration is less than or equal to the preset gravity acceleration threshold, and the maximum strain value of the wall meets the numerical range of the yield strain of the conveying pipeline wall, then it is set as the core safety control target.

[0016] Preferably, the specific method for constructing the prediction-decision-feedback deep reinforcement learning control framework is as follows: the dynamic response parameters of the conveying pipeline predicted by the hybrid prediction model at a future preset time step and the current pipeline operating parameters are used together as the state space, and the adjustable parameters of the active control actuator are used as the action space. A reward function is set based on the state space and the action space. The reward function sets an output value according to whether the predicted dynamic response parameters of the conveying pipeline at a future preset time step in the state space meet the set core safety control target. If the predicted dynamic response parameters of the conveying pipeline at a future preset time step in the state space meet the set core safety control target, the reward function outputs a positive reward value; otherwise, the reward function outputs a negative penalty value. This is how the prediction-decision-feedback deep reinforcement learning control framework is constructed.

[0017] Preferably, the method for generating the optimal control command and driving the actuator to respond is as follows: the deep reinforcement learning agent outputs the corresponding control parameter value in the action space according to the current state space, generates the corresponding control command for the active control actuator based on the control parameter value, and transmits the control command to the active control actuator in real time through the industrial bus and completes the execution within a specified time.

[0018] Step 4. Online Model Update and Maintenance: Regularly update and maintain the parameters of the hybrid prediction model and the deep reinforcement learning agent.

[0019] Preferably, the method for periodically updating and maintaining the parameters of the hybrid prediction model and the deep reinforcement learning agent is as follows: within a preset period, the newly collected internal fluid parameters, pipe structure parameters, and external environmental parameters affecting the dynamic response of the conveying pipe are preprocessed and added to the training set. Incremental training is performed on the hybrid prediction model and the deep reinforcement learning agent. When the prediction error of the hybrid prediction model on the test set is greater than a preset prediction error threshold or the control response delay is greater than a preset allowable delay threshold, full training of the hybrid prediction model and the deep reinforcement learning agent is triggered to update their parameters.

[0020] Before triggering incremental or full training of the hybrid prediction model and deep reinforcement learning agent, the signal quality of each sensor channel is evaluated.

[0021] The signal-to-noise ratio (SNR) and packet loss rate of each sensor are obtained from the local database. If the SNR of a certain sensor channel is continuously less than the preset SNR threshold or the packet loss rate is greater than the preset packet loss rate threshold within the target period, the sensor is determined to be faulty.

[0022] When a single sensor is determined to be faulty, the real-time data from the remaining normal sensors is input into the pre-trained hybrid prediction model. The predicted values ​​of the hybrid prediction model for the fault point parameters are used as reconstructed data to replace the real data of the faulty sensor and input into the deep reinforcement learning agent to generate control commands.

[0023] When it is determined that three or more sensors fail simultaneously or the prediction error of the hybrid prediction model increases to the preset warning threshold within the target period, the deep reinforcement learning agent control command generation function is suspended. The system directly issues a set of preset basic safety action commands to the active control actuator and triggers an audible and visual alarm to request manual intervention.

[0024] A second aspect of the present invention provides a system for intelligent vibration prediction and active control of a conveying pipeline, comprising a data acquisition and preprocessing module: used to acquire internal fluid parameters, pipeline structural parameters and external environmental parameters that affect the dynamic response of the conveying pipeline, and to preprocess the time-series data of each acquired parameter.

[0025] The dynamic response prediction module based on a hybrid deep learning model is used to construct a hybrid prediction model of the physical information neural network PINN and the long short-term memory network LSTM. The model is trained and its parameters are optimized based on preprocessed time series data, and the dynamic response parameters of the conveying pipeline at future time steps are predicted.

[0026] The reinforcement learning-based adaptive active control module is used to set the core safety control objectives based on the predicted dynamic response parameters of the pipeline and to construct a prediction-decision-feedback deep reinforcement learning control framework. Through real-time interaction between the deep learning agent and the pipeline system, it generates optimal control commands and drives the actuators to respond.

[0027] Online Model Update and Maintenance Module: Used to periodically update and maintain the parameters of the hybrid prediction model and the deep reinforcement learning agent.

[0028] The beneficial effects of the present invention are as follows: (1) The first part of the present invention: high-frequency acquisition by multiple sensors ensures the comprehensiveness of data, preprocessing improves the quality of data, lays the foundation for subsequent links, the hybrid prediction model accurately outputs key parameters such as peak vibration acceleration, the reinforcement learning framework dynamically generates control instructions, suppresses vibration over-limit in advance, ensures the integrity of pipeline structure, reduces fatigue damage, optimizes control energy consumption, adapts to the safe operation requirements of different specifications and working conditions of transmission pipelines in industrial scenarios, and has clear practical application value.

[0029] (2) The second part of the present invention: Breaking through the limitations of large prediction deviation and lag of traditional finite element modeling and PID control, the invention innovatively integrates the physical information neural network PINN and the long short-term memory network LSTM to construct a hybrid prediction model, which captures both time-series dependencies and applies physical constraints, thereby improving prediction accuracy and reliability. It also constructs a deep reinforcement learning control framework to achieve adaptive optimization of control strategies, solves the problem of lag in control of sudden vibrations, and forms a full-process intelligent system of prediction-control-update, fully explores the value of data, and improves the generalization ability of the model.

[0030] (3) The third part of the present invention: By deeply integrating predictive safety objectives with adaptive control, an intelligent control system capable of suppressing sudden vibrations in a forward-looking, precise and efficient manner is constructed, thereby comprehensively surpassing the limitations of traditional passive response control. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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.

[0032] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0033] Figure 2 This is a schematic diagram of the system modules of the present invention.

[0034] Figure 3 This is an overall framework diagram of the present invention. Detailed Implementation

[0035] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Reference Figure 1 As shown, the first aspect of the present invention provides an intelligent vibration prediction and active control method for a conveying pipeline, including step 1. Data acquisition and preprocessing: acquiring internal fluid parameters, pipeline structural parameters and external environmental parameters that affect the dynamic response of the conveying pipeline, and preprocessing the time series data of each acquired parameter.

[0037] The internal fluid parameters include the fluid velocity and pressure inside the conveying pipe; the structural parameters of the conveying pipe include the vibration acceleration and strain of the conveying pipe; and the external environmental parameters include the vibration of the conveying pipe support and the ambient temperature.

[0038] The strain refers to the structural deformation of the conveying pipe measured by strain gauges deployed on the pipe wall, which is used to characterize the stress state of the pipe wall under the action of internal fluid and external load.

[0039] It should be noted that the flow velocity and pressure in the internal fluid parameters are measured by ultrasonic flow meters installed in the straight sections of the conveying pipeline and pressure sensors installed at pumps, valves and elbows, respectively. The vibration acceleration in the pipeline structure parameters is measured by acceleration sensors installed at the midpoint of the span of the conveying pipeline, elbows and valves. The vibration of the pipeline support and the ambient temperature in the external environmental parameters are measured by acceleration sensors installed on the support of the conveying pipeline and temperature sensors installed on the pipe wall or in the environment, respectively.

[0040] It should be noted that the various types of sensors deployed at key locations in the transmission pipeline need to collect data at a frequency of 100Hz to ensure the capture of high-frequency dynamic responses, such as high-frequency vibrations caused by fluid pulsation.

[0041] The pipeline dynamic response parameters include peak vibration acceleration and maximum wall strain.

[0042] In a specific embodiment of the present invention, the preprocessing of the time series data of each parameter is specifically performed as follows: abnormal data points that do not conform to the 3σ criterion are removed from the time series data of each parameter collected by each sensor, and the time series data of each parameter is denoised using wavelet threshold denoising technology. Feature engineering is performed on the processed time series data of each parameter, and a feature dataset for model training is constructed by extracting its time domain features and frequency domain features. The feature dataset is divided into a training set, a validation set, and a test set according to a preset ratio.

[0043] It should be noted that the 3σ criterion is a classic statistical data analysis method. Its basic principle is based on the assumption of normal distribution, which regards data points that deviate from the mean by more than three standard deviations as low-probability events and discards them. This invention applies this criterion to efficiently and automatically identify and remove outliers caused by sensor momentary failures or electromagnetic interference, providing a high-quality data foundation for subsequent analysis. This criterion is a well-known technology in this field, and those skilled in the art can flexibly configure it according to specific scenario requirements, which will not be described in detail here.

[0044] It should be noted that wavelet thresholding denoising technology decomposes the signal into different scales through wavelet transform and performs thresholding on the wavelet coefficients at each scale, thereby achieving effective separation of signal and noise. This invention specifically points out that it is applied to the time-series data denoising of various parameters of the conveying pipeline to remove environmental noise and high-frequency interference and extract effective signals that reflect the true dynamic state of the conveying pipeline. The above threshold settings are all common knowledge in the prior art, and those skilled in the art can flexibly configure them based on conventional experience and specific scenarios without affecting the implementation and reproduction of the technical solution of this invention.

[0045] It should be noted that the extraction methods and specific indicators of the time-domain and frequency-domain features involved in this invention are well-known technologies in the field of signal processing. In time-domain analysis, indicators such as mean, variance, root mean square, peak value, and kurtosis are used to quantify the statistical laws and waveform characteristics of signals. In frequency-domain analysis, conventional indicators such as spectral centroid, frequency standard deviation, and spectral energy are used to reveal the frequency components and energy distribution of signals through techniques such as fast Fourier transform. The definitions, calculation methods, and engineering significance of these features are common knowledge in related technical fields and are used as key steps in constructing hybrid prediction models. Therefore, the selection and combination of specific features can be flexibly configured by those skilled in the art based on conventional experience and specific scenarios, without affecting the complete implementation of the core technical solution and the realization of the technical effect of this invention.

[0046] It should be noted that preprocessing the time-series data of each collected parameter ensures the accuracy and usability of the data, providing high-quality input for the prediction of the hybrid prediction model.

[0047] For example, the dataset can be divided into training set, validation set and test set according to a preset ratio, such as dividing the dataset into training set, validation set and test set in a ratio of 7:2:1.

[0048] Step 2. Dynamic response prediction based on a hybrid deep learning model: Construct a hybrid prediction model of a physical information neural network (PINN) and a long short-term memory network (LSTM). Train and optimize the model based on preprocessed time series data, and predict the dynamic response parameters of the conveying pipeline in future time steps.

[0049] In a specific embodiment of the present invention, the method for constructing a hybrid prediction model of a Physical Information Neural Network (PINN) and a Long Short-Term Memory (LSTM) network is as follows: the hybrid prediction model is constructed by the synergistic effect of the PINN and the LSTM. The input layer receives the preprocessed feature dataset, and the input dimension is set according to the number of indicators of time-domain and frequency-domain features extracted by feature engineering. The LSTM network, as a hidden layer, captures the temporal dependency of the pipeline dynamic response through an internal control mechanism. The fully connected layer is used to fuse the features output by the LSTM network. The PINN constraint layer, as the top layer, performs physical constraints by using the control equation of the pipeline dynamics as a regularization term of the loss function. The output layer outputs the predicted values ​​of the pipeline dynamic response parameters for future time steps.

[0050] In one specific embodiment, the input dimension is set according to the number of indicators of time-domain and frequency-domain features extracted by feature engineering. The specific setting method is as follows: based on the sensors deployed at key monitoring points, time-domain and frequency-domain features are extracted from the 100Hz time-series data collected by each sensor channel. The total number of features is determined according to the number of statistical indicators of time-domain and frequency-domain features and the number of key monitoring points. The total number of features is the product of the number of key monitoring points and the number of feature indicators of each key monitoring point. This value is the input dimension of the input layer of the hybrid prediction model.

[0051] For example, the input dimension is set according to the number of indicators of time-domain and frequency-domain features extracted by feature engineering. For instance, feature extraction is performed on 100Hz time-series data collected by sensors deployed at key monitoring points of the transmission pipeline. The time-domain features are selected from five statistical indicators: mean, variance, root mean square, peak value, and kurtosis. The frequency-domain features are extracted from five indicators: spectral centroid, frequency standard deviation, and energy of three main frequency bands through fast Fourier transform. Each key monitoring point obtains 10 feature indicators. The number of key monitoring points is determined according to actual monitoring needs. When there are 4 key monitoring points, the total number of features is the product of the number of key monitoring points and the number of feature indicators for each key monitoring point, i.e., 4 × 10 = 40 dimensions. This value is the input dimension of the input layer of the hybrid prediction model.

[0052] For example, key monitoring points include: bends in the transmission pipeline, midpoints of the transmission pipeline span, etc.

[0053] It should be noted that the gate control mechanism of the Long Short-Term Memory (LSTM) network achieves the function of capturing temporal dependencies through the coordinated operation of the input gate, forget gate, and output gate. This process does not require detailed description of its internal parameter configuration to achieve the objective of this invention. The application of LSTM in this invention focuses on its temporal modeling capabilities in industrial scenarios, rather than structural improvements to the algorithm itself, and does not affect the complete implementation of the core technical solution and the realization of the technical effect of this invention.

[0054] The governing equations of the dynamics of the conveying pipeline are not a single, fixed equation, but a general term for a class of physical laws that can describe the dynamic behavior of the conveying pipeline, such as the Euler-Bernoulli equation, the heat conduction equation, and the pipeline vibration equation. Those skilled in the art can flexibly configure them based on conventional experience and specific scenarios without affecting the implementation and reproduction of the technical solution of this invention.

[0055] It should be noted that the PINN constraint layer of the physical information neural network uses the control equations of the pipeline dynamics as the regularization term of the loss function to impose physical constraints in order to force the output of the hybrid prediction model to meet physical laws, avoid the prediction results from deviating from the actual physical meaning, and achieve prediction results with high accuracy, high reliability and strong generalization ability.

[0056] In a specific embodiment of the present invention, the method for training the model and optimizing its parameters is as follows: a composite loss function is constructed by adding a mean squared error formula and a physical constraint formula to the hybrid prediction model. ,in These are the weighting coefficients for the physical constraint formula; Mean square error formula This represents the deviation between the predicted values ​​of the hybrid prediction model and the actual data, where... Indicates the first The true values ​​of the dynamic response parameters of the pipeline corresponding to each data point. Indicates the mixed prediction model for the first The predicted values ​​of the dynamic response parameters of the pipeline corresponding to each data point. , It is a positive integer greater than 2. This represents the total number of data points in the training set.

[0057] Physical constraint formula This represents the deviation between the output of the hybrid prediction model and the physical equations, where The governing equations that govern the dynamics of the transport pipeline are represented. Indicates the hybrid prediction model in the 1st... The predicted physical field quantities and their first and second derivatives for each physical constraint point. , It is a positive integer greater than 2. This represents the total number of sample points randomly selected within the computational domain of the transport pipeline to apply physical constraints.

[0058] During the training of the hybrid prediction model, the Adam optimizer is used to set the initial learning rate, and the learning rate is dynamically adjusted through the cosine annealing strategy. If the composite loss function value on the validation set does not decrease within the preset number of training iterations, the training of the hybrid prediction model is stopped, which constitutes the optimization of the model parameters.

[0059] It should be noted that the weighting coefficients in the physical constraint formula It is obtained through a systematic hyperparameter tuning process based on validation set evaluation, with values ​​between 0.1 and 1.0, used to balance the contributions of data fitting and physical constraints, and to avoid overfitting or underfitting of the model.

[0060] It should be noted that the settings This is to embed physical laws as soft constraints into the hybrid prediction model to ensure that the prediction results conform to physical reality.

[0061] For example, physical field quantities such as fluid pressure and velocity, or displacement, stress, and strain of a transport pipeline structure, and the derivatives of physical field quantities such as pressure gradient, velocity divergence, and strain derivatives, reflect the changing laws of the physical field.

[0062] It should be noted that the computational domain of the transmission pipeline refers to the geometric space in which the entire transmission pipeline is located. The physical constraint points are not sensors, but virtual coordinate points selected within the geometric space in which the entire transmission pipeline is located, in order to check whether the prediction results of the hybrid prediction model satisfy the physical equations.

[0063] It should be noted that the Adam optimizer is an adaptive learning rate gradient descent algorithm that combines the advantages of momentum and RMSProp. It dynamically adjusts the learning rate of each parameter by calculating the first and second moments of the gradient. It can automatically adapt to the characteristics of different parameters, avoiding the tediousness of manual parameter tuning. It is particularly suitable for processing nonlinear, high-dimensional time-series data such as the dynamic response of conveying pipelines, thereby accelerating the training process and reducing the risk of getting trapped in local optima.

[0064] It should be noted that the initial learning rate needs to be set to 0.001 because 0.001 is a commonly used empirical value that can balance stability and optimization efficiency in deep learning tasks. It also represents the step size at the beginning of training the hybrid prediction model. Furthermore, the learning rate determines the magnitude of parameter updates. If it is too large, it may cause the loss function to oscillate or diverge. If it is too small, convergence will be slow.

[0065] It should be noted that cosine annealing is a method for dynamically adjusting the learning rate. It periodically reduces the learning rate by simulating the decay law of the cosine function. That is, the learning rate starts from the initial value and gradually decreases to near zero according to the cosine curve as the number of training iterations increases. The advantage of this strategy is that it can prevent the mixed prediction model from converging to a suboptimal point too early. In the early stage of training, a higher learning rate helps to quickly explore the parameter space, while a lower learning rate in the later stage allows for fine-tuning and improves the model's generalization ability.

[0066] It should be noted that by adding the mean square error formula and the physical constraint formula to the hybrid prediction model to form a composite loss function, and by synergistically integrating data-driven fitting and physical law constraints, the single prediction of the traditional model is transformed into an intelligent prediction with high accuracy, strong generalization ability and physical reliability. This ensures that the prediction results strictly follow the physical laws under any operating conditions, thereby providing a solid and reliable decision basis for subsequent adaptive active control, and fundamentally guaranteeing the safety and long-term effectiveness of the transmission pipeline system.

[0067] In a specific embodiment of the present invention, the method for predicting the dynamic response parameters of the conveying pipeline in future time steps is as follows: the time series data of each parameter after preprocessing in the historical and current time steps are input into the trained hybrid prediction model; the temporal dependency of the dynamic response of the conveying pipeline is captured by the Long Short-Term Memory (LSTM) network in the hidden layer and its dynamic evolution law is learned; and the constraint layer of the Physical Information Neural Network (PINN) ensures that each prediction step in the prediction process satisfies the constraints of the dynamic control equation of the conveying pipeline; finally, the dynamic response parameters of the conveying pipeline in future time steps, including the peak value of vibration acceleration and the maximum strain value of the wall, are directly generated through the output layer.

[0068] For example, the temporal dependence of the dynamic response of the transmission pipeline can be captured. For instance, when the pump starts or stops or the valve is quickly adjusted, a pressure pulsation is generated in the transmission pipeline. This pressure wave will propagate along the transmission pipeline at a certain wave speed and be reflected and superimposed at the boundaries such as bends and supports. Therefore, the peak pressure or vibration acceleration measured by the sensor at a certain point downstream of the transmission pipeline will appear slightly later than that of the upstream source point. The Long Short-Term Memory (LSTM) network can learn this fixed time delay pattern, that is, the state of the downstream point depends on the state of the upstream point tens or even hundreds of milliseconds ago.

[0069] For example, the dynamic evolution of the dynamic response of a transmission pipeline can be captured. In actual working conditions, the transmission pipeline may be subjected to multiple excitations from pumps, valves and the external environment at the same time. When the vibration waves generated by these excitations meet in the transmission pipeline, nonlinear superposition and interference will occur, resulting in an extremely complex dynamic response. The powerful nonlinear fitting ability of the Long Short-Term Memory (LSTM) network can learn the comprehensive evolution law under the combined action of multiple factors, thereby accurately predicting the peak value of vibration acceleration at future moments in complex input signals.

[0070] It should be noted that by using a hybrid prediction model to accurately predict the future evolution trend of pipeline dynamic response parameters, intelligent and forward-looking judgment of pipeline vibration can be achieved, ultimately providing a reliable data foundation for the decision generation of adaptive active control.

[0071] Step 3. Reinforcement learning-based adaptive active control: Set the core safety control target based on the predicted dynamic response parameters of the pipeline, and construct a prediction-decision-feedback deep reinforcement learning control framework. Through real-time interaction between the deep learning agent and the pipeline system, generate the optimal control command and drive the actuator to respond.

[0072] In a specific embodiment of the present invention, the method for setting the core safety control target is as follows: based on the predicted future time step dynamic response parameters of the conveying pipeline, if the peak value of its vibration acceleration is less than or equal to a preset gravity acceleration threshold, and the maximum strain value of the wall meets the numerical range of the yield strain of the conveying pipeline wall, then it is set as the core safety control target.

[0073] It should be noted that setting core safety objectives aims to prioritize the structural integrity of the transmission pipeline, avoid the accumulation of fatigue damage caused by exceeding response limits, thereby significantly extending the service life of the transmission pipeline, reducing energy consumption of equipment such as electromagnetic actuators, and improving control efficiency.

[0074] In a specific embodiment of the present invention, the method for constructing the prediction-decision-feedback deep reinforcement learning control framework is as follows: the dynamic response parameters of the conveying pipeline at a future preset time step predicted by the hybrid prediction model and the current pipeline operating parameters are used together as the state space, and the adjustable parameters of the active control actuator are used as the action space. A reward function is set based on the state space and the action space. The reward function sets an output value according to whether the predicted dynamic response parameters of the conveying pipeline at a future preset time step in the state space meet the set core safety control target. If the predicted dynamic response parameters of the conveying pipeline at a future preset time step in the state space meet the set core safety control target, the reward function outputs a positive reward value; otherwise, the reward function outputs a negative penalty value. This is how the prediction-decision-feedback deep reinforcement learning control framework is constructed.

[0075] The current pipeline operating parameters include internal fluid parameters, pipeline structural parameters, and external environmental parameters that affect the dynamic response of the transport pipeline.

[0076] For example, adjustable parameters of the active control actuator include: the output force of the electromagnetic actuator and the damping coefficient of the adjustable damper.

[0077] For example, the preset time steps in the future may be 10 time steps, 15 time steps, etc.

[0078] In a specific embodiment of the present invention, the method for generating the optimal control command and driving the actuator to respond is as follows: the deep reinforcement learning agent outputs the corresponding control parameter value in the action space according to the current state space, generates the corresponding control command for the active control actuator based on the control parameter value, and transmits the control command to the active control actuator in real time through the industrial bus and completes the execution within a specified time.

[0079] It should be noted that the deep reinforcement learning agent outputs the corresponding control parameter values ​​in the action space based on the current state space because the policy network inside the agent, which has been optimized and trained by the reward function, can achieve intelligent mapping from complex state information to precise control commands. The policy network performs instantaneous calculations based on these inputs and directly maps the optimal control parameter values ​​through its internalized reward function decision logic.

[0080] For example, control commands for the corresponding active control actuator of the pipeline system are generated based on control parameter values. For instance, when the status of the pipeline system shows that the vibration acceleration at a certain bend in the pipeline is predicted to exceed the limit and the current pressure rises sharply, the strategy network outputs a command to apply a force of 1500N to the corresponding active control actuator A03. When the status of the pipeline system shows that the wall strain at the midpoint of a certain span of the pipeline is predicted to continue to increase and approach 80% of the material yield strain, and the spectrum analysis shows that the main vibration energy is concentrated in the low frequency band, the strategy network calculates and outputs a command to adjust the damping coefficient of the adjustable damper to 1200N·s / m.

[0081] It should be noted that the actions of the active control actuator directly affect the flow pipeline system, changing its dynamic state. Sensors deployed on the pipeline monitor these state changes in real time and feed the updated parameter data back to the data acquisition and preprocessing module, thereby updating the state space on which the deep reinforcement learning agent depends, forming a real-time interactive closed loop of prediction-decision-execution-feedback.

[0082] Step 4. Online Model Update and Maintenance: Regularly update and maintain the parameters of the hybrid prediction model and the deep reinforcement learning agent.

[0083] Reference Figure 3 As shown in the specific embodiment of the present invention, the method for periodically updating and maintaining the parameters of the hybrid prediction model and the deep reinforcement learning agent is as follows: within a preset period, the newly collected internal fluid parameters, pipe structure parameters, and external environmental parameters affecting the dynamic response of the conveying pipe are preprocessed and added to the training set. Incremental training is performed on the hybrid prediction model and the deep reinforcement learning agent. When the prediction error of the hybrid prediction model on the test set is greater than a preset prediction error threshold or the control response delay is greater than a preset allowable delay threshold, full training of the hybrid prediction model and the deep reinforcement learning agent is triggered to update their parameters.

[0084] Before triggering incremental or full training of the hybrid prediction model and deep reinforcement learning agent, the signal quality of each sensor channel is evaluated.

[0085] The signal-to-noise ratio (SNR) and packet loss rate of each sensor are obtained from the local database. If the SNR of a certain sensor channel is continuously less than the preset SNR threshold or the packet loss rate is greater than the preset packet loss rate threshold within the target period, the sensor is determined to be faulty.

[0086] When a single sensor is determined to be faulty, the real-time data from the remaining normal sensors is input into the pre-trained hybrid prediction model. The predicted values ​​of the hybrid prediction model for the fault point parameters are used as reconstructed data to replace the real data of the faulty sensor and input into the deep reinforcement learning agent to generate control commands.

[0087] When it is determined that three or more sensors fail simultaneously or the prediction error of the hybrid prediction model increases to the preset warning threshold within the target period, the deep reinforcement learning agent control command generation function is suspended. The system directly issues a set of preset basic safety action commands to the active control actuator and triggers an audible and visual alarm to request manual intervention.

[0088] The signal-to-noise ratio (SNR) is a core indicator for measuring signal quality. It is the ratio of useful signal power to background noise power. In this invention, useful signal power refers to effective data that reflects the true state of the transmission pipeline, and background noise power refers to all interference signals.

[0089] The data packet loss rate is an indicator that measures the integrity and reliability of data transmission. It is the ratio of the number of lost data packets to the total number of data packets that should have been received per unit of time.

[0090] For example, the system directly issues a set of preset safety basic action commands to the active control actuator. For instance, if the system detects that multiple vibration acceleration sensors at key monitoring points of the conveying pipeline fail simultaneously, and the prediction error of the hybrid prediction model increases to more than 15% within the target period, the system issues a command to the electromagnetic actuator located at the key monitoring point, numbered EMA-03: output a constant damping force opposite to the vibration velocity of the conveying pipeline, with a force value of 800 Newtons and a duration of 5 seconds. If the system detects abnormal data from multiple sensors used to monitor the wall strain at the midpoint of the conveying pipeline span, and the ambient temperature sensor also fails, the system issues a command to the adjustable damper located at the midpoint of the conveying pipeline span, numbered VSD-01: lock the damping coefficient to the maximum level, i.e., 1200 N·s / m, and maintain this state until manual intervention.

[0091] Reference Figure 2As shown, a second aspect of the present invention provides a system for intelligent vibration prediction and active control of a conveying pipeline, comprising a data acquisition and preprocessing module: used to acquire internal fluid parameters, pipeline structural parameters and external environmental parameters that affect the dynamic response of the conveying pipeline, and to preprocess the time-series data of each acquired parameter.

[0092] The dynamic response prediction module based on a hybrid deep learning model is used to construct a hybrid prediction model of the physical information neural network PINN and the long short-term memory network LSTM. The model is trained and its parameters are optimized based on preprocessed time series data, and the dynamic response parameters of the conveying pipeline at future time steps are predicted.

[0093] The reinforcement learning-based adaptive active control module is used to set the core safety control objectives based on the predicted dynamic response parameters of the pipeline and to construct a prediction-decision-feedback deep reinforcement learning control framework. Through real-time interaction between the deep learning agent and the pipeline system, it generates optimal control commands and drives the actuators to respond.

[0094] Online Model Update and Maintenance Module: Used to periodically update and maintain the parameters of the hybrid prediction model and the deep reinforcement learning agent.

[0095] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.

[0096] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.

Claims

1. A method for intelligent vibration prediction and active control of a transmission pipeline, characterized in that, Includes the following steps: Step 1. Data Acquisition and Preprocessing: Acquire internal fluid parameters, pipe structure parameters, and external environmental parameters that affect the dynamic response of the conveying pipeline, and preprocess the time-series data of each parameter. Step 2. Dynamic response prediction based on a hybrid deep learning model: Construct a hybrid prediction model of physical information neural network PINN and long short-term memory network LSTM, train and optimize the model based on preprocessed time series data, and predict the dynamic response parameters of the conveying pipeline in future time steps; Step 3. Reinforcement learning-based adaptive active control: Set the core safety control target based on the predicted dynamic response parameters of the conveying pipeline, and construct a prediction-decision-feedback deep reinforcement learning control framework. Through real-time interaction between the deep learning agent and the conveying pipeline system, generate the optimal control command and drive the actuator to respond. Step 4. Online Model Update and Maintenance: Regularly update and maintain the parameters of the hybrid prediction model and the deep reinforcement learning agent.

2. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 1, characterized in that, The specific method for preprocessing the time-series data of each collected parameter is as follows: Abnormal data points that do not conform to the 3σ criterion are removed from the time series data of each parameter collected by each sensor. The time series data of each parameter is denoised using wavelet thresholding. Feature engineering is performed on the processed time series data of each parameter. The feature dataset for model training is constructed by extracting its time domain features and frequency domain features. The feature dataset is divided into training set, validation set and test set according to a preset ratio.

3. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 2, characterized in that, The specific method for constructing the hybrid prediction model of the Physical Information Neural Network (PINN) and the Long Short-Term Memory Network (LSTM) is as follows: A hybrid prediction model is constructed by synergistically combining the Physical Information Neural Network (PINN) and the Long Short-Term Memory (LSTM) network. The input layer receives the preprocessed feature dataset, and the input dimension is set according to the number of indicators of time-domain and frequency-domain features extracted by feature engineering. The LSTM network serves as a hidden layer, capturing the temporal dependencies of the pipeline dynamic response through its internal control mechanism. The fully connected layer is used to fuse the features output by the LSTM network. The PINN constraint layer serves as the top layer, applying physical constraints by using the control equations of the pipeline dynamics as regularization terms in the loss function. The output layer outputs the predicted values ​​of the pipeline dynamic response parameters for future time steps.

4. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 3, characterized in that, The specific method for training and optimizing the model parameters is as follows: A composite loss function is constructed by adding the mean squared error formula and the physical constraint formula to the hybrid prediction model. ,in These are the weighting coefficients for the physical constraint formula; Mean square error formula This represents the deviation between the predicted values ​​of the hybrid prediction model and the actual data, where... Indicates the first The true values ​​of the dynamic response parameters of the pipeline corresponding to each data point. Indicates the mixed prediction model for the first The predicted values ​​of the dynamic response parameters of the pipeline corresponding to each data point. , It is a positive integer greater than 2. This represents the total number of data points in the training set; Physical constraint formula This represents the deviation between the output of the hybrid prediction model and the physical equations, where The governing equations that govern the dynamics of the pipeline are represented. Indicates the hybrid prediction model in the 1st... The predicted physical field quantities and their first and second derivatives for each physical constraint point. , It is a positive integer greater than 2. This represents the total number of sample points randomly selected within the computational domain of the transport pipeline to apply physical constraints. During the training of the hybrid prediction model, the Adam optimizer is used to set the initial learning rate, and the learning rate is dynamically adjusted through the cosine annealing strategy. If the composite loss function value on the validation set does not decrease within the preset number of training iterations, the training of the hybrid prediction model is stopped, which constitutes the optimization of the model parameters.

5. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 4, characterized in that, The specific method for predicting the dynamic response parameters of the flow-carrying pipeline at future time steps is as follows: The time-series data of each parameter after preprocessing in the historical and current time steps are input into the trained hybrid prediction model. The Long Short-Term Memory (LSTM) network in the hidden layer captures the temporal dependence of the dynamic response of the conveying pipe and learns its dynamic evolution law. The PINN constraint layer of the physical information neural network ensures that each prediction step in the prediction process satisfies the constraints of the dynamic control equation of the conveying pipe. Finally, the dynamic response parameters of the conveying pipe in the future time step are directly generated through the output layer, including the peak value of vibration acceleration and the maximum strain value of the wall.

6. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 5, characterized in that, The specific method for setting the core security control target is as follows: If the peak vibration acceleration of the pipeline is less than or equal to the preset gravity acceleration threshold, and the maximum strain value of the wall meets the numerical range of the yield strain of the pipeline wall, then it is set as the core safety control target.

7. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 6, characterized in that, The specific method for constructing the prediction-decision-feedback deep reinforcement learning control framework is as follows: The dynamic response parameters of the pipeline at a future preset time step predicted by the hybrid prediction model and the current pipeline operating parameters are used together as the state space, and the adjustable parameters of the active control actuator are used as the action space. A reward function is set based on the state space and the action space. The reward function sets the output value according to whether the predicted dynamic response parameters of the pipeline at a future preset time step in the state space meet the set core safety control target. If the predicted dynamic response parameters of the pipeline at a future preset time step in the state space meet the set core safety control target, the reward function outputs a positive reward value, otherwise the reward function outputs a negative penalty value. In this way, a deep reinforcement learning control framework of prediction-decision-feedback is constructed.

8. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 7, characterized in that, The specific method for generating optimal control commands and driving the actuator to respond is as follows: The deep reinforcement learning agent outputs the corresponding control parameter values ​​in the action space based on the current state space, generates control instructions for the corresponding active control actuator based on the control parameter values, and transmits the control instructions to the active control actuator in real time through the industrial bus and completes the execution within a specified time.

9. The intelligent vibration prediction and active control method for a transmission pipeline according to claim 8, characterized in that, The specific method for periodically updating and maintaining the parameters of the hybrid prediction model and the deep reinforcement learning agent is as follows: Within a preset period, the newly collected internal fluid parameters, pipe structure parameters, and external environmental parameters affecting the dynamic response of the conveying pipe are preprocessed and added to the training set. Incremental training is performed on the hybrid prediction model and the deep reinforcement learning agent. When the prediction error of the hybrid prediction model on the test set is greater than the preset prediction error threshold or the control response delay is greater than the preset allowable delay threshold, full training of the hybrid prediction model and the deep reinforcement learning agent is triggered to update their parameters. Before triggering incremental or full training of the hybrid prediction model and deep reinforcement learning agent, the signal quality of each sensor channel is evaluated. The signal-to-noise ratio and data loss rate of each sensor are obtained from the local database. If the signal-to-noise ratio of a certain sensor channel is continuously less than the preset signal-to-noise ratio threshold or the data loss rate is greater than the preset data loss rate threshold within the target period, the sensor is determined to be faulty. When a single sensor is determined to be faulty, the real-time data of the remaining normal sensors is input into the pre-trained hybrid prediction model. The predicted values ​​of the hybrid prediction model for the fault point parameters are used as the reconstructed data to replace the real data of the faulty sensor and are input into the deep reinforcement learning agent to generate control commands. When it is determined that three or more sensors fail simultaneously or the prediction error of the hybrid prediction model increases to the preset warning threshold within the target period, the deep reinforcement learning agent control command generation function is suspended. The system directly issues a set of preset basic safety action commands to the active control actuator and triggers an audible and visual alarm to request manual intervention.

10. A system for implementing the intelligent vibration prediction and active control method for a transmission pipeline according to any one of claims 1-9, characterized in that, include: Data acquisition and preprocessing module: used to acquire internal fluid parameters, pipe structure parameters and external environmental parameters that affect the dynamic response of the conveying pipeline, and to preprocess the time series data of each acquired parameter; The dynamic response prediction module based on a hybrid deep learning model is used to construct a hybrid prediction model of the physical information neural network PINN and the long short-term memory network LSTM. The model is trained and its parameters are optimized based on preprocessed time series data, and the dynamic response parameters of the conveying pipeline at future time steps are predicted. The reinforcement learning-based adaptive active control module is used to set the core safety control objectives based on the predicted dynamic response parameters of the conveying pipeline, and to construct a prediction-decision-feedback deep reinforcement learning control framework. Through real-time interaction between the deep learning agent and the conveying pipeline system, it generates the optimal control commands and drives the actuators to respond. Online Model Update and Maintenance Module: Used to periodically update and maintain the parameters of the hybrid prediction model and the deep reinforcement learning agent.

Images