Intelligent control method for screw vacuum pump based on digital twinning

Abstract

This invention discloses an intelligent control method for screw vacuum pumps based on digital twins, belonging to the field of intelligent control. The specific steps of this intelligent control method are as follows: Ⅰ. Unified acquisition of multi-source sensor data and establishment of a time reference; Ⅱ. Construction of a state simulation model and simultaneous initialization of the model state. This invention enables control decisions to no longer rely on limited sensor data, significantly improving the completeness and accuracy of state perception. It ensures that the digital twin maintains high consistency and stability under complex operating conditions and long-term operation. It can capture minute deformations and local anomalies that are difficult to identify using traditional methods. Compared with traditional control methods that target pressure or flow, it is more fundamental and targeted, enabling complex flow analysis to meet real-time control requirements, improving the overall system response speed, breaking through the limits of traditional steady-state control, and further reducing energy loss and improving operating efficiency while ensuring stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, and in particular to an intelligent control method for screw vacuum pumps based on digital twins. Background Technology

[0002] Screw vacuum pumps, as a crucial core component for dry vacuum generation, are widely used in high-end industrial fields such as semiconductor manufacturing, pharmaceuticals, chemicals, precision machining, and new energy. They achieve continuous gas delivery and compression through one or more pairs of meshing rotors, offering advantages such as compact structure, stable pumping speed, and oil-free operation. However, in actual operation, due to the intense unsteady flow, gap leakage, thermal expansion coupling, and complex vortex structure evolution accompanying gas compression, significant energy dissipation and flow instability often occur within the pump chamber, leading to decreased system efficiency, increased temperature rise, and accelerated component wear. Traditional screw vacuum pump control methods mainly rely on fixed parameter settings or experience-based closed-loop adjustment strategies, such as constant speed operation or simple frequency modulation control based on pressure feedback. These methods typically assume steady-state conditions and lack the ability to characterize the internal flow field evolution mechanism, making them difficult to respond to transient disturbances and complex operating condition changes. Meanwhile, since the internal state of the pump chamber (such as local velocity distribution, vortex structure, gap leakage path, etc.) is difficult to observe directly, the control system generally suffers from the problem of "limited measurable variables and numerous unmeasurable states", which leads to a significant deviation between control decisions and actual energy loss. Therefore, it is particularly important to invent an intelligent control method for screw vacuum pumps based on digital twins.

[0003] Existing intelligent control methods for screw vacuum pumps rely too heavily on limited sensing data for control decisions, reducing the completeness and accuracy of state perception. They cannot guarantee that the digital twin will maintain high consistency and stability under complex operating conditions and long-term operation. Furthermore, traditional methods struggle to identify minute deformations and local anomalies, and complex flow analysis cannot meet real-time control requirements, reducing the overall system response speed. At the same time, they cannot further reduce energy loss and improve operating efficiency while ensuring stability. Therefore, we propose an intelligent control method for screw vacuum pumps based on digital twins. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a digital twin-based intelligent control method for screw vacuum pumps.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A digital twin-based intelligent control method for screw vacuum pumps, comprising the following specific steps:

[0007] Ⅰ. Unified collection of multi-source sensor data and establishment of time reference: Multiple types of sensors are deployed in various parts of the screw vacuum pump to collect various types of sensor data in a unified manner, and a global time reference is established based on the network time protocol to perform time alignment and synchronization correction on sensor data with different sampling frequencies.

[0008] II. Constructing a state simulation model and initializing the model state: After completing data alignment, construct a state simulation model of the screw vacuum pump using the processed multi-source sensor data, and reconstruct the various operating variables of the screw vacuum pump in real time.

[0009] III. Establish a virtual game in the cloud and simulate counterfactual scenarios in parallel: Based on the state simulation model, establish a corresponding virtual game in the cloud and generate multiple counterfactual operating scenarios based on the current state. When transient disturbance trends are detected in the operating parameters, simulate different control strategies and generate multiple sets of candidate control strategies.

[0010] IV. Analyze the real-time high-frequency ultrasonic signals and establish a virtual gap field: While generating candidate control strategies, deeply analyze the real-time high-frequency ultrasonic signals and invert the three-dimensional vibration morphology and instantaneous elastic deformation of the rotor surface. Then, map the inversion results into a spatially continuous virtual gap field.

[0011] V. Dynamic gap adjustment control based on the established virtual gap field: Based on the virtual gap field, the continuous state change is converted into a pulse frequency encoded signal by simulating the neuron pulse firing mechanism. Subsequently, the rotor position is adjusted according to the generated pulse frequency encoded signal.

[0012] VI. Online identification of pump cavity entropy production rate and analysis of flow field structure: After the gap control is stabilized, the entropy production rate distribution inside the screw vacuum pump is calculated in real time. By constructing the entropy production rate field and spatial gradient distribution, the region where energy loss is concentrated is identified, and the unstable vortices or unfavorable flow modes existing in the pump cavity are analyzed.

[0013] VII. Dynamically adjust the control parameters of the screw vacuum pump and verify them through a state simulation model: Based on the obtained entropy yield distribution, dynamically adjust the control strategy of the screw vacuum pump, and feed the optimized control strategy back to the state simulation model for verification. Simultaneously, apply it to the corresponding screw vacuum pump, and then collect new data again through various sensors to form new data input.

[0014] As a further aspect of the present invention, the various types of sensing data mentioned in step I specifically include temperature, pressure, vibration, rotational speed, and ultrasound.

[0015] As a further aspect of the present invention, the specific steps of step II, which involve constructing a state simulation model of the screw vacuum pump using the processed multi-source sensor data and reconstructing the various operating variables of the screw vacuum pump in real time, are as follows:

[0016] S1.1: After completing the time alignment of various sensor data, the scale of each sensor data is uniformly processed. The sampled values ​​from different sensor channels are weighted and fused according to a preset time base to obtain a unified observation vector at different times.

[0017] S1.2: Standardize and smooth each observation component in the obtained unified observation vector to generate stable input data with consistent scale. Then, based on the processed input data, perform parametric modeling of the gap field and thermal field of the screw vacuum pump under various operating states, and construct the state vector of the state simulation model of the screw vacuum pump based on the modeling results.

[0018] S1.3: Based on the structural parameters, dynamic relationships and thermal deformation relationships of the screw vacuum pump, construct the corresponding state evolution relationship, and then set the equivalent thermal expansion correction relationship to directly introduce the thermal state change into the gap field mapping in order to establish a state simulation model that includes the thermal-deformation-gap coupling relationship.

[0019] S1.4: The state of the screw vacuum pump is predicted by the state simulation model, and the prediction error covariance matrix corresponding to the current state prediction is constructed. Then, the Kalman gain is calculated based on the current state prediction, the actual state and the prediction error covariance matrix, and the state simulation model is corrected.

[0020] S1.5: Identify each parameter in the state simulation model based on historical samples and current residuals. After parameter identification is completed, obtain the state set of the state simulation model at the current time and calculate the residual of the current state simulation model. If the residual exceeds the preset threshold, it indicates that there is a deviation in the current twin model, and the state estimation or parameter identification process is re-executed.

[0021] As a further aspect of the present invention, the specific form of the state simulation model described in S1.3 is as follows:

[0022]

[0023] In the formula, Representative moment Spatial coordinates The distribution of gaps at the location; Represents the nominal gap curve; Representing the One gap modal coefficient; Represents the basis functions of the space; Represents the number of gap modes; Representative moment In spatial coordinates Temperature distribution at that location; Represents the reference temperature field; Representing the One thermal modal coefficient; Represents the thermal field basis functions; This represents the number of thermal modes.

[0024] As a further aspect of the present invention, the state set described in S1.5 includes rotor displacement, rotational speed, gap field, thermal field, and updated parameter information.

[0025] As a further aspect of the present invention, step III involves establishing a corresponding virtual game entity in the cloud based on a state simulation model, generating multiple counterfactual operating scenarios based on the current state, and performing the following specific steps to deduce different control strategies when a transient disturbance trend is detected in the operating parameters:

[0026] S2.1: Extract the original state features at the current moment from the latest running state output by the state simulation model, and then map the original initial features to the game initial state through the state coding operator. Then, based on the current game initial state and based on various possibilities, construct different virtual initial conditions to establish corresponding counterfactual scenarios, and define a disturbance intensity score for each scenario.

[0027] S2.2: Based on the current operating state and the previous operating state, calculate the current state jump variable. If the state jump variable is greater than the preset disturbance trigger threshold, the screw vacuum pump is determined to have entered the transient disturbance stage and immediately enters the virtual simulation. At the same time, establish corresponding reward functions for each counterfactual scenario to obtain the reward value of each simulation step in real time. Under multiple scenario conditions, calculate the robust objective value of each strategy through the robust objective function.

[0028] S2.3: When entering the transient disturbance stage, the virtual initial conditions corresponding to each counterfactual scenario are taken as the root node, and a tree search is performed based on the upper confidence bound principle. The next branch is selected step by step, and each state-action combination in the search process is represented by a node. The search is stopped when the preset maximum search time is reached. Multiple rolling simulations are performed based on the branches generated in this round of search, and the cumulative benefit of the action in different counterfactual scenarios is calculated. Then, the simulation results are fed back to the tree node value.

[0029] S2.4: Based on the highest cumulative reward in the current tree search process, update the policy parameters using the policy gradient form, repeat the tree search and policy update until the change value of the policy parameters converges to the preset range, stop the tree search and policy update, and output the candidate control policies for each counterfactual scenario;

[0030] S2.5: Calculate the worst value of each candidate policy in each counterfactual scenario, compare the worst values ​​of each candidate policy, select the candidate policy with the highest worst value as the final control policy output, and retain the second-best candidate policy to enter the next round of policy update.

[0031] As a further aspect of the present invention, the specific calculation formula for the return function in S2.2 is as follows:

[0032]

[0033] In the formula, Representing the In the scenario, the first The reward value for each simulation step; Representing the In the scenario, the first Efficiency benefit term for each simulation step; Representing the In the scenario, the first An unstable penalty term for each deduction step; Representing the In the scenario, the first Constraint violations in each deduction step; , , These represent the weights corresponding to efficiency gain items, instability penalty items, and constraint violation items, respectively.

[0034] The specific calculation formula for the robust objective function described in S2.2 is as follows:

[0035]

[0036] In the formula, Representative strategy The robust target value; Represents expectations; Represents the number of steps in the deduction process; Represents the discount factor;

[0037] The specific manifestation of the upper confidence boundary principle described in S2.3 is as follows:

[0038]

[0039] In the formula, This represents the best candidate action selected at the current node; Represents a set of actions; Representative actions The average value; Represents the exploration coefficient; This represents the total number of visits to the current parent node; Representative actions Number of visits;

[0040] The specific calculation formula for the update strategy parameters described in S2.4 is as follows:

[0041]

[0042] In the formula, Representing the Strategy parameters at the time of the next update; Represents the learning rate; Represents the state Select action The probability of; This represents the current optimal rolling return; Represents the baseline term.

[0043] As a further aspect of the present invention, the various possibilities mentioned in S2.1 include sudden temperature rise, load fluctuation, gas property shift, sudden increase in mechanical friction, etc.

[0044] As a further aspect of the present invention, the specific steps of mapping the inversion result to a spatially continuous virtual gap field in step IV are as follows:

[0045] S3.1: The original echo signal is acquired in real time by an ultrasonic sensor array arranged at the position of the screw vacuum pump. The continuous echo signal is discretized according to a fixed sampling period and then divided into multiple short-time analysis frames. Then, the echo signal of each frame is bandpass filtered and the echo signal of each frame is envelope demodulated to obtain the corresponding envelope signal.

[0046] S3.2: The short-time Fourier transform is used to project the acquired envelope signals of each frame onto the time-frequency plane, and the energy distribution of different frequency bands changes with time is detected. At the same time, the local maximum trajectory is extracted based on the detection results to form the main ridge feature sequence. Then, the envelope signal, main ridge feature and energy distribution voiceprint features are spliced ​​into a unified observation vector in the order within the same frame and scale normalized.

[0047] S3.3: Based on the existing vibration modes of each rotor surface, a set of candidate modal basis functions are pre-constructed, and the constructed modal basis functions are stacked into a dictionary matrix. Then, based on the constructed dictionary matrix, a linear relationship between the observation vector and the modal coefficients is established. After that, a corresponding sparse prior is added to each modal coefficient, and each coefficient has an independent precision parameter. Finally, the posterior distribution of each modal coefficient is calculated in combination with the actual observation of the current frame.

[0048] S3.4: After obtaining each posterior distribution, update the precision parameters of each modal coefficient through the evidence maximization rule. Then, according to the decision threshold, remove modes with precision parameters lower than the decision threshold, establish an active mode set and output the effective modal coefficients. Then, based on the effective modal coefficients and through the corresponding modal basis functions, reconstruct the vibration displacement field of the rotor surface in polar coordinates.

[0049] S3.5: Based on the established vibration displacement field, the vibration displacement is converted into instantaneous elastic deformation. Then, according to the reconstructed vibration displacement and elastic deformation, the corresponding local gap value is calculated on the corresponding discrete node. Subsequently, radial basis interpolation is used to extend the discrete node into a continuous field to generate a complete virtual gap field.

[0050] As a further aspect of the present invention, the specific steps of step VI, namely, calculating the entropy yield distribution inside the screw vacuum pump in real time, identifying regions of concentrated energy loss by constructing an entropy yield field and spatial gradient distribution, and analyzing unstable vortices or unfavorable flow modes existing in the pump cavity, are as follows:

[0051] S4.1: After the gap control enters a steady state, multiple flow field snapshots are extracted from the pump chamber computation domain of the screw vacuum pump. The velocity, pressure and temperature distributions in each flow field snapshot are organized into a unified sample. Then, all snapshots are combined into a sample matrix. Singular value decomposition is performed on the sample matrix. The number of main modes retained is determined by the energy retention rate, and each retained main mode is used as a reduced-order basis.

[0052] S4.2: The control equations for continuous flow within the pump chamber of a screw vacuum pump are discretized onto grid nodes to form a computable full-dimensional state. Then, based on the reduced valence base, the order reduction coefficients corresponding to each reduced valence base, and the mean vector of the full-dimensional state, a full-dimensional state equation is established. The reduced valence base is then subjected to a weighted orthogonal projection to obtain the corresponding state evolution equation.

[0053] S4.3: The implicit time-progression scheme is used to iterate the state evolution equation quickly, and after obtaining the latest order reduction coefficient, it is remapped back to the full-dimensional space to obtain the solved three-dimensional flow field. Then, the entropy productivity of each spatial unit in the three-dimensional flow field of the pump cavity is calculated, and the entropy productivity values ​​of all spatial units are statistically analyzed to establish a complete entropy productivity field.

[0054] S4.4: Calculate the spatial gradient of each spatial unit based on the established entropy yield field, and set the gradient magnitude. After obtaining the entropy yield field and its gradient, calculate the current high entropy threshold value through the mean entropy yield, the mean entropy yield, and the corresponding threshold coefficient, and mark the region where the entropy yield is greater than the high entropy threshold value as the high entropy loss region.

[0055] S4.5: Collect the velocity vector of each spatial unit and calculate the vortex index of the velocity field of each spatial unit. At the same time, construct rotation and strain criteria. When the vortex index is higher than the preset vortex threshold and the rotation and strain criteria are greater than 0, the corresponding region is determined to be an unstable vortex candidate region.

[0056] As a further aspect of the present invention, the specific steps of step VII, which involve dynamically adjusting the screw vacuum pump control strategy based on the acquired entropy yield distribution and simultaneously feeding the optimized control strategy back to the state simulation model for verification, are as follows:

[0057] S5.1: Arrange the entropy yield values ​​of each discrete unit in the pump chamber of the screw vacuum pump into candidate paths according to the spatial connectivity relationship, then calculate the cumulative cost of each candidate path, and then set the operating status of the screw vacuum pump at each control time based on historical data, and establish a recursive nonlinear prediction model, and set the rolling optimization objective function corresponding to the nonlinear prediction model.

[0058] S5.2: Obtain the inverter carrier frequency command and synchronous gear phase difference command for each control moment in the current control strategy, input them into the nonlinear prediction model to predict the operating state of the screw vacuum pump, calculate the optimization target value under the current control quantity through the rolling optimization objective function, adjust the control quantity, and re-predict the state through the nonlinear prediction model.

[0059] S5.3: Repeat the control quantity adjustment iteration until the optimization target value decreases to within the preset range, then stop the iteration, take the control quantity corresponding to the minimum optimization target value as the optimal control quantity at the current control moment, gradually update the control quantity at each control moment, and output the updated optimal control strategy;

[0060] S5.4: Based on the pulsation amplitude, pulsation frequency, pulsation initial phase, and control sampling period, calculate the frequency pulsation amount for the corresponding control time. Then, superimpose the corresponding frequency pulsation amount onto the inverter carrier frequency command of the optimal control quantity for the corresponding control time in the optimal control strategy, and generate the final inverter carrier frequency command. At the same time, calculate the rotor's inertial response based on the final inverter carrier frequency command, and then calculate the vortex repeatability intensity in each control cycle in real time. If the vortex repeatability intensity decreases compared to the previous control moment, it indicates that the current control quantity can suppress the regeneration of periodic vortices; otherwise, readjust the control quantity for the current control cycle.

[0061] S5.5: Based on the optimal control strategy after pulsation correction, the rolling time domain method is used to output only the first action at the current moment, and the screw vacuum pump operating state is recalculated in the next cycle. Then, the rolling optimization objective is solved again using the new feedback state, and the current control strategy is adjusted.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] This intelligent control method for screw vacuum pumps based on digital twins performs time alignment of multi-source sensor data, scale unification and weighted fusion of data from different channels to construct a unified observation vector. After standardization and smoothing, a stable input is formed. Based on this input, the gap field and thermal field of the screw vacuum pump are parametrically modeled, establishing a state simulation model that includes structural parameters, dynamic relationships, and thermal deformation coupling. Dynamic model correction and parameter identification are achieved through state prediction and Kalman correction. When the model residual exceeds the limit, reestimation is triggered. Subsequently, state features are extracted and mapped to the initial state of the game, constructing multiple counterfactual scenarios. When transient disturbances are detected, virtual deduction based on Monte Carlo tree search and policy gradient is triggered, outputting a robust optimal control strategy. Simultaneously, acoustic features are extracted using ultrasonic signals, and rotor vibration and the virtual gap field are reconstructed through sparse Bayesian inversion. After the gap stabilizes, a reduced-order flow... The field model rapidly solves three-dimensional flow and calculates the entropy production rate field, identifies high-loss regions and unstable vortex structures, and then constructs a nonlinear model predictive control framework with the minimum entropy production path as the objective. It jointly optimizes the inverter frequency and gear phase, and superimposes micro-amplitude speed pulsations to suppress periodic vortices. Finally, it continuously updates the control strategy through rolling optimization, which enables control decisions to no longer rely on limited sensing quantities, significantly improving the completeness and accuracy of state perception. It ensures that the digital twin maintains high consistency and stability under complex working conditions and long-term operation, and can capture small deformations and local anomalies that are difficult to identify by traditional methods. Compared with traditional control methods that target pressure or flow, it is more essential and targeted, enabling complex flow analysis to meet real-time control requirements, improving the overall system response speed, breaking through the limits of traditional steady-state control, and further reducing energy loss and improving operating efficiency while ensuring stability. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0065] Figure 1 This is a flowchart of the intelligent control method for screw vacuum pumps based on digital twins proposed in this invention. Detailed Implementation

[0066] Example 1, referring to Figure 1 A digital twin-based intelligent control method for screw vacuum pumps is described, with the following specific steps:

[0067] Unified collection of multi-source sensor data and establishment of time reference: Multiple types of sensors are deployed in various parts of the screw vacuum pump to collect various types of sensor data in a unified manner, and a global time reference is established based on the network time protocol to perform time alignment and synchronization correction on sensor data with different sampling frequencies.

[0068] Construct a state simulation model and initialize the model state: After data alignment is completed, construct a state simulation model of the screw vacuum pump using the processed multi-source sensor data, and reconstruct the various operating variables of the screw vacuum pump in real time.

[0069] Specifically, after time alignment of various sensor data, the data is scaled uniformly. Sampled values ​​from different sensor channels are weighted and fused according to a preset time base to obtain a unified observation vector at different times. The observed components in the obtained unified observation vector are standardized and smoothed to generate stable input data with consistent scale. Then, based on the processed input data, parametric modeling of the gap field and thermal field of the screw vacuum pump is performed for each operating state. Based on the modeling results, a state vector for the state simulation model of the screw vacuum pump is constructed. According to the structural parameters, dynamic relationships, and thermal deformation relationships of the screw vacuum pump, corresponding state evolution relationships are constructed. Finally, an equivalent thermal expansion correction relationship is set to adjust the thermal state... The state change is directly introduced into the gap field mapping to establish a state simulation model that includes the thermal-deformation-gap coupling relationship. The state simulation model is used to predict the state of the screw vacuum pump and construct the prediction error covariance matrix corresponding to the current state prediction. Then, the Kalman gain is calculated based on the current state prediction, the actual state, and the prediction error covariance matrix, and the state simulation model is corrected. Based on historical samples and the current residual, the parameters in the state simulation model are identified. After the parameter identification is completed, the state set of the state simulation model at the current time is obtained, and the residual of the current state simulation model is calculated. If the residual exceeds the preset threshold, it indicates that there is a deviation in the current twin model, and the state estimation or parameter identification process is re-executed.

[0070] A virtual game entity is established in the cloud, and counterfactual scenarios are simulated in parallel: Based on the state simulation model, a corresponding virtual game entity is established in the cloud, and multiple counterfactual operating scenarios are generated based on the current state. When transient disturbance trends are detected in the operating parameters, different control strategies are simulated to generate multiple sets of candidate control strategies.

[0071] Specifically, the original state features at the current moment are extracted from the latest operating state output by the state simulation model. Then, the original initial features are mapped to the initial state of the game using a state encoding operator. Based on the current initial state and various possibilities, different virtual initial conditions are constructed to establish corresponding counterfactual scenarios. A disturbance intensity score is defined for each scenario. Based on the current operating state and the previous operating state, the current state jump variable is calculated. If the state jump variable > a preset disturbance trigger threshold, the screw vacuum pump is determined to have entered the transient disturbance stage and immediately enters virtual simulation. Simultaneously, a corresponding reward function is established for each counterfactual scenario to obtain the reward value of each simulation step in real time. Under multiple scenario conditions, the robust objective value of each strategy is calculated using a robust objective function. When entering the transient disturbance stage, the virtual initial conditions corresponding to each counterfactual scenario are used as root nodes, and based on the upper confidence level... The tree search is performed according to the bounded principle, progressively selecting the next branch. Each state-action combination in the search process is represented as a node until the preset maximum search time is reached. The search then stops, and multiple rolling simulations are performed based on the branches generated in this round of search. The cumulative benefit of the action in different counterfactual scenarios is calculated. Subsequently, the simulation results are fed back to the tree node value. Based on the highest cumulative benefit in the current tree search process, the policy parameters are updated using the policy gradient. The tree search and policy update are repeated until the change value of the policy parameters converges to a preset range. The tree search and policy update are then stopped, and candidate control policies for each counterfactual scenario are output. The worst value of each candidate policy in each counterfactual scenario is calculated. The worst values ​​of each candidate policy are compared, and the candidate policy with the highest worst value is selected as the final control policy output. The second-best candidate is retained, and the next round of policy update begins.

[0072] Example 2, refer to Figure 1 A digital twin-based intelligent control method for screw vacuum pumps is described, with the following specific steps:

[0073] Analyze the high-frequency ultrasonic signals acquired in real time and establish a virtual gap field: While generating candidate control strategies, deeply analyze the high-frequency ultrasonic signals acquired in real time, and invert the three-dimensional vibration morphology and instantaneous elastic deformation of the rotor surface. Then, map the inversion results into a spatially continuous virtual gap field.

[0074] Specifically, an ultrasonic sensor array positioned at the location of the screw vacuum pump acquires raw echo signals in real time. The continuous echo signals are discretized according to a fixed sampling period and then divided into multiple short-time analysis frames. Each frame's echo signal is then bandpass filtered, and envelope demodulation is performed on the filtered echo signals to obtain the corresponding envelope signals. A short-time Fourier transform is used to project the acquired envelope signals of each frame onto the time-frequency plane, and the energy distribution of different frequency bands over time is detected. Simultaneously, local maximum trajectories are extracted based on the detection results to form a main ridge feature sequence. The envelope signals, main ridge features, and acoustic signature features of the energy distribution are then concatenated in the same frame to form a unified observation vector, which is then scaled. Based on the existing vibration modes of each rotor surface, a set of candidate modal basis functions is pre-constructed, and these constructed modal basis functions are stacked column-wise into a dictionary matrix. Finally, based on the constructed... The dictionary matrix is ​​used to establish a linear relationship between the observation vector and the modal coefficients. Then, a corresponding sparse prior is added to each modal coefficient, and each coefficient has an independent precision parameter. The posterior distribution of each modal coefficient is calculated in combination with the actual observation of the current frame. After obtaining each posterior distribution, the precision parameter of each modal coefficient is updated by the evidence maximization rule. Then, according to the decision threshold, modes with precision parameters lower than the decision threshold are removed, an active mode set is established, and the effective modal coefficients are output. Then, based on the effective modal coefficients and the corresponding modal basis functions, the vibration displacement field of the rotor surface in polar coordinates is reconstructed. Based on the established vibration displacement field, the vibration displacement is converted into instantaneous elastic deformation. Then, according to the reconstructed vibration displacement and elastic deformation, the corresponding local gap value is calculated on the corresponding discrete node. Finally, radial basis interpolation is used to extend the discrete node into a continuous field to generate a complete virtual gap field.

[0075] Furthermore, it should be noted that the specific implementation process of the evidence maximization rule is as follows: Under the current accuracy parameter value, the posterior distribution of each modality coefficient is obtained by combining the observation data. The posterior distribution includes the posterior mean and posterior variance. Then, the accuracy parameter of each modality is recalculated using the posterior mean and posterior variance. It is then checked whether the change between the old and new accuracy parameters meets the preset threshold. If not, the accuracy parameter of each modality is recalculated. After repeated iterations, the accuracy parameters of the modalities with strong observation and explanatory power are retained. The specific calculation formula is as follows:

[0076]

[0077] In the formula, The updated precision parameters; The total number of frames participating in the joint estimation; For the first Frame posterior mean The One component; The posterior covariance matrix is ​​the first... One diagonal element; The value is positive to prevent the denominator from being zero; then, the active mode set is filtered based on the threshold.

[0078]

[0079] And output the effective modal coefficients:

[0080]

[0081] In the formula, A set of active modality indices; To determine the threshold; These are the effective modal coefficients of the final output; To retain only the components corresponding to the active index.

[0082] Based on the established virtual gap field, dynamic gap adjustment control is performed: based on the virtual gap field, the continuous state change is converted into a pulse frequency encoded signal by simulating the neuron pulse firing mechanism. Subsequently, the rotor position is scaled based on the generated pulse frequency encoded signal.

[0083] Further explanation is needed. The difference between the gap value of each spatial node in the virtual gap field and the corresponding reference gap is calculated to obtain the gap deviation distribution. Then, spatial filtering and temporal smoothing are performed on each gap deviation to extract each deviation region that reflects the rotor's tendency to rub or deviate. Then, each deviation region is compressed into a set of state input vectors arranged by position. Subsequently, the state input vector is sent to the input layer of the spiking neural network. The input layer maps each gap deviation to the membrane potential accumulation of the corresponding neuron. The larger the deviation, the faster the membrane potential rises; the smaller the deviation, the slower the membrane potential rises. When the membrane potential accumulation of any neuron reaches a preset threshold, a pulse is generated, and the membrane potential of the corresponding neuron is reset or partially attenuated after the pulse is emitted.

[0084] In addition, it should be noted that if the gap deviation at any position continues to increase, it indicates that the risk is rising, and the corresponding neuron will generate pulses more frequently; if the gap deviation tends to stabilize or decrease, the pulse firing frequency will decrease.

[0085] Online identification of pump cavity entropy production rate and analysis of flow field structure: After the gap control stabilizes, the entropy production rate distribution inside the screw vacuum pump is calculated in real time. By constructing the entropy production rate field and spatial gradient distribution, the region where energy loss is concentrated is identified, and the unstable vortices or unfavorable flow modes existing in the pump cavity are analyzed.

[0086] Specifically, after the gap control enters a steady state, multiple flow field snapshots at various times are extracted from the computational domain of the screw vacuum pump's pump chamber. The velocity, pressure, and temperature distributions in each snapshot are then organized into a unified sample. All snapshots are then combined into a sample matrix, and singular value decomposition is performed on the matrix. The number of retained principal modes is determined by the energy retention rate, and each retained principal mode serves as a reduced-order basis. The continuous flow control equations within the screw vacuum pump's pump chamber are discretized onto grid nodes to form a computable full-dimensional state. Then, a full-dimensional state equation is established based on the reduced-order basis, the corresponding reduction coefficients of each reduced-order basis, and the mean vector of the full-dimensional state. A weighted orthogonal projection is performed on the reduced-order basis to obtain the corresponding state evolution equation. An implicit time-progression scheme is used to rapidly iterate the state evolution equation, and the result is obtained at the optimal time. After the new order reduction coefficients are applied, they are remapped back to the full-dimensional space to obtain the solved three-dimensional flow field. Then, the entropy productivity of each spatial unit in the three-dimensional flow field of the pump cavity is calculated, and the entropy productivity values ​​of all spatial units are statistically analyzed to establish a complete entropy productivity field. Based on the established entropy productivity field, the spatial gradient of each spatial unit is calculated, and the gradient magnitude is set. After obtaining the entropy productivity field and its gradient, the current high entropy threshold value is calculated using the mean entropy productivity, the mean entropy productivity, and the corresponding threshold coefficient. Regions with entropy productivity > high entropy threshold value are marked as high entropy loss regions. The velocity vectors of each spatial unit are collected, and the vortex index of the velocity field of each spatial unit is calculated. At the same time, rotation and strain criteria are constructed. When the vortex index is higher than the preset vorticity threshold, and the rotation and strain criteria are greater than 0, the corresponding region is determined to be an unstable vortex candidate region.

[0087] The control parameters of the screw vacuum pump are dynamically adjusted and verified through a state simulation model: Based on the obtained entropy yield distribution, the control strategy of the screw vacuum pump is dynamically adjusted, and the optimized control strategy is fed back to the state simulation model for verification and applied to the corresponding screw vacuum pump. Then, new data input is formed by collecting data through various sensors.

[0088] Specifically, the entropy productivity values ​​of each discrete unit within the pump chamber of the screw vacuum pump are arranged into candidate paths according to their spatial connectivity. The cumulative cost of each candidate path is then calculated. Based on historical data, the operating state of the screw vacuum pump at each control time is set, and a recursive nonlinear prediction model is established. A rolling optimization objective function corresponding to this nonlinear prediction model is set. The inverter carrier frequency command and synchronous gear phase difference command at each control time in the current control strategy are obtained and input into the nonlinear prediction model to predict the operating state of the screw vacuum pump. The optimized target value under the current control quantity is calculated using the rolling optimization objective function, and the control quantity is adjusted. The state prediction is then performed again using the nonlinear prediction model. This iterative adjustment of the control quantity is repeated until the decrease in the optimized target value converges to a preset range. The iteration stops then, and the control quantity corresponding to the minimum optimized target value is taken as the optimal control quantity at the current control time. This process is repeated to update the control state at each control time. The system calculates the frequency pulsation amount for the corresponding control time based on the pulsation amplitude, pulsation frequency, pulsation initial phase, and control sampling period. Then, it superimposes the corresponding frequency pulsation amount onto the inverter carrier frequency command of the optimal control amount for the corresponding control time in the optimal control strategy, and generates the final inverter carrier frequency command. Simultaneously, it calculates the rotor's inertial response based on the final inverter carrier frequency command, and then calculates the vortex repeatability intensity in each control cycle in real time. If the vortex repeatability intensity decreases compared to the previous control moment, it indicates that the current control amount can suppress the regeneration of periodic vortices; otherwise, it readjusts the control amount for the current control cycle. Based on the optimal control strategy after pulsation correction, it uses a rolling time domain method to output only the first action at the current moment, and recalculates the screw vacuum pump operating state in the next cycle. Subsequently, it uses the new feedback state to iteratively solve the rolling optimization objective and adjust the current control strategy.

[0089] Furthermore, it should be noted that, based on the current operating status of the screw vacuum pump, rotor angular velocity, pump chamber pressure, temperature status, clearance sensitivity status, and vortex intensity are collected and organized into a state vector. Then, based on the current inverter carrier frequency command and synchronous gear phase difference command, a corresponding control vector is established. Subsequently, a nonlinear prediction model is generated offline using historical samples, and a corresponding rolling optimization objective function is set, the specific form of which is as follows:

[0090]

[0091] In the formula, This is the state vector for the next time step; This is the state transition matrix; The input matrix; This is the current state vector; This is the current control vector; For disturbance terms; It is a nonlinear mapping function;

[0092] The specific calculation formula for the rolling optimization objective function is as follows:

[0093]

[0094] In the formula, Represents the current time window; Representing the The cost of the candidate entropy yield path corresponding to each step; Represents the amount of control change; Represents the desired state; Represents constraint and penalty items; These represent the weighting coefficients.

[0095] It should be further explained that the specific implementation method for calculating the rotor's inertial response based on the final inverter carrier frequency command is as follows:

[0096] The final inverter carrier frequency command after pulsation correction is collected. If the motor drives the rotor through the number of pole pairs, the corresponding equivalent synchronous angular velocity is obtained according to the current final inverter carrier frequency command. The actual response is then described by the rotor dynamics equation. After that, the established rotor dynamics equation is discretized, and the reference angular velocity of the rotor in response to the frequency command change, i.e. the inertial requirement, is used as the inertial response quantity.

[0097] Furthermore, the faster the final frequency command changes, the greater the angular acceleration the rotor needs to keep up with, resulting in a larger inertial response; conversely, when the frequency command changes gradually, the inertial response is smaller. This can be used to determine whether the current pulsation correction will cause the rotor to enter an excessively strong dynamic following state.

Claims

1. A smart control method for screw vacuum pumps based on digital twins, characterized in that, The specific steps of this intelligent control method are as follows: Ⅰ. Unified collection of multi-source sensor data and establishment of time reference: Multiple types of sensors are deployed in various parts of the screw vacuum pump to collect various types of sensor data in a unified manner, and a global time reference is established based on the network time protocol to perform time alignment and synchronization correction on sensor data with different sampling frequencies. II. Constructing a state simulation model and initializing the model state: After completing data alignment, construct a state simulation model of the screw vacuum pump using the processed multi-source sensor data, and reconstruct the various operating variables of the screw vacuum pump in real time. III. Establish a virtual game in the cloud and simulate counterfactual scenarios in parallel: Based on the state simulation model, establish a corresponding virtual game in the cloud and generate multiple counterfactual operating scenarios based on the current state. When transient disturbance trends are detected in the operating parameters, simulate different control strategies and generate multiple sets of candidate control strategies. IV. Analyze the real-time high-frequency ultrasonic signals and establish a virtual gap field: While generating candidate control strategies, deeply analyze the real-time high-frequency ultrasonic signals and invert the three-dimensional vibration morphology and instantaneous elastic deformation of the rotor surface. Then, map the inversion results into a spatially continuous virtual gap field. V. Dynamic gap adjustment control based on the established virtual gap field: Based on the virtual gap field, the continuous state change is converted into a pulse frequency encoded signal by simulating the neuron pulse firing mechanism. Subsequently, the rotor position is adjusted according to the generated pulse frequency encoded signal. VI. Online identification of pump cavity entropy production rate and analysis of flow field structure: After the gap control is stabilized, the entropy production rate distribution inside the screw vacuum pump is calculated in real time. By constructing the entropy production rate field and spatial gradient distribution, the region where energy loss is concentrated is identified, and the unstable vortices or unfavorable flow modes existing in the pump cavity are analyzed. VII. Dynamically adjust the control parameters of the screw vacuum pump and verify them through a state simulation model: Based on the obtained entropy yield distribution, dynamically adjust the control strategy of the screw vacuum pump, and feed the optimized control strategy back to the state simulation model for verification. Simultaneously, apply it to the corresponding screw vacuum pump, and then collect new data again through various sensors to form new data input.

2. The intelligent control method for a screw vacuum pump based on digital twins according to claim 1, characterized in that, The specific steps for constructing a state simulation model of the screw vacuum pump using the processed multi-source sensor data in step II, and for reconstructing the various operating variables of the screw vacuum pump in real time, are as follows: S1.1: After completing the time alignment of various sensor data, the sensor data is uniformly processed by weighting and fusing the sampled values ​​from different sensor channels according to a preset time base to obtain a unified observation vector at different times. S1.2: Standardize and smooth each observation component in the obtained unified observation vector to generate stable input data with consistent scale. Then, based on the processed input data, perform parametric modeling of the gap field and thermal field of the screw vacuum pump under various operating states, and construct the state vector of the state simulation model of the screw vacuum pump based on the modeling results. S1.3: Based on the structural parameters, dynamic relationships and thermal deformation relationships of the screw vacuum pump, construct the corresponding state evolution relationship, and then set the equivalent thermal expansion correction relationship to directly introduce the thermal state change into the gap field mapping in order to establish a state simulation model that includes the thermal-deformation-gap coupling relationship. S1.4: The state of the screw vacuum pump is predicted by the state simulation model, and the prediction error covariance matrix corresponding to the current state prediction is constructed. Then, the Kalman gain is calculated based on the current state prediction, the actual state and the prediction error covariance matrix, and the state simulation model is corrected. S1.5: Identify each parameter in the state simulation model based on historical samples and current residuals. After parameter identification is completed, obtain the state set of the state simulation model at the current time and calculate the residual of the current state simulation model. If the residual exceeds the preset threshold, it indicates that there is a deviation in the current twin model, and the state estimation or parameter identification process is re-executed.

3. The intelligent control method for a screw vacuum pump based on digital twins according to claim 2, characterized in that, Step III involves establishing a corresponding virtual game entity in the cloud based on the state simulation model, generating multiple counterfactual operating scenarios based on the current state, and performing the following specific steps to deduce different control strategies when a transient disturbance trend is detected in the operating parameters: S2.1: Extract the original state features at the current moment from the latest running state output by the state simulation model, and then map the original initial features to the game initial state through the state coding operator. Then, based on the current game initial state and based on various possibilities, construct different virtual initial conditions to establish corresponding counterfactual scenarios, and define a disturbance intensity score for each scenario. S2.2: Based on the current operating state and the previous operating state, calculate the current state jump variable. If the state jump variable is greater than the preset disturbance trigger threshold, the screw vacuum pump is determined to have entered the transient disturbance stage and immediately enters the virtual simulation. At the same time, establish corresponding reward functions for each counterfactual scenario to obtain the reward value of each simulation step in real time. Under multiple scenario conditions, calculate the robust objective value of each strategy through the robust objective function. S2.3: When entering the transient disturbance stage, the virtual initial conditions corresponding to each counterfactual scenario are taken as the root node, and a tree search is performed based on the upper confidence bound principle. The next branch is selected step by step, and each state-action combination in the search process is represented by a node. The search is stopped when the preset maximum search time is reached. Multiple rolling simulations are performed based on the branches generated in this round of search, and the cumulative benefit of the action in different counterfactual scenarios is calculated. Then, the simulation results are fed back to the tree node value. S2.4: Based on the highest cumulative reward in the current tree search process, update the policy parameters using the policy gradient form, repeat the tree search and policy update until the change value of the policy parameters converges to the preset range, stop the tree search and policy update, and output the candidate control policies for each counterfactual scenario; S2.5: Calculate the worst value of each candidate policy in each counterfactual scenario, compare the worst values ​​of each candidate policy, select the candidate policy with the highest worst value as the final control policy output, and retain the second-best candidate policy to enter the next round of policy update.

4. The intelligent control method for a screw vacuum pump based on digital twins according to claim 1, characterized in that, The specific steps for mapping the inversion results to a spatially continuous virtual gap field, as described in step IV, are as follows: S3.1: The original echo signal is acquired in real time by an ultrasonic sensor array arranged at the position of the screw vacuum pump. The continuous echo signal is discretized according to a fixed sampling period and then divided into multiple short-time analysis frames. Then, the echo signal of each frame is bandpass filtered and the echo signal of each frame is envelope demodulated to obtain the corresponding envelope signal. S3.2: The short-time Fourier transform is used to project the acquired envelope signals of each frame onto the time-frequency plane, and the energy distribution of different frequency bands changes with time is detected. At the same time, the local maximum trajectory is extracted based on the detection results to form the main ridge feature sequence. Then, the envelope signal, main ridge feature and energy distribution voiceprint features are spliced ​​into a unified observation vector in the order within the same frame and scale normalized. S3.3: Based on the existing vibration modes of each rotor surface, a set of candidate modal basis functions are pre-constructed, and the constructed modal basis functions are stacked into a dictionary matrix. Then, based on the constructed dictionary matrix, a linear relationship between the observation vector and the modal coefficients is established. After that, a corresponding sparse prior is added to each modal coefficient, and each coefficient has an independent precision parameter. Finally, the posterior distribution of each modal coefficient is calculated in combination with the actual observation of the current frame. S3.4: After obtaining each posterior distribution, update the precision parameters of each modal coefficient through the evidence maximization rule. Then, according to the decision threshold, remove modes with precision parameters lower than the decision threshold, establish an active mode set and output the effective modal coefficients. Then, based on the effective modal coefficients and through the corresponding modal basis functions, reconstruct the vibration displacement field of the rotor surface in polar coordinates. S3.5: Based on the established vibration displacement field, the vibration displacement is converted into instantaneous elastic deformation. Then, according to the reconstructed vibration displacement and elastic deformation, the corresponding local gap value is calculated on the corresponding discrete node. Subsequently, radial basis interpolation is used to extend the discrete node into a continuous field to generate a complete virtual gap field.

5. The intelligent control method for a screw vacuum pump based on digital twins according to claim 1, characterized in that, The specific steps for step VI, namely, to calculate the entropy yield distribution inside the screw vacuum pump in real time, identify regions of concentrated energy loss by constructing an entropy yield field and spatial gradient distribution, and analyze unstable vortices or unfavorable flow modes in the pump cavity, are as follows: S4.1: After the gap control enters a steady state, multiple flow field snapshots are extracted from the pump chamber computation domain of the screw vacuum pump. The velocity, pressure and temperature distributions in each flow field snapshot are organized into a unified sample. Then, all snapshots are combined into a sample matrix. Singular value decomposition is performed on the sample matrix. The number of main modes retained is determined by the energy retention rate, and each retained main mode is used as a reduced-order basis. S4.2: The control equations for continuous flow within the pump chamber of a screw vacuum pump are discretized onto grid nodes to form a computable full-dimensional state. Then, based on the reduced valence base, the order reduction coefficients corresponding to each reduced valence base, and the mean vector of the full-dimensional state, a full-dimensional state equation is established. The reduced valence base is then subjected to a weighted orthogonal projection to obtain the corresponding state evolution equation. S4.3: The implicit time-progression scheme is used to iterate the state evolution equation quickly, and after obtaining the latest order reduction coefficient, it is remapped back to the full-dimensional space to obtain the solved three-dimensional flow field. Then, the entropy productivity of each spatial unit in the three-dimensional flow field of the pump cavity is calculated, and the entropy productivity values ​​of all spatial units are statistically analyzed to establish a complete entropy productivity field. S4.4: Calculate the spatial gradient of each spatial unit based on the established entropy yield field, and set the gradient magnitude. After obtaining the entropy yield field and its gradient, calculate the current high entropy threshold value through the mean entropy yield, the mean entropy yield, and the corresponding threshold coefficient, and mark the region where the entropy yield is greater than the high entropy threshold value as the high entropy loss region. S4.5: Collect the velocity vector of each spatial unit and calculate the vortex index of the velocity field of each spatial unit. At the same time, construct rotation and strain criteria. When the vortex index is higher than the preset vortex threshold and the rotation and strain criteria are greater than 0, the corresponding region is determined to be an unstable vortex candidate region.

6. The intelligent control method for a screw vacuum pump based on digital twins according to claim 5, characterized in that, The specific steps of step VII, which involve dynamically adjusting the screw vacuum pump control strategy based on the obtained entropy yield distribution and feeding the optimized control strategy back to the state simulation model for verification, are as follows: S5.1: Arrange the entropy yield values ​​of each discrete unit in the pump chamber of the screw vacuum pump into candidate paths according to the spatial connectivity relationship, then calculate the cumulative cost of each candidate path, and then set the operating status of the screw vacuum pump at each control time based on historical data, and establish a recursive nonlinear prediction model, and set the rolling optimization objective function corresponding to the nonlinear prediction model. S5.2: Obtain the inverter carrier frequency command and synchronous gear phase difference command for each control moment in the current control strategy, input them into the nonlinear prediction model to predict the operating state of the screw vacuum pump, calculate the optimization target value under the current control quantity through the rolling optimization objective function, adjust the control quantity, and re-predict the state through the nonlinear prediction model. S5.3: Repeat the control quantity adjustment iteration until the optimization target value decreases to within the preset range, then stop the iteration, take the control quantity corresponding to the minimum optimization target value as the optimal control quantity at the current control moment, gradually update the control quantity at each control moment, and output the updated optimal control strategy; S5.4: Based on the pulsation amplitude, pulsation frequency, pulsation initial phase, and control sampling period, calculate the frequency pulsation amount for the corresponding control time. Then, superimpose the corresponding frequency pulsation amount onto the inverter carrier frequency command of the optimal control quantity for the corresponding control time in the optimal control strategy, and generate the final inverter carrier frequency command. At the same time, calculate the rotor's inertial response based on the final inverter carrier frequency command, and then calculate the vortex repeatability intensity in each control cycle in real time. If the vortex repeatability intensity decreases compared to the previous control moment, it indicates that the current control quantity can suppress the regeneration of periodic vortices; otherwise, readjust the control quantity for the current control cycle. S5.5: Based on the optimal control strategy after pulsation correction, the rolling time domain method is used to output only the first action at the current moment, and the screw vacuum pump operating state is recalculated in the next cycle. Then, the rolling optimization objective is solved again using the new feedback state, and the current control strategy is adjusted.

Images