Multi-source sensing oil cylinder clamping stress unevenness prevention control method

Abstract

The invention relates to the technical field of hydraulic servo cooperative control, in particular to a multi-source sensing oil cylinder clamping stress unevenness prevention control method which is applied to a framework comprising a sensing module, a control system and multiple sets of servo hydraulic oil cylinders and comprises the steps that displacement, speed, pressure and pipe network flow of all oil cylinders are collected to construct a low-dimensional state sequence; mapping the state vector into a high-dimensional state vector by using an observation dictionary matrix; constructing a Kupman linear prediction equation in combination with the vector, and calculating a predicted shear stress vector of the target structural member; calculating a difference value between a safety yield stress threshold value and the predicted stress to construct a control barrier function, and generating a hard constraint affine boundary equation; and a cost function is constructed by fusing the trajectory deviation and control increment penalty, convex optimization solution is executed under the rigid constraint condition, an optimal control vector is obtained and issued to an oil cylinder for execution, and a control closed loop is maintained. According to the method, the stress extreme value is pre-judged and intercepted in advance, and plastic failure of the component is effectively prevented.

Description

Technical Field

[0001] This invention relates to the field of hydraulic servo collaborative control technology, and in particular to a multi-source sensing method for preventing uneven stress during cylinder clamping. Background Technology

[0002] In the aerospace and new energy fields, multi-directional coordinated demolding of complex structural components requires the scheduling of multiple sets of servo cylinders. Due to the strong fluid-structure interaction interference between hydraulic network pressure fluctuations and asymmetric frictional forces, existing hysteresis feedback control mechanisms heavily rely on the online iteration of nonlinear Jacobian matrices. Limited by the computing power bottleneck of industrial controllers, complex calculus calculations result in instruction delays of tens of milliseconds, making it impossible for the system to predict and intervene in coupled operating conditions in advance. This leads to imbalances in the movements of each axis, and the accumulation of mechanical distortion energy at local nodes causes shear stress to exceed the material's yield limit, ultimately resulting in plastic failure or surface tensile damage to the structural components. Summary of the Invention

[0003] To overcome the above shortcomings, this invention provides a multi-source sensing method for controlling uneven stress in cylinder clamping, which aims to improve the problem that traditional multi-cylinder control mostly uses hysteresis feedback, which causes components to be prone to plastic failure due to lag in computing power and inability to prevent stress mutations in advance.

[0004] This invention provides the following technical solution: a multi-source sensing method for preventing uneven stress during cylinder clamping, applied to an architecture including a multi-source sensing module, a control system, and multiple sets of servo hydraulic cylinders, comprising the following steps: S1. Based on the multi-source sensing module, collect the displacement, speed, cavity pressure and total flow of hydraulic pipeline of each servo hydraulic cylinder to construct a low-dimensional state sequence; S2. Using the observation dictionary matrix composed of polynomials and radial basis functions, the low-dimensional state sequence is dynamically decomposed and transformed into a high-dimensional state vector. S3. Combining the high-dimensional state vector, construct the Koopman linear prediction equation using the identified high-dimensional state transfer matrix and control input matrix, and obtain the predicted shear stress vector of the target structural component through the observation mapping matrix. S4. Calculate the difference between the safe yield stress threshold and the predicted shear stress vector to construct the control barrier function, introduce a convergence coefficient to reconstruct it into a recursive inequality, and generate a hard constraint affine boundary equation that restricts the control input vector. S5. Construct a quadratic cost function by combining trajectory tracking deviation and control increment penalty, and use the hard constraint affine boundary equation as a hard constraint condition to perform convex optimization solution to obtain the optimal control vector. S6. The optimal control vector is sent to each servo hydraulic cylinder for execution. After the action is completed, the acquisition operation of the low-dimensional state sequence is triggered again to maintain the control closed loop.

[0005] By adopting the above technical solution, the Koopman prediction equation is constructed and combined with the safety yield threshold to generate the hard constraint boundary equation for convex optimization solution, thereby realizing the early prediction and interception of stress extrema. This improves the problem that traditional multi-cylinder control mostly uses hysteresis feedback, which causes the components to be prone to plastic failure due to the lag in computing power and the inability to prevent stress mutations in advance.

[0006] Optionally, in S1, based on the multi-source sensing module, the displacement, velocity, cavity pressure, and total flow rate of each servo hydraulic cylinder are collected to construct a low-dimensional state sequence, including: According to the set unified time sampling benchmark, the real-time displacement data, real-time speed data, working chamber transient pressure data, and transient total flow data of each servo hydraulic cylinder are extracted synchronously. The real-time displacement data, real-time velocity data, transient pressure data of the working chamber, and transient total flow data at the same time section are concatenated and encapsulated according to a set physical dimension order to generate a transient state vector corresponding to a single time point. The transient state vectors from multiple consecutive acquisition cycles are arranged and combined in time sequence to construct a low-dimensional state sequence that reflects the dynamic evolution of the physical system.

[0007] Optionally, in S2, the step of using the observation dictionary matrix composed of polynomials and radial basis functions to perform dynamic mode decomposition transformation on the low-dimensional state sequence and map it to a high-dimensional state vector includes: Extract each state variable from the low-dimensional state sequence, and substitute each state variable into a preset polynomial basis function to generate a first mapping feature subset; Each of the aforementioned state variables is substituted into a radial basis function based on a set state center to generate a second mapping feature subset; The first mapping feature subset and the second mapping feature subset are concatenated by dimension to form the observation dictionary matrix; The inner product space projection operation is performed on the discrete data points in the low-dimensional state sequence using the observation dictionary matrix, and the output is the high-dimensional state vector with a dimension parameter higher than that of the low-dimensional state sequence.

[0008] Optionally, in S3, constructing the Koopman linear prediction equation by combining the high-dimensional state vector with the identified high-dimensional state transfer matrix and control input matrix includes: Retrieve the high-dimensional state transfer matrix and the control input matrix extracted offline using least-squares approximation based on the system's historical operation dataset; By performing matrix multiplication between the high-dimensional state transfer matrix and the high-dimensional state vector at the current sampling time, the autonomous evolution component characterizing the autonomous linear evolution of the system state is obtained. By performing matrix multiplication between the control input matrix and the control input vector at the current sampling time, the external intervention component characterizing the external control's intervention on the system state is obtained; The Koopman linear prediction equation for predicting the state evolution result at the next discrete time step is constructed by performing a linear algebraic superposition of the autonomous evolution component and the external intervention component.

[0009] Optionally, in S3, obtaining the predicted shear stress vector of the target structural member through the observation mapping matrix includes: Extract the observation mapping matrix that was previously generated based on the three-dimensional finite element stiffness model calibration of the target structural component; Extract the high-dimensional predicted state vector for the next discrete time step, derived from the Koopman linear prediction equation; The observation mapping matrix is ​​used to perform a dimension reduction linear mapping transformation operation on the high-dimensional predicted state vector, and the high-dimensional predicted state vector located in the high-dimensional Hilbert function space is inversely transformed into the physical three-dimensional task space, and the predicted shear stress vector corresponding to each stress node of the target structural component in the next discrete time step is extracted.

[0010] Optionally, in S4, calculating the difference between the safe yield stress threshold and the predicted shear stress vector to construct the control barrier function includes: Extract the safe yield stress threshold pre-calibrated according to the allowable stress limit physical properties of the target structural component material; The safe yield stress threshold is set as the upper limit boundary of the system safety. The subtraction operation is performed between each predicted shear stress component in the predicted shear stress vector and the safe yield stress threshold to obtain the stress margin vector corresponding to each stress node. Based on the non-negative characteristics of each data element in the stress margin vector, a physical safety set for which the target system does not undergo plastic deformation is defined, and the spatial boundary equation enclosing the physical safety set is defined as the control barrier function under the boundary of the continuous domain.

[0011] Optionally, in S4, the introduction of convergence coefficients to reconstruct it into a recursive inequality, generating the hard-constrained affine boundary equations that restrict the control input vector, includes: Introduce the convergence coefficient that determines the rate at which the system state approaches the boundary of the physical security set; The recursive inequality that embodies the forward invariance characteristic is constructed using the current barrier function value, the barrier function value at the next prediction time, and the convergence coefficient. Substitute the control input variables in the Koopman linear prediction equation into the recursive inequality and expand it to separate the linear correlation term and system constant term containing the control input variables. Reconstruct the linear inequality for the control input vector and use the linear inequality as the hard constraint affine boundary equation.

[0012] Optionally, in S5, the construction of the quadratic cost function by combining trajectory tracking deviation and control increment penalty includes: Extract the set reference demolding displacement trajectory sequence, compare the difference between the reference demolding displacement trajectory sequence and the actual displacement state value fed back by the multi-source sensing module, and generate the trajectory tracking deviation; Extract the control input change rate that reflects the smoothness of the action of each servo hydraulic cylinder, quantify the control input change rate in combination with the set positive definite weight matrix, and generate the control increment penalty; The quadratic cost function is constructed by performing a weighted summation operation on the quadratic term of the trajectory tracking deviation and the quadratic term of the control increment penalty, in the form of a standard quadratic function.

[0013] Optionally, in S5, the hard-constrained affine boundary equation is used as a hard constraint condition for convex optimization to obtain the optimal control vector, including: The constructed quadratic cost function is set as the minimum objective function of the convex optimization algorithm; The hard-constrained affine boundary equation is set as an absolute physical boundary condition that cannot be violated during the optimization iteration process of the convex optimization algorithm. The interior point method is used to perform an iterative search operation in the feasible control input space that satisfies the absolute physical boundary conditions, calculate the control input solution that minimizes the quadratic cost function, and establish the control input solution as the optimal control vector.

[0014] Optionally, in S6, the process of sending the optimal control vector to each servo hydraulic cylinder for execution, and then re-triggering the acquisition operation of the low-dimensional state sequence after the action is completed includes: The digital commands in the optimal control vector are converted into matching analog voltage drive signals; The analog voltage drive signal is transmitted in parallel to the servo proportional valve control terminal inside each servo hydraulic cylinder, driving each servo hydraulic cylinder to perform the associated physical displacement action according to the analog voltage command. At the end of the control cycle of each servo hydraulic cylinder completing a single drive command, a trigger interrupt signal is generated. The trigger interrupt signal is used to wake up the data sampling array of the multi-source sensing module and force the entry into the physical parameter extraction and low-dimensional state sequence construction process of the next discrete time step.

[0015] The present invention has the following beneficial effects: 1. In this invention, by constructing the Koopman prediction equation and combining it with the safety yield threshold to generate the hard constraint boundary equation for convex optimization solution, stress extremum prediction and interception are achieved, thereby improving the problem that traditional multi-cylinder control mostly uses hysteresis feedback, which causes components to be prone to plastic failure due to lag in computing power and inability to prevent stress mutations in advance.

[0016] 2. In this invention, dynamic pattern decomposition and dimensionality enhancement are performed by using a dictionary matrix containing polynomials and radial basis functions, thereby converting nonlinear physical quantities into linear features. This improves the problem that traditional nonlinear deductions mostly use matrix inversion, which is extremely time-consuming due to iterative calculations, resulting in severe delays in the issuance of control commands.

[0017] 3. In this invention, a quadratic cost function is constructed by fusing trajectory tracking deviation and control increment penalty to perform optimization solution, thereby achieving a comprehensive quantitative balance between demolding accuracy and motion smoothness. This improves the problem that traditional displacement control mostly adopts a single following strategy, which lacks constraints on motion intensity and thus causes pipeline impact.

[0018] 4. In this invention, an interrupt signal is generated at the moment the single drive of the hydraulic cylinder is completed to wake up the sampling array, thereby forcibly aligning the time axis of the end point of the physical action with the start point of the next acquisition. This improves the problem that traditional cyclic acquisition mostly uses timed polling, which is prone to misalignment with the actual action due to the fixed period, thus causing sampling errors. Attached Figure Description

[0019] Figure 1 The flowchart shows the multi-source sensing hydraulic cylinder clamping stress unevenness control method proposed in this invention. Figure 2 The flowchart shows the low-dimensional state sequence for constructing the multi-source sensing hydraulic cylinder clamping anti-stress unevenness control method proposed in this invention. Figure 3 This is a flowchart of the high-dimensional state vector mapping of the multi-source sensing hydraulic cylinder clamping anti-stress unevenness control method proposed in this invention. Figure 4 This is a flowchart illustrating the predicted shear stress vector calculation process of the multi-source sensing hydraulic cylinder clamping anti-stress unevenness control method proposed in this invention. Detailed Implementation

[0020] The technical solutions in 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.

[0021] Example 1: In the first embodiment of the present invention, the present invention provides a multi-source sensing method for preventing uneven stress during cylinder clamping, which is applied to an architecture including a multi-source sensing module, a control system, and multiple sets of servo hydraulic cylinders, such as... Figures 1-4 As shown, it includes the following steps: S1. Based on the multi-source sensing module, the displacement, speed, cavity pressure and total flow of the hydraulic pipeline of each servo hydraulic cylinder are collected to construct a low-dimensional state sequence; Furthermore, in S1, the displacement, velocity, cavity pressure, and total flow rate of each servo hydraulic cylinder are collected based on the multi-source sensing module to construct a low-dimensional state sequence, including: According to the set unified time sampling benchmark, the real-time displacement data, real-time speed data, transient pressure data of the working chamber, and transient total flow data of the hydraulic pipeline network of each servo hydraulic cylinder are extracted synchronously. Real-time displacement data, real-time velocity data, transient pressure data of the working chamber, and transient total flow data at the same time section are concatenated and encapsulated according to the set physical dimension order to generate a transient state vector corresponding to a single time point. The transient state vectors from multiple consecutive acquisition cycles are arranged and combined in time sequence to construct a low-dimensional state sequence that reflects the dynamic evolution of the physical system.

[0022] Specifically, in the scenario of multi-directional hydraulic cylinder collaborative control, there is a strong nonlinear fluid-structure interaction interference between the total flow fluctuation of the hydraulic pipeline network and the mechanical motion of each hydraulic cylinder actuator. Step S1, as the data foundation of the entire control method, plays a core role in converting heterogeneous physical analog quantities into mathematical sequences with rigorous logical relationships, providing complete boundary input conditions for subsequent high-dimensional reduction decoupling and state prediction.

[0023] The continuous analog electrical signals acquired by the multi-source sensing module serve as the initial input. The control system triggers a unified time base at its core to extract discrete digital features. Subsequently, the extracted parameters are cascaded and recombined to output a multi-dimensional transient state vector. Finally, a sliding data window is used to extract a continuous vector array, outputting a low-dimensional state sequence containing a temporal feature matrix, which is then processed by the subsequent dimensionality reduction algorithm module.

[0024] Due to inherent physical delays in the communication lines between various sensors, data asynchrony can lead to input disorder in the control model. Synchronously extracting multidimensional data according to a set unified time sampling benchmark completely eliminates sampling phase differences between sensors at different locations and of different types. This ensures that the extracted mechanical kinematic features and hydraulic fluid dynamic features are strictly aligned on the absolute time axis, avoiding misjudgments of coupling states caused by timestamp misalignment.

[0025] After data alignment is completed, the problem of unifying the expression of physical quantities with different dimensions needs to be addressed. A specific sampling time is defined as... The total number of servo hydraulic cylinders configured in the system is set to .set up Representative includes Real-time displacement data vector of each hydraulic cylinder position information. (Setting) Representative characterization Real-time speed data vectors for the movement speed of each hydraulic cylinder. (Settings) Representatives reflected Transient pressure data vector of the working chamber of each hydraulic cylinder driving force. (Setting) The transient total flow rate of the hydraulic pipeline network represents the overall power supply status of the system.

[0026] The parameters mentioned above at the same time section are cascaded and encapsulated according to the dimensional order of physical space constraints and fluid energy dissipation to establish the first... Transient state vector for each sampling period The specific vector concatenation equation is expressed as follows: ; The letter in the upper right corner of the equation This indicates that the transpose operation is performed on the concatenated matrix. The transient state vector constructed through this operation fundamentally breaks the parameter isolation of a single component, and completely describes the distribution relationship of the overall mechanical potential energy, kinetic energy, and hydraulic fluid field energy of the system within a single time slice.

[0027] Since a single transient state vector can only characterize a static snapshot of the physical system, it lacks the ability to represent the evolutionary laws. The system further defines the time sliding window length of the continuous acquisition cycle as... The low-dimensional state sequence is defined as a matrix by arranging and combining multiple consecutive transient state vectors within the sliding window according to their temporal evolution order. The sequence combination expression is presented as: ; The letters in the equation The transient state vector at the starting point of the current analysis window, represented by the letter... To analyze the transient state vector at the end of the window.

[0028] The processing logic of upscaling transient data to a time-series matrix endows static physical quantities with dynamic evolution characteristics. The low-dimensional state sequence fully records the true trajectory of fluid wave dynamics characteristics propagating along the time axis, providing the necessary underlying time series foundation for subsequent data-driven dynamic mode decomposition. This directly avoids the response lag and action interference defects caused by the lack of understanding of the global evolution of the system in traditional single-variable proportional-integral-derivative control strategies.

[0029] S2. Using the observation dictionary matrix composed of polynomials and radial basis functions, the low-dimensional state sequence is dynamically decomposed and transformed into a high-dimensional state vector. Furthermore, in S2, the observation dictionary matrix composed of polynomials and radial basis functions is used to perform dynamic mode decomposition transformation on the low-dimensional state sequence, mapping it to a high-dimensional state vector including: Extract each state variable from the low-dimensional state sequence, and substitute each state variable into a preset polynomial basis function to generate the first mapping feature subset; Each state variable is substituted into the radial basis function based on the set state center to generate the second mapping feature subset; The first and second mapping feature subsets are concatenated by dimension to form the observation dictionary matrix; The inner product space projection operation is performed on the discrete data points in the low-dimensional state sequence using the observation dictionary matrix, and the output is a high-dimensional state vector with a higher dimension parameter than the low-dimensional state sequence dimension parameter.

[0030] Specifically, the data input / output flow is as follows: The input of this step receives the low-dimensional state sequence established in the previous step. The underlying algorithm extracts and splits the discrete state variables within the sequence, substituting them into the polynomial calculation logic and the radial basis function calculation logic respectively for dimensionality-upgrading. After processing, an observation dictionary matrix is ​​generated, composed of features from two types of basis functions. The output uses this matrix to perform spatial projection, outputting a high-dimensional state vector with its dimension forcibly expanded, which serves as the sole data input source for subsequent prediction evolution equations.

[0031] In multi-cylinder servo hydraulic control scenarios, severe nonlinear fluid-structure interaction (FSI) exists between cylinder chamber pressure, pipeline flow, and mechanical end displacement. If future state predictions are directly based on data from a low-dimensional physical space, the system must rely on complex nonlinear Jacobian partial derivative iterations for solution, which consumes a significant amount of programmable logic controller (PLC) computation cycles, leading to fatal physical lag in servo valve control command issuance. The core function of this step, dynamic mode decomposition and eigenmaps, is to project low-dimensional nonlinear physical quantities onto a high-dimensional Hilbert function space using specific mathematical basis functions. Utilizing Koopman operator-related control theory, the strongly nonlinear coupled dynamic behavior, originally in low-dimensional space, can be precisely equivalent to globally linear dynamic behavior in high-dimensional space. This completely eliminates the computational bottleneck of online inversion of nonlinear equations, paving the way for subsequent matrix linear prediction within extremely short clock cycles.

[0032] Let the low-dimensional state variables of the low-dimensional state sequence at a specific sampling discrete time be... Define the original physical dimension parameter of this state variable as: .

[0033] Extracting low-dimensional state variables Substituting the polynomial basis function into a preset set, performing exponentiation and cross-term operations, generates the first mapping feature subset. The expanded expression for the first mapping feature subset is: ; in Expression of the first Polynomial basis elements of order 1 The total number of polynomial basis function terms in the system configuration is shown in the upper right corner. Expressing the transpose of matrices and vectors.

[0034] Extracting the same low-dimensional state variables Substituting into the radial basis function set, a second mapping feature subset is generated to capture local nonlinear mutation features. The specific mathematical formula for the radial basis functions used at the bottom layer is as follows: ; in Representing the The output values ​​of the radial basis solution. Represents the first set of parameters pre-defined based on historical fluid dynamics data. The coordinates of the state center point are variables. The width parameter represents the rate of decay within the local effective range of the radial basis function. The Euclidean distance norm represents the distance between the current low-dimensional state variable vector and the set state center vector. This represents the exponential law with the natural constant as its base. The results from each computation node are aggregated to generate a second subset of mapping features. ;in This represents the total number of radial basis functions used in the solution.

[0035] Retrieve the first subset of generated mapping features and the second mapping feature subset By performing dimension concatenation and combination operations according to the column vector arrangement rules, an observation dictionary matrix with both global and local perception capabilities is constructed. The concatenated relational expression is presented as follows: .

[0036] Using a constructed, fixed observation dictionary matrix Perform inner product spatial projection mapping on the discrete data points contained in the low-dimensional state sequence. The output is a high-dimensional state vector. The specific matrix mapping correlation is as follows: The high-dimensional state vector output by the solution. Current dimension parameters Equal to the number of polynomial bases With radial base number The accumulated value. This is due to the total number of configured feature base items. Strictly larger than the original physical dimension parameters The system completes the data upscaling transformation within the physical control logic.

[0037] S3. Combining the high-dimensional state vector, the Koopman linear prediction equation is constructed using the identified high-dimensional state transfer matrix and control input matrix, and the predicted shear stress vector of the target structural component is obtained through the observation mapping matrix. Furthermore, in S3, by combining the high-dimensional state vector and utilizing the identified high-dimensional state transfer matrix and control input matrix, the Koopman linear prediction equation is constructed, including: Retrieve the high-dimensional state transfer matrix and control input matrix extracted offline using least-squares approximation identification based on the system's historical operation dataset; By performing matrix multiplication between the high-dimensional state transfer matrix and the high-dimensional state vector at the current sampling time, the autonomous evolution components characterizing the autonomous linear evolution of the system state are obtained. By performing matrix multiplication between the control input matrix and the control input vector at the current sampling time, the external intervention component characterizing the external control's intervention on the system state is obtained; By performing a linear algebraic superposition of the autonomous evolution component and the external intervention component, a Koopman linear prediction equation is constructed to predict the state evolution result of the next discrete time step.

[0038] In S3, the predicted shear stress vector of the target structural member is obtained by observing the mapping matrix, including: Extract the observation mapping matrix generated in advance based on the three-dimensional finite element stiffness model calibration of the target structural component; Extract the high-dimensional predicted state vector for the next discrete time step derived from the Koopman linear prediction equation; By performing a dimension reduction linear mapping transformation operation on the high-dimensional predicted state vector using the observation mapping matrix, the high-dimensional predicted state vector located in the high-dimensional Hilbert function space is inversely transformed into the physical three-dimensional task space, and the predicted shear stress vector of each stress node of the corresponding target structural component in the next discrete time step is extracted.

[0039] Specifically, in multi-cylinder hydraulic synchronous demolding, traditional online solutions using nonlinear prediction models incur significant computational burdens and communication delays. The core function of step S3 is to completely eliminate real-time iterative calculations of the nonlinear Jacobian partial derivative matrix. The control algorithm transforms the extremely complex fluid wave dynamics and spatial interference motion of mechanical components into a purely high-dimensional linear algebraic superposition process, forcibly compressing the computation time from the traditional tens of milliseconds to the microsecond level. Combined with dimensionality reduction extraction techniques, the control algorithm can overcome the physical differences between hydraulic fluid parameters and solid deformation, proactively sensing the stress extrema of the mold's stress nodes, thus providing a window of opportunity for the servo system to intervene early.

[0040] The algorithm's input side extracts the high-dimensional state vector generated by the mapping in the preceding steps and the control input vector issued by the system at the current moment. The data is then processed within the underlying processor using forward evolution multiplication with matrix coefficients extracted offline from the system's historical runtime dataset, outputting the predicted state data within the high-dimensional space. Finally, a dimensionality-reduced mapping matrix calibrated by the finite element model is introduced to perform an inverse projection operation, directly generating the predicted shear stress vector of the target structural component within the corresponding real physical three-dimensional task space.

[0041] During the implementation phase of constructing the Koopman linear prediction equation, the system storage unit pre-loads the high-order matrix parameters extracted through historical operating data and least squares approximation. The extracted high-dimensional state transfer matrix is ​​set as follows: The extracted control input matrix is ​​set as follows: The system extracts the corresponding high-dimensional state vector at the current discrete-time sampling point. and control input vector .

[0042] The central processing unit executes matrix multiplication logic, transferring high-dimensional states to matrices. With high-dimensional state vector Multiplication yields an autonomous evolutionary component. This autonomous evolutionary component characterizes the system's natural physical decay trend and inertial trajectory under no external command intervention. The computational unit synchronously transmits the control input matrix... With control input vector Multiplying these components yields the external intervention component. This external intervention component represents the forced state change value triggered by the injection of commands such as the current servo valve opening into the system. Performing a linear algebraic superposition on the aforementioned two components constructs the Koopman linear prediction equation describing the system's evolution state at the next discrete time step. The specific mathematical expression is set as follows: ; in This represents the high-dimensional predicted state vector for the next discrete time step obtained through the derivation. All evolutionary derivation calculations are strictly limited to basic matrix multiplication and addition operations.

[0043] The high-dimensional predicted state vector output by the prediction equation derivation Located within the abstract Hilbert function space, it cannot be directly used for physical boundary comparison in mechanical anti-eccentric loading. The system continues to retrieve the observation mapping matrix pre-generated based on the three-dimensional finite element stiffness model of the target structural component, setting the retrieved observation mapping matrix to... .

[0044] The central processing unit will observe the mapping matrix. With high-dimensional prediction state vector Perform a multiplication operation to execute a dimension reduction linear mapping transformation. The specific mathematical expression for the dimension reduction solution process is defined as follows: ; in This represents the predicted shear stress vector of each stress node of the corresponding target structural component at the next discrete time step after reverse transformation to the physical three-dimensional task space.

[0045] Through the aforementioned matrix space mapping process, the abstract high-dimensional mathematical prediction features are accurately reduced to the actual stress distribution values ​​that the target structural component will experience within a very short time span in the future. The calculated predicted shear stress vector provides precise physical target parameters for subsequent steps to establish a rigid protective boundary to prevent plastic micro-deformation of the mold material.

[0046] S4. Calculate the difference between the safe yield stress threshold and the predicted shear stress vector to construct the control barrier function, introduce a convergence coefficient to reconstruct it into a recursive inequality, and generate a hard constraint affine boundary equation that restricts the control input vector. Furthermore, in S4, calculating the difference between the safe yield stress threshold and the predicted shear stress vector to construct the control barrier function includes: Extract the safe yield stress threshold pre-calibrated according to the physical properties of the allowable stress limit of the target structural component material; The safe yield stress threshold is set as the upper limit boundary of the system safety. The subtraction operation is performed between each predicted shear stress component in the predicted shear stress vector and the safe yield stress threshold to obtain the stress margin vector corresponding to each stress node. Based on the non-negative characteristics of each data element in the stress margin vector, the physical safety set of the target system that does not undergo plastic deformation is defined, and the spatial boundary equation of the envelope physical safety set is defined as the control barrier function under the continuous domain boundary.

[0047] In S4, convergence coefficients are introduced to reconstruct it into a recursive inequality, generating the hard-constrained affine boundary equations that restrict the control input vector, including: Introduce a convergence coefficient that determines the rate of evolution of the system state approaching the boundary of the physical safety set; We construct a recursive inequality that reflects forward invariance using the current barrier function value, the barrier function value at the next prediction time, and the convergence coefficient. Substituting the control input variables in the Koopman linear prediction equation into the recursive inequality, we can separate the linear correlation terms containing the control input variables and the system constant terms, reconstruct them into linear inequalities for the control input vector, and use the linear inequalities as hard-constrained affine boundary equations.

[0048] Specifically, the core technical action of this step is to construct an insurmountable mathematical constraint boundary within the physical control link base. In the multi-directional demolding conditions of complex structural components, local extreme shear stresses often exceed the material's yield strength within milliseconds, triggering irreversible microscopic tearing or plastic failure. Stress prediction alone can only serve as a post-event warning. By introducing control barrier functions and affine boundary reconstruction logic, the early warning mechanism is completely transformed into a proactive, hard-line defense mechanism. The system establishes strict mathematical inequalities to physically isolate and eliminate any control command vectors that pose a potential risk of damaging the mold before they reach the servo valve actuator.

[0049] The data input receives the high-dimensional state vector generated by the previous steps and the predicted shear stress vector extracted through dimensionality reduction. The underlying computational core synchronously retrieves the material safety yield stress threshold and the set convergence coefficient, which are stored in non-volatile memory. The data flows through difference comparison logic to generate stress margin and continuous domain functions, which are then fed into a discrete-time algebraic iteration array for dimensionality reduction expansion. The output finally generates a linear inequality that restricts the range of values ​​for the control input vector. This linear inequality directly serves as the absolute boundary constraint condition for the downstream quadratic programming solution module.

[0050] At the technical implementation level, the central processing unit retrieves the preset safety yield stress threshold based on the material properties of the target structural component. The aforementioned safe yield stress threshold is set as the upper limit boundary for the system's anti-eccentric load operation. The computational logic extracts the predicted shear stress vector. The stress margin vector is obtained by performing a subtraction difference calculation. Based on physical principles, the absence of plastic deformation in a structural component is equivalent to all elements within the stress margin vector remaining non-negative. The spatial boundary equation enclosing this non-negative safety set is defined as the control barrier function. The specific relational equation configuration is as follows: ; The equation contains... The control barrier function, representing the current discrete time step, calculates the state value. Express the target structural component's safe yield stress threshold. The expression system obtains the predicted shear stress vector at the current moment through deduction.

[0051] To achieve proactive action blocking, the computational logic extracts a pre-defined convergence coefficient. The convergence coefficient determines the hysteresis rate of the system's operating state as it approaches the boundary of the physically safe set. This is achieved using the barrier function value at the current discrete moment. And the barrier function value for the next predicted discrete time step. A recursive inequality embodying the characteristics of forward invariance theory is constructed to ensure that the system's evolution trajectory is firmly contained within the safe set. This recursive inequality is expressed as follows: ; The internal relation The state value derived from the control barrier function represents the next prediction time step. The range of values ​​is strictly greater than zero and less than or equal to one.

[0052] The control system further extracts the previously constructed Koopman linear prediction equations. The high-dimensional state transitivity and control input intervention relations are substituted into the aforementioned recursive inequalities for expansion. Algebraic terms containing control input variables are extracted and linear correlation terms are established, while the remaining known state variables are grouped into system constant terms. Rearranging and simplification operations are performed to reconstruct the equations for the control input vector. The hard-constrained affine boundary equations. The reconstructed derivation generates linear inequalities of the form: ; in The observation mapping matrix represents the reduction of the high-dimensional function space to a three-dimensional task space. This represents the control input matrix that describes the characteristics of external intervention. This represents the control input vector at the current time step, which is waiting to be optimized and solved. This represents the identity matrix, where all elements on the diagonal are one. This represents a high-dimensional state transfer matrix that describes the natural evolution of states. This represents the high-dimensional state vector extracted at the current moment. This hard-constrained affine boundary equation cuts off the generation paths of all illegal drive commands, endowing the multi-cylinder servo hydraulic system with an absolute control barrier from the algorithm's underlying layers to ensure the preservation of mold structural components from damage.

[0053] S5. Construct a quadratic cost function by combining trajectory tracking deviation and control increment penalty, and use the hard constraint affine boundary equation as a hard constraint condition to perform convex optimization solution to obtain the optimal control vector. Furthermore, in S5, the quadratic cost function constructed by combining trajectory tracking deviation and control increment penalty includes: Extract the set reference demolding displacement trajectory sequence, compare the difference between the reference demolding displacement trajectory sequence and the actual displacement state value fed back by the multi-source sensing module, and generate trajectory tracking deviation; Extract the rate of change of control input reflecting the smoothness of the action of each servo hydraulic cylinder, quantify the rate of change of control input by combining it with the set positive definite weight matrix, and generate control increment penalty; By performing a weighted summation of the quadratic terms of the trajectory tracking deviation and the quadratic terms of the control increment penalty, a quadratic cost function in the standard form of a mathematical quadratic form is constructed.

[0054] In S5, the hard-constrained affine boundary equations are used as hard constraints for convex optimization to obtain the optimal control vector, including: The constructed quadratic cost function is set as the minimum objective function of the convex optimization algorithm; The hard-constrained affine boundary equation is set as an absolute physical boundary condition that cannot be violated during the optimization iteration process of the convex optimization algorithm. The interior point method is used to perform iterative search operations in the feasible control input space that satisfies the absolute physical boundary conditions, calculate the control input solution that minimizes the quadratic cost function, and establish the control input solution as the optimal control vector.

[0055] Specifically, the algorithm input in this step receives the reference demolding displacement trajectory sequence issued by the central control baseline, the actual displacement state values ​​fed back by the underlying sensors, and the hard-constrained affine boundary equations derived from the previous steps. The central processing unit receives the aforementioned three sets of data and performs quadratic function encapsulation and interior-point spatial optimization operations. The output generates a unique optimal control vector corresponding to the current control cycle, which serves as the final drive digital command for the physical servo valve actuator response.

[0056] In the complex control mechanism of multiple servo hydraulic cylinders operating in parallel, simply pursuing absolute overlap of the demolding trajectory can easily trigger high-frequency abrupt jitter at the hydraulic valve ports, thereby inducing hydraulic shock in the pipeline network. The core function of constructing the quadratic cost function in step S5 is to quantify and balance the two major control indicators: spatial position tracking accuracy and physical motion smoothness. By combining hard constraints with convex optimization, the system establishes the ultimate stress protection mechanism against material damage as an inviolable underlying logic. The control architecture can automatically search for the optimal drive command that balances motion accuracy and energy consumption within the mathematically feasible domain, ensuring absolute mold safety.

[0057] The control system extracts the reference demolding displacement trajectory sequence values ​​at the current moment and simultaneously extracts the actual displacement state values ​​fed back by the multi-source sensing module. The reference demolding displacement trajectory sequence values ​​and the actual displacement state values ​​are subtracted to generate the trajectory tracking deviation matrix corresponding to the current sampling period. This trajectory tracking deviation matrix is ​​set as follows: .

[0058] The control logic continues to extract the difference between the current control command and the control command at the previous time step, defining this feature parameter as the control input rate of change matrix. This control input rate of change matrix is ​​then set as follows: A first positive definite weight matrix representing the trajectory tracking accuracy is introduced. And the second positive definite weight matrix representing the smoothness of the motion. By integrating the trajectory tracking deviation matrix with the control input rate of change matrix, a quadratic cost function in the standard form of a quadratic form is constructed. The specific mathematical equation is as follows: ; in Express the transpose of the trajectory tracking deviation matrix. This is expressed in the transpose of the control input rate of change matrix. The calculated value of this quadratic cost function directly reflects the combined execution error and control energy loss generated by the current demolding action.

[0059] During the implementation phase of optimizing and solving the extreme value control command, the underlying controller will use the aforementioned quadratic cost function. The objective function is set as the minimum value search function of the convex optimization algorithm. The hard-constrained affine boundary equations containing the system's limit safety boundary are extracted. The coefficient matrix of these boundary equations is set as follows. The boundary constant matrix is ​​set as Control input vector Forced to strictly satisfy linear inequalities The absolute physical boundary conditions.

[0060] The underlying computational core calls the interior-point search algorithm. Based on the principle of penalized iterative computation, the interior-point method performs multiple calculations along the gradient descent path of the objective function within the constrained polyhedral feasible region. Under the premise of satisfying the absolute physical boundary conditions of the linear inequalities, the system calculates the quadratic cost function. The solution for the control input variable with the minimum value is obtained. At the end of the operation, this control input solution is extracted and established as the optimal control vector to be sent to the dynamic execution module.

[0061] S6. Send the optimal control vector to each servo hydraulic cylinder for execution and drive. After the action is completed, re-trigger the acquisition operation of the low-dimensional state sequence to maintain the control closed loop. Furthermore, in S6, the optimal control vector is sent to each servo hydraulic cylinder for execution, and the low-dimensional state sequence acquisition operation is re-triggered after the action is completed, including: Convert the digital commands in the optimal control vector into matching analog voltage drive signals; The analog voltage drive signal is transmitted in parallel to the servo proportional valve control terminal inside each servo hydraulic cylinder, driving each servo hydraulic cylinder to perform the associated physical displacement action according to the analog voltage command. At the end of the control cycle of each servo hydraulic cylinder completing a single drive command, a trigger interrupt signal is generated. The trigger interrupt signal is used to wake up the data sampling array of the multi-source sensing module and force the entry into the physical parameter extraction and low-dimensional state sequence construction process of the next discrete time step.

[0062] Specifically, the algorithm layer outputs the optimal control vector as the digital input parameters for the underlying hardware. The central controller's analog-to-digital converter outputs an analog voltage drive signal that flows to the physical actuator. At the end of the specified action cycle, the control loop containing the actuator outputs a trigger interrupt signal that flows to the sensing hardware layer.

[0063] Step S6 constitutes the final action channel for mapping the virtual control domain to the physical domain. All previous mathematical optimization derivations rely on this step to realize the actual mechanical actions that block extreme stresses. After the control system obtains the optimal control vector in the discrete mathematical domain, it performs hardware-level analog-to-digital conversion analysis. The instruction mapping operation relationship follows the set linear transformation equation: ; Among the variables This represents the vector of analog voltage drive signals transmitted to each servo hydraulic cylinder. (Variable) This represents the hardware gain scaling matrix for the built-in analog-to-digital conversion channel of the central control system. (Variables) This represents the optimal digital control vector determined through convex optimization. Variables This represents the inherent electrical zero-point bias voltage vector at the control terminal of the servo proportional valve.

[0064] After acquiring the analog voltage drive signal, the system uses a parallel transmission mechanism to send data. This transmission logic eliminates the time difference in command reception between execution nodes under conventional serial communication mechanisms, ensuring that multiple servo hydraulic cylinders receive the drive level and perform physical displacement actions at the same absolute point in time. This operation eliminates secondary nonlinear interference caused by inconsistent arrival of action commands at the hardware communication level.

[0065] The closed-loop action at the end of the control cycle relies on an interrupt signal for completion. Conventional timed polling sampling mechanisms continuously consume the controller's underlying computing power and introduce sampling phase errors. This step involves the underlying logic circuit directly sending an interrupt level signal to the data sampling array of the multi-source sensing module at the instant a single drive action is completed. Upon receiving this signal, the data sampling array is forcibly awakened and immediately enters the physical parameter extraction process for the next discrete time step. This interrupt wake-up mechanism ensures seamless alignment of the end point of the physical action with the start point of the next data acquisition on the time axis, guaranteeing the timestamp accuracy of subsequent low-dimensional state sequence reconstruction and maintaining the rigorous cyclic operation of the anti-stress unevenness control logic.

[0066] Example 2: The application scenario involves the multi-directional collaborative demolding and clamping process of large, complex, irregularly shaped thin-walled structural components in the aerospace and new energy vehicle manufacturing fields. In this actual production scenario, for molds with multi-dimensional nested cavity features, multiple sets of servo hydraulic cylinders must be scheduled to perform concurrent displacement movements along different spatial coordinate axes to achieve precise core pulling and component detachment. As multiple sets of cylinders simultaneously pump fluid within an extremely short start-up window, the main hydraulic pipeline network will experience a strong fluid-structure interaction, resulting in a sudden drop in pressure and drastic fluctuations in flow. At the same time, the transient physical frictional resistance between the complex irregularly shaped structural components and the contact surfaces of the mold cavity exhibits extremely strong spatial asymmetry and nonlinear step-change characteristics. Under the dual interference constraints of transient drop in hydraulic power source and nonlinear disturbance of mechanical space load, traditional displacement or force feedback control frameworks, due to computational bottlenecks, cannot predict and compensate for multi-physics domain coupled conditions. This easily leads to severe imbalance in the actual drag force output along various action axes, forcing weak nodes of structural components that have not yet completely escaped rigid constraints to be subjected to malignantly accumulated mechanical distortion energy instantaneously. This causes the local extreme shear stress to instantly exceed the safe yield limit of the target material, ultimately leading to irreversible microcrystalline lattice fracture, surface tensile damage, or overall plastic failure of the structural components. To solve the above problems, this invention employs a multi-source sensing hydraulic cylinder clamping anti-stress unevenness control method, the structure of which is as follows: Figure 1 As shown. The specific implementation process of this method is as follows: The system synchronously collects multidimensional mechanical and fluid parameters to establish a low-level state sequence, and then forces it to be mapped to a high-dimensional function space through dynamic pattern decomposition, transforming it into a global linear model, thus completely overcoming the computational delay bottleneck of nonlinear solutions. Subsequently, relying on ultra-fast linear matrix deduction, it obtains the predicted shear stress of the future structural components in advance, and constructs a control barrier function in combination with the material yield threshold, generating a hard-constrained affine boundary equation to intercept dangerous commands. Under this insurmountable absolute safety boundary, the system integrates trajectory deviation and motion change rate to establish a quadratic cost function and performs convex optimization to obtain the optimal control vector that balances demolding accuracy and smoothness. Finally, the command is issued to the low-level hydraulic cylinder to execute physical actions and immediately triggers a new round of sampling, achieving seamless synchronization between the time axis of the virtual algorithm and the physical execution, and relying on continuous data looping to prevent plastic failure of the structural components.

[0067] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-source sensing method for preventing uneven stress during cylinder clamping, applied to an architecture including a multi-source sensing module, a control system, and multiple sets of servo hydraulic cylinders, characterized in that... Includes the following steps: S1. Based on the multi-source sensing module, collect the displacement, speed, cavity pressure and total flow of hydraulic pipeline of each servo hydraulic cylinder to construct a low-dimensional state sequence; S2. Using the observation dictionary matrix composed of polynomials and radial basis functions, the low-dimensional state sequence is dynamically decomposed and transformed into a high-dimensional state vector. S3. Combining the high-dimensional state vector, construct the Koopman linear prediction equation using the identified high-dimensional state transfer matrix and control input matrix, and obtain the predicted shear stress vector of the target structural component through the observation mapping matrix. S4. Calculate the difference between the safe yield stress threshold and the predicted shear stress vector to construct the control barrier function, introduce a convergence coefficient to reconstruct it into a recursive inequality, and generate a hard constraint affine boundary equation that restricts the control input vector. S5. Construct a quadratic cost function by combining trajectory tracking deviation and control increment penalty, and use the hard constraint affine boundary equation as a hard constraint condition to perform convex optimization solution to obtain the optimal control vector. S6. The optimal control vector is sent to each servo hydraulic cylinder for execution. After the action is completed, the acquisition operation of the low-dimensional state sequence is triggered again to maintain the control closed loop.

2. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S1, the displacement, velocity, cavity pressure, and total flow rate of each servo hydraulic cylinder are collected based on the multi-source sensing module to construct a low-dimensional state sequence, including: According to the set unified time sampling benchmark, the real-time displacement data, real-time speed data, working chamber transient pressure data, and transient total flow data of each servo hydraulic cylinder are extracted synchronously. The real-time displacement data, real-time velocity data, transient pressure data of the working chamber, and transient total flow data at the same time section are concatenated and encapsulated according to a set physical dimension order to generate a transient state vector corresponding to a single time point. The transient state vectors from multiple consecutive acquisition cycles are arranged and combined in time sequence to construct a low-dimensional state sequence that reflects the dynamic evolution of the physical system.

3. The multi-source sensing method for preventing uneven stress during cylinder clamping as described in claim 1, characterized in that, In S2, the dynamic mode decomposition transformation of the low-dimensional state sequence into a high-dimensional state vector using the observation dictionary matrix composed of polynomials and radial basis functions includes: Extract each state variable from the low-dimensional state sequence, and substitute each state variable into a preset polynomial basis function to generate a first mapping feature subset; Each of the aforementioned state variables is substituted into a radial basis function based on a set state center to generate a second mapping feature subset; The first mapping feature subset and the second mapping feature subset are concatenated by dimension to form the observation dictionary matrix; The inner product space projection operation is performed on the discrete data points in the low-dimensional state sequence using the observation dictionary matrix, and the output is the high-dimensional state vector with a dimension parameter higher than that of the low-dimensional state sequence.

4. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S3, combining the high-dimensional state vector, the Koopman linear prediction equation is constructed using the identified high-dimensional state transfer matrix and the control input matrix, including: Retrieve the high-dimensional state transfer matrix and the control input matrix extracted offline using least-squares approximation based on the system's historical operation dataset; By performing matrix multiplication between the high-dimensional state transfer matrix and the high-dimensional state vector at the current sampling time, the autonomous evolution component characterizing the autonomous linear evolution of the system state is obtained. By performing matrix multiplication between the control input matrix and the control input vector at the current sampling time, the external intervention component characterizing the external control's intervention on the system state is obtained; The Koopman linear prediction equation for predicting the state evolution result at the next discrete time step is constructed by performing a linear algebraic superposition of the autonomous evolution component and the external intervention component.

5. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S3, obtaining the predicted shear stress vector of the target structural member through the observation mapping matrix includes: Extract the observation mapping matrix that was previously generated based on the three-dimensional finite element stiffness model calibration of the target structural component; Extract the high-dimensional predicted state vector for the next discrete time step, derived from the Koopman linear prediction equation; The observation mapping matrix is ​​used to perform a dimension reduction linear mapping transformation operation on the high-dimensional predicted state vector, and the high-dimensional predicted state vector located in the high-dimensional Hilbert function space is inversely transformed into the physical three-dimensional task space, and the predicted shear stress vector corresponding to each stress node of the target structural component in the next discrete time step is extracted.

6. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S4, calculating the difference between the safe yield stress threshold and the predicted shear stress vector to construct the control barrier function includes: Extract the safe yield stress threshold pre-calibrated according to the allowable stress limit physical properties of the target structural component material; The safe yield stress threshold is set as the upper limit boundary of the system safety. The subtraction operation is performed between each predicted shear stress component in the predicted shear stress vector and the safe yield stress threshold to obtain the stress margin vector corresponding to each stress node. Based on the non-negative characteristics of each data element in the stress margin vector, a physical safety set for which the target system does not undergo plastic deformation is defined, and the spatial boundary equation enclosing the physical safety set is defined as the control barrier function under the boundary of the continuous domain.

7. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 6, characterized in that, In S4, the introduction of convergence coefficients to reconstruct it into a recursive inequality, generating the hard-constrained affine boundary equations for the restricted control input vector, includes: Introduce the convergence coefficient that determines the rate at which the system state approaches the boundary of the physical security set; The recursive inequality that embodies the forward invariance characteristic is constructed using the current barrier function value, the barrier function value at the next prediction time, and the convergence coefficient. Substitute the control input variables in the Koopman linear prediction equation into the recursive inequality and expand it to separate the linear correlation term and system constant term containing the control input variables. Reconstruct the linear inequality for the control input vector and use the linear inequality as the hard constraint affine boundary equation.

8. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S5, the construction of the quadratic cost function by combining trajectory tracking deviation and control increment penalty includes: Extract the set reference demolding displacement trajectory sequence, compare the difference between the reference demolding displacement trajectory sequence and the actual displacement state value fed back by the multi-source sensing module, and generate the trajectory tracking deviation; Extract the control input change rate that reflects the smoothness of the action of each servo hydraulic cylinder, quantify the control input change rate in combination with the set positive definite weight matrix, and generate the control increment penalty; The quadratic cost function is constructed by performing a weighted summation operation on the quadratic term of the trajectory tracking deviation and the quadratic term of the control increment penalty, in the form of a standard quadratic function.

9. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S5, the hard-constrained affine boundary equation is used as a hard constraint condition for convex optimization to obtain the optimal control vector, including: The constructed quadratic cost function is set as the minimum objective function of the convex optimization algorithm; The hard-constrained affine boundary equation is set as an absolute physical boundary condition that cannot be violated during the optimization iteration process of the convex optimization algorithm. The interior point method is used to perform an iterative search operation in the feasible control input space that satisfies the absolute physical boundary conditions, calculate the control input solution that minimizes the quadratic cost function, and establish the control input solution as the optimal control vector.

10. The multi-source sensing method for preventing uneven stress during cylinder clamping according to claim 1, characterized in that, In S6, the optimal control vector is sent to each servo hydraulic cylinder for execution, and the low-dimensional state sequence acquisition operation is re-triggered after the action is completed, including: The digital commands in the optimal control vector are converted into matching analog voltage drive signals; The analog voltage drive signal is transmitted in parallel to the servo proportional valve control terminal inside each servo hydraulic cylinder, driving each servo hydraulic cylinder to perform the associated physical displacement action according to the analog voltage command. At the end of the control cycle of each servo hydraulic cylinder completing a single drive command, a trigger interrupt signal is generated. The trigger interrupt signal is used to wake up the data sampling array of the multi-source sensing module and force the entry into the physical parameter extraction and low-dimensional state sequence construction process of the next discrete time step.

Images