An adaptive control parameter optimization method for a multi-field industrial control system

Abstract

The application discloses a kind of adaptive control parameter optimization methods for multi-field industrial control system, belong to industrial control and intelligent control technical field;By collecting the multi-source heterogeneous time series data in multi-field industrial control system, a unified system state representation is constructed;The non-invasive virtual state evolution of multiple candidate control parameters is deduced, and the corresponding system response trajectory is generated;Introduce the preferred parameter verification mechanism based on physical residual and parameter risk potential, and screen out the parameter combination that violates physical constraints or has high risk;Based on spectral entropy and spectral geometric center, the frequency domain topology gating analysis is carried out, the stability and safety of the control parameters are constrained in frequency domain, and the application of the control parameters is triggered only when the frequency domain stability condition is met. Realize the safe, robust adaptive optimization of multi-field industrial control system control parameters without interfering with the existing production process, improve the reliability and engineering practicability of complex industrial control system parameter setting.

Description

Technical Field

[0001] This invention relates to the field of industrial control and intelligent control technology, specifically to an adaptive control parameter optimization method for multi-domain industrial control systems. Background Technology

[0002] An adaptive control parameter optimization method for multi-domain industrial control systems refers to a method that, during the operation of an industrial control system, for subsystems simultaneously containing multiple physical mechanisms and dynamic characteristics, optimizes control parameters dynamically by uniformly modeling and analyzing the system's operating state without directly interfering with the actual control loop. This method typically combines system state perception, parameter response prediction, constraint verification, and stability analysis to achieve adaptive adjustment of control parameters according to changes in operating conditions. By applying this method, the consistency and robustness of control performance can be improved while ensuring system safety and stability, reducing the cost of manual parameter tuning, and enhancing the reliability and intelligence level of multi-domain industrial control systems under complex operating conditions and long-term operation.

[0003] However, in existing industrial control systems, the tuning of control parameters typically relies on manual experience, offline debugging, or optimization methods based on a single subsystem model. On the one hand, manual or semi-automatic tuning methods are inefficient and difficult to adapt to changes in operating conditions; on the other hand, some data-driven adaptive parameter optimization methods often require online perturbation trial and error in the actual control loop, which can easily interfere with the production process and even cause system instability. In addition, existing methods usually focus on optimizing time-domain control errors, lacking a comprehensive assessment of the long-term evolution risks of the system after parameter adjustment, violations of physical constraints, and frequency domain stability issues, making it difficult to meet the safety, reliability, and continuous operation requirements of industrial control systems in multiple fields.

[0004] Specifically, existing adaptive control parameter optimization methods for industrial control systems have technical problems such as reliance on online disturbance trial and error, intrusion of parameter adjustment process into production control loop, difficulty in unified state representation when crossing thermodynamic subsystems, motion drive subsystems and fluid hydraulic subsystems, and lack of ability to predict the long-term evolution risk of the system after parameter adjustment.

[0005] Existing methods for verifying optimal parameters have technical problems. They rely solely on a single control error or short-term performance index for parameter selection, ignoring the physical constraints of the system. This results in parameters that are theoretically optimal but cause energy anomalies, actuator overloads, or control instability in actual systems.

[0006] Existing frequency domain topology gating analysis methods suffer from technical problems, such as relying solely on a single spectral amplitude threshold or fixed bandwidth limitation, making it difficult to fully reflect changes in the frequency domain structure of the control signal, and failing to effectively identify high-frequency resonances or potential instability modes introduced by parameter adjustments. Summary of the Invention

[0007] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an adaptive control parameter optimization method for multi-domain industrial control systems. The technical solution adopted by this invention is as follows: This invention provides an adaptive control parameter optimization method for multi-domain industrial control systems, which includes the following steps:

[0008] Step S1: Reconstructing the features of heterogeneous data manifold;

[0009] Step S2: Evolution of virtual state of control parameters;

[0010] Step S3: Validation of optimal parameters based on physical constraints and parameter deviation cost modeling;

[0011] Step S4: Frequency domain topology gating analysis.

[0012] Further, in step S1, the heterogeneous data manifold feature reconstruction is used to collect and eliminate the differences in different physical dimensions in multi-domain industrial control systems and establish a unified system state feature vector. Specifically, it involves collecting multi-source heterogeneous time-series data from multi-domain industrial control systems, calculating the geodesic distance and energy distribution of the data in high-dimensional space, constructing system manifold features, and obtaining a holographic manifold feature vector characterizing the current operating state of the system.

[0013] The reconstruction of heterogeneous data manifold features specifically includes the following steps:

[0014] Step S11: Data acquisition alignment, specifically, parallel acquisition of raw time-series data of temperature, servo position, drive current and fluid pressure, and frequency alignment processing of data from each channel using a unified hardware timestamp to obtain a time-domain synchronized multidimensional raw state vector.

[0015] Step S12: Energy distribution construction, specifically, setting a fixed-length sliding time window, extracting a segment of the multidimensional original state vector within the window, and calculating its sample covariance matrix, where the diagonal elements of the matrix represent the autocorrelation energy of each subsystem, and the off-diagonal elements represent the cross-correlation coupling strength between different subsystems, thus obtaining the local covariance matrix that represents the local dynamic characteristics of the system.

[0016] Step S13: Manifold feature calculation, specifically, calculating the geodesic distance between the local covariance matrix and the steady-state reference covariance matrix in the Riemannian manifold space at the current moment, and concatenating the geodesic distance with the principal eigenvalues ​​of the local covariance matrix to obtain a holographic manifold feature vector containing system geometric distance information and energy principal component information.

[0017] Further, in step S2, the virtual state evolution of the control parameters is used to generate potential response trajectories for different combinations of control parameters in the future without intrusion. Specifically, it involves using the holographic manifold feature vector as the initial state to construct a time-series prediction model, loading multiple sets of candidate control parameters, and performing multi-step forward virtual iterative calculations to obtain a set of virtual state evolution trajectories corresponding to different candidate parameters, including the following steps:

[0018] Step S21: Parameter space sampling, used to determine the range of control parameters that need to be virtually tested and generate candidate samples. Specifically, a preset disturbance range is set with the control parameters of the current system operation as the center, and multiple sets of evenly distributed candidate control parameter combinations are generated within this range using the Latin hypercube sampling method to obtain the candidate parameter set to be evaluated.

[0019] Step S22: Heterogeneous feature tensor fusion, used to construct a fusion input tensor adapted to the prediction model dimension. Specifically, the holographic manifold feature vector is used as a dynamic channel, each set of static parameters in the candidate parameter set is expanded in the time dimension, and the two are spliced ​​and fused in the feature dimension to obtain a composite input feature tensor corresponding to each set of candidate parameters.

[0020] Step S23: Recursive temporal reasoning, used to perform recursive reasoning calculations for the future state of the system. Specifically, it calls a preset long short-term memory network model, inputs the composite input feature tensor into the model, and performs rolling forward prediction according to a preset time step to obtain the original prediction sequence containing the changing trends of the key physical quantities of the system.

[0021] Step S24: Trajectory reconstruction, used to generate standardized virtual response trajectory objects. Specifically, the original prediction sequence is denormalized and its physical dimensions are restored. The predicted values ​​of the controlled variables and the predicted values ​​of the actuator output are encapsulated into time-series trajectory objects to obtain the final set of virtual state evolution trajectories for subsequent verification.

[0022] Further, in step S3, the optimized parameter verification is used to select the best control parameters from two dimensions: physical feasibility and strategy optimality. Specifically, it employs a parameter optimization method based on residual screening and risk potential field, uses physical conservation equations to perform compliance verification on the set of virtual state evolution trajectories, eliminates trajectories that violate physical constraints, constructs a parameter deviation cost function, calculates the deviation between the compliant trajectory and the theoretical optimal target, selects the parameters corresponding to the trajectory with the smallest cumulative parameter deviation cost, and obtains the set of verification target control parameters, including the following steps:

[0023] Step S31: Physical compliance verification, specifically, constructing the Hamiltonian energy residual equation based on the energy conservation law of the multi-domain system, calculating the energy violation index of each trajectory in the set of virtual state evolution trajectories, and determining the trajectories whose index exceeds the preset physical tolerance threshold as non-physical trajectories and removing them, thus obtaining a set of valid trajectories that are physically compliant;

[0024] Step S32: Construct the parameter deviation potential energy and performance cost function. Specifically, construct a comprehensive cost function that includes a trajectory tracking error term and a parameter deviation potential energy term. The parameter deviation potential energy term uses a risk weighting matrix to calculate the Mahalanobis distance of the candidate parameter relative to the current operating parameter, thereby characterizing the potential risk cost of parameter adjustment and obtaining the comprehensive evaluation value corresponding to each effective trajectory.

[0025] Step S33: Strategy optimization, specifically, minimizing the comprehensive evaluation value in the set of effective trajectories, selecting the control parameter combination corresponding to the trajectory with the lowest cumulative cost, and obtaining the theoretically optimal parameter candidate;

[0026] Step S34: Verify the generation of the target control parameter set. Specifically, verify the non-emptiness of the effective trajectory set. When it is non-empty, encapsulate the theoretically optimal parameter candidates. When it is empty, directly encapsulate the current running parameters and trigger an anomaly flag to obtain the final verification target control parameter set.

[0027] Further, in step S4, the frequency domain topology gating analysis is used to perform security and stability interception on the target control parameter set at the signal frequency domain level. Specifically, it generates a pre-simulated control signal based on the target control parameter set, and uses an adaptive frequency domain security gating method based on spectral entropy topology features to perform frequency domain transformation on the pre-simulated control signal to extract spectral topology features, determine whether there are high-frequency resonances or unstable energy peaks, and generate a blocking signal if the safety threshold is exceeded; otherwise, it generates a conducting signal to obtain a frequency domain gating command with a security identifier, including the following steps:

[0028] Step S41: Short-time virtual pre-simulation signal generation, specifically, taking the current system state as the starting point, using the verification target control parameter set to perform short-time closed-loop iterative simulation on the underlying control loop, applying a preset step excitation signal, generating a pre-simulation control signal sequence containing the dynamic response characteristics of the new parameters, and obtaining a short-time time-domain pre-simulation sequence;

[0029] Step S42: Power spectral density function construction, specifically, windowing and fast Fourier transform are performed on the pre-simulated control signal sequence to calculate the power spectral density, and the power spectral density is normalized to satisfy the mathematical properties of the probability density function, thereby constructing a spectral probability density function that reflects the shape of the signal energy distribution in the frequency domain, and obtaining normalized frequency domain topological distribution data;

[0030] Step S43: Spectral entropy topological feature extraction, specifically, based on the spectral probability density function, the spectral entropy feature value is calculated using the information entropy formula to characterize the frequency domain order of the control signal; at the same time, the spectral geometric centroid feature value is calculated using the weighted average method to characterize the center position of the frequency domain distribution of the control energy, thus obtaining a frequency domain security feature vector composed of spectral entropy and spectral centroid.

[0031] Step S44: Dual-threshold safety topology gating, specifically, comparing the spectral entropy feature value with a preset system stability entropy threshold, and comparing the spectral geometric centroid feature value with a preset system bandwidth cutoff frequency. If either feature value exceeds the threshold, an instability risk is determined and a blocking signal is generated; if both feature values ​​are within the threshold range, a conduction signal is generated, resulting in a frequency domain gating instruction with a safety identifier. Based on the frequency domain gating instruction, a control parameter update or retention mechanism is triggered, so that the control parameters are written to the controller only when the frequency domain stability condition is met.

[0032] The beneficial effects achieved by the present invention using the above solution are as follows:

[0033] (1) In view of the technical problems in the existing adaptive control parameter optimization methods of industrial control systems, such as reliance on online disturbance trial and error, intrusion of parameter adjustment process into production control loop, difficulty in unified state representation when crossing thermodynamic subsystem, motion drive subsystem and fluid hydraulic subsystem, and lack of ability to predict the long-term evolution risk of the system after parameter adjustment, this solution creatively adopts a comprehensive intelligent multi-domain industrial control system parameter optimization method that combines virtual state evolution of control parameters, verification of optimized parameters and frequency domain topology gating analysis. By performing virtual forward deduction on multiple sets of candidate control parameters without actual intervention in the control loop, and completing the dual screening of physical compliance and frequency domain stability before writing the parameters, the parameter optimization process and the actual production control process are decoupled.

[0034] (2) In view of the technical problem that the existing parameter selection verification methods only select parameters based on a single control error or short-term performance index, ignoring the physical constraints of the system, resulting in parameters that are theoretically optimal but cause energy anomalies, actuator overload or control instability in the actual system, this solution creatively adopts a parameter selection method based on residual screening and risk potential energy field. It realizes that when evaluating the virtual state evolution trajectory of control parameters, a physical residual verification mechanism is introduced to eliminate "non-physical trajectories" that violate energy balance or power constraints, and further characterizes the adjustment risk of candidate parameters relative to the current operating parameters by parameter deviation potential energy term;

[0035] (3) In view of the technical problems in the existing frequency domain topology gating analysis methods, which rely solely on a single spectral amplitude threshold or fixed bandwidth limitation, making it difficult to fully reflect the changes in the frequency domain structure of the control signal and effectively identify high-frequency resonance or potential instability modes introduced by parameter adjustment, this scheme creatively adopts an adaptive frequency domain security gating method based on spectral entropy topology features. This method realizes an adaptive frequency domain security gating method based on spectral entropy topology features. By simultaneously introducing spectral entropy features and spectral geometric centroid features, the orderliness of the frequency domain energy distribution of the control signal and the center frequency position are jointly determined. Attached Figure Description

[0036] Figure 1 A flowchart illustrating an adaptive control parameter optimization method for multi-domain industrial control systems provided by this invention;

[0037] Figure 2 A flowchart illustrating the process of reconstructing the heterogeneous data manifold features in step S1;

[0038] Figure 3 This is a flowchart illustrating the evolution of the virtual state of the control parameters in step S2.

[0039] Figure 4 A flowchart illustrating the process of verifying the optimal parameters for step S3, which is based on physical constraints and parameter deviation costs.

[0040] Figure 5 This is a flowchart illustrating the frequency domain topology gating analysis in step S4.

[0041] 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. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0043] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0044] Example 1, see Figure 1 This invention provides an adaptive control parameter optimization method for multi-domain industrial control systems, which includes the following steps:

[0045] Step S1: Reconstructing the features of heterogeneous data manifold;

[0046] Step S2: Evolution of virtual state of control parameters;

[0047] Step S3: Validation of optimal parameters based on physical constraints and parameter deviation cost modeling;

[0048] Step S4: Frequency domain topology gating analysis.

[0049] By performing the above operations, this solution addresses the technical problems in existing adaptive control parameter optimization methods for industrial control systems, such as reliance on online disturbance trial and error, intrusion of parameter adjustment processes into production control loops, difficulty in unifying state representation across thermodynamic subsystems, motion drive subsystems, and fluid hydraulic subsystems, and lack of ability to predict long-term evolution risks of the system after parameter adjustment. This solution creatively adopts a comprehensive intelligent multi-domain industrial control system parameter optimization method that combines virtual state evolution of control parameters, optimal parameter verification, and frequency domain topology gating analysis. By performing virtual forward extrapolation on multiple sets of candidate control parameters without actual intervention in the control loop, and completing dual screening of physical compliance and frequency domain stability before writing the parameters, the parameter optimization process is decoupled from the actual production control process.

[0050] Taking the industrial control scenario of parallel operation of injection molding machine barrel temperature control and robotic arm servo drive as an example, this solution can predict the impact of different PID parameter combinations on temperature fluctuation and servo oscillation trend in advance without interrupting heating and motion control. Only parameters that meet stability and performance requirements are applied to the control loop, thereby significantly improving the safety and engineering feasibility of parameter adaptive optimization in multi-domain industrial control systems.

[0051] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the heterogeneous data manifold feature reconstruction is used to collect and eliminate the differences of different physical dimensions in multi-domain industrial control systems and establish a unified system state feature vector. Specifically, it involves collecting multi-source heterogeneous time-series data from multi-domain industrial control systems, calculating the geodesic distance and energy distribution of the data in high-dimensional space, constructing system manifold features, and obtaining a holographic manifold feature vector that characterizes the current operating state of the system.

[0052] In this embodiment, the multi-domain industrial control system specifically covers the following physical subsystems and their corresponding data acquisition objects:

[0053] Thermodynamic subsystems specifically refer to temperature-controlled objects with high inertia and slow response characteristics, such as the barrel heating section of an injection molding machine, a semiconductor diffusion furnace, or a reflow oven; the collected data includes real-time temperature, set temperature, and heating power output percentage;

[0054] Preferably, Table 1 is a sample table of original data of the thermodynamic subsystem taking the barrel temperature control system of a precision injection molding machine as an example. As shown in the table, by synchronously collecting data from the "nozzle section" and the "plasticizing section", the thermal conduction coupling effect between different temperature zones can be captured. The "ambient temperature" introduced as an independent dimension is intended to provide an external disturbance reference for subsequent manifold feature reconstruction, so that the algorithm can effectively eliminate the interference of ambient temperature change on the internal thermal balance equation of the system when calculating the covariance matrix, thereby establishing a more accurate thermodynamic steady-state model of the barrel heating process.

[0055] Table 1. Example of raw data for the thermodynamic subsystem of a precision injection molding machine barrel temperature control system;

[0056]

[0057] The motion drive subsystem specifically refers to servo or stepper driven objects with high-frequency response characteristics, such as robotic arm joints, CNC machine tool feed axes, or conveyor belts; the collected data includes: servo encoder position, real-time speed, bus voltage, q-axis current, and torque current;

[0058] Preferably, Table 2 is a sample table of original data of the motion drive subsystem taking the X-axis feed system of a five-axis CNC machine tool as an example. As shown in the table, the correlation analysis between "q-axis torque current" and "position following error" can directly reflect the torque output rigidity of the motor when overcoming the cutting load. At the same time, monitoring "DC bus voltage" is not only for recording energy consumption, but more importantly, it is used as an energy boundary constraint condition in the subsequent virtual state evolution to prevent the prediction model from generating false high-acceleration trajectories that exceed the power supply capacity of the driver, thus ensuring the full-dimensional state representation from the micro current loop to the macro position loop.

[0059] Table 2. Example of raw data for motion drive subsystem taking the X-axis feed system of a five-axis CNC machine tool as an example;

[0060]

[0061] The fluid hydraulic subsystem specifically refers to objects involving pressure and flow control, such as hydraulic pump stations, pneumatic clamps, or coolant circulation loops; the data collected includes: pipeline pressure, instantaneous flow rate, and valve opening.

[0062] Preferably, Table 3 is a sample table of original data for the fluid hydraulic subsystem, taking the injection circuit of a cold chamber die-casting machine as an example. As shown in the table, in hydraulic control, the "proportional valve spool opening" and "instantaneous flow rate" are not simply linearly related, but are deeply constrained by the pressure difference provided by the "system main oil circuit pressure" and the fluid viscosity change caused by the "hydraulic oil temperature". Therefore, this embodiment, by jointly collecting these five sets of physical quantities, enables the subsequent Riemann manifold measurement algorithm to map the viscosity drift caused by oil temperature into the geodesic distance change in the manifold space, thereby achieving adaptive correction of the pressure transmission characteristics of the hydraulic system under different working conditions, and solving the technical problem that single pressure data cannot accurately describe the system impedance characteristics;

[0063] Table 3. Sample original data of the fluid hydraulic subsystem, taking the injection circuit of a cold chamber die casting machine as an example;

[0064]

[0065] The reconstruction of heterogeneous data manifold features specifically includes the following steps:

[0066] Step S11: Data acquisition alignment, specifically, the sensor interfaces of the thermodynamic subsystem, motion drive subsystem and fluid dynamics subsystem are connected respectively, and the raw time-series data of temperature, servo position, drive current and fluid pressure are collected in parallel. The data of each channel are frequency aligned using a unified hardware timestamp to obtain a time-domain synchronized multidimensional raw state vector.

[0067] Step S12: Energy distribution construction, specifically, setting a fixed-length sliding time window, extracting a segment of the multidimensional original state vector within the window, and calculating its sample covariance matrix, where the diagonal elements of the matrix represent the autocorrelation energy of each subsystem, and the off-diagonal elements represent the cross-correlation coupling strength between different subsystems, thus obtaining the local covariance matrix that represents the local dynamic characteristics of the system.

[0068] Step S13: Manifold feature calculation, specifically based on the affine invariant Riemannian metric algorithm, calculates the geodesic distance between the local covariance matrix at the current moment and the steady-state reference covariance matrix in the Riemannian manifold space, and concatenates the geodesic distance with the principal eigenvalues ​​of the local covariance matrix to obtain a holographic manifold feature vector containing system geometric distance information and energy principal component information;

[0069] Preferably, this scheme uses Riemann geometric distance for calculation, and the calculation formula is as follows:

[0070] ;

[0071] In the formula, It is the geodesic distance. It is the steady-state baseline covariance matrix. It is a local covariance matrix, where k is the time index. This refers to the Frobenius norm operation on the matrix, where n is the dimension of the covariance matrix, equal to the total number of sensor acquisition channels, and i is the dimension index of the covariance matrix. It is a matrix pair Generalized eigenvalues;

[0072] Further preferably, in order to construct a holographic manifold feature vector containing system geometric distance information and energy principal component information, this step concatenates the geodesic distances calculated above with the principal eigenvalue vectors of the local covariance matrix to obtain the holographic manifold feature vector. The calculation formula is as follows:

[0073] ;

[0074] In the formula, It is the eigenvector of the holographic manifold. is the first principal eigenvalue of the covariance matrix, and m is the total number of principal eigenvalue vectors. It is the vector transpose operator.

[0075] Example 3, see Figure 1 and Figure 3This embodiment is based on the above embodiment. In step S2, the virtual state evolution of the control parameters is used to generate potential response trajectories for different combinations of control parameters in the future without intrusion. Specifically, it uses the holographic manifold feature vector as the initial state to construct a time-series prediction model, loads multiple sets of candidate control parameters, and performs multi-step forward virtual iterative calculations to obtain a set of virtual state evolution trajectories corresponding to different candidate parameters. The steps include:

[0076] Step S21: Parameter space sampling, used to determine the range of control parameters that need to be virtually tested and generate candidate samples. Specifically, a preset disturbance range is set with the control parameters of the current system operation as the center, and multiple sets of evenly distributed candidate control parameter combinations are generated within this range using the Latin hypercube sampling method to obtain the candidate parameter set to be evaluated.

[0077] Preferably, the candidate parameter set to be evaluated includes not only proportional coefficients, integral coefficients, and differential coefficients, but also feedforward gain coefficients; when setting the perturbation range, for highly sensitive differential coefficients, the perturbation range is limited to within ±10%, while for proportional coefficients, the perturbation range is set to ±20%, thereby constructing a non-uniform parameter search space and improving search efficiency.

[0078] More preferably, the specific number of samples for the Latin hypercube sampling is set between 30 and 100; a "minimum distance constraint" is introduced during the sampling process to ensure that the distance between any two generated candidate parameters in Euclidean space is greater than a preset threshold, thereby avoiding the generation of overly similar redundant parameter combinations and ensuring the sparsity and representativeness of the samples in the solution space.

[0079] To more intuitively illustrate the effect of parameter space sampling, Table 4 shows typical candidate parameter sampling examples generated based on the current baseline parameters. Taking the temperature control loop of a typical high-inertia thermal process industrial control object as an example (such as a precision injection molding equipment), the baseline PID parameter is K. p =12.0, K i =0.5, K d =2.5; as shown in the table, the differential coefficient K for high sensitivity is... d The algorithm limits its perturbation to a narrow range of ±10% (i.e., 2.25~2.75), while the scaling factor K... p This provides a wider space for exploration; this non-uniform sampling strategy avoids virtual high-frequency oscillations caused by excessively large differential terms;

[0080] Table 4. Examples of candidate control parameter sets based on non-uniform Latin hypercube sampling;

[0081]

[0082] Step S22: Heterogeneous feature tensor fusion, used to construct a fusion input tensor adapted to the prediction model dimension. Specifically, the holographic manifold feature vector is used as a dynamic channel, each set of static parameters in the candidate parameter set is expanded in the time dimension, and the two are spliced ​​and fused in the feature dimension to obtain a composite input feature tensor corresponding to each set of candidate parameters.

[0083] Preferably, the dimension construction of the composite input feature tensor follows the (B, T, D) format, where B is the batch size (corresponding to the number of candidate parameter groups), T is the time step, and D is the total feature dimension; the concatenation and fusion process specifically involves combining the (T, D) dimension tensors into a single array. manifold Using the holographic manifold feature vectors of dimension (1, D) as the base layer, the holographic manifold feature vectors of dimension (1, D) are used as the base layer. param) The static control parameters are copied and extended to (T, D) param The parameter matrix of ) is concatenated with the parameter matrix of ) along the feature dimension to obtain D = D manifold +D param The composite tensor;

[0084] More preferably, before splicing and fusion, the static control parameters are nonlinearly scaled based on the Sigmoid function and mapped to the (0,1) interval so that their numerical magnitude is consistent with the energy distribution characteristics in the holographic manifold feature vector, preventing the neural network gradient from vanishing or exploding due to excessive numerical differences.

[0085] Step S23: Recursive temporal reasoning, used to perform recursive reasoning calculations for the future state of the system. Specifically, it calls a preset long short-term memory network model, inputs the composite input feature tensor into the model, and performs rolling forward prediction according to a preset time step to obtain the original prediction sequence containing the changing trends of the key physical quantities of the system.

[0086] Preferably, the pre-built long short-term memory network model adopts a two-layer stacked architecture. The first LSTM layer contains 128 hidden units for extracting short-term fluctuation features; the second LSTM layer contains 64 hidden units for extracting long-term trend features; the model output layer is connected to a fully connected layer, and the activation function adopts a linear rectified unit or a hyperbolic tangent function to adapt to the value range of different physical quantities.

[0087] More preferably, the specific step size of the rolling forward prediction is set to 50 to 200 control cycles in the future; in the recursive reasoning process, real historical data is used to assist in correction in the first 10 steps of the prediction, and the subsequent steps rely entirely on the prediction output of the previous moment as the input of the next moment, so as to simulate the long-term evolution trend while ensuring short-term accuracy.

[0088] Step S24: Trajectory reconstruction, used to generate standardized virtual response trajectory objects, specifically involves inverse normalization and physical dimension restoration of the original prediction sequence, and encapsulating the predicted values ​​of the controlled variables and the predicted values ​​of the actuator output into a time-series trajectory object, to obtain the final set of virtual state evolution trajectories for subsequent verification.

[0089] Preferably, the standardization process includes performing destandardization on the original prediction sequence using the pre-stored mean and variance of the training set; the encapsulated time-series trajectory object is a structured data packet that contains not only the prediction curve of the controlled variable, but also the prediction curve of the executor output variable.

[0090] More preferably, when encapsulating the time-series trajectory object, the system automatically includes a confidence decay mask; the mask value decays exponentially with the increase of the prediction step size, and is used to reduce the weight of the long-term prediction result when calculating the cost function in the subsequent step S3, thereby reflecting the physical fact that the uncertainty of the prediction model increases over time.

[0091] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the preferred parameter verification is used to select the best control parameters from two dimensions: physical feasibility and strategy optimality. Specifically, it adopts a parameter selection method based on residual screening and risk potential field, uses physical conservation equations to perform compliance verification on the set of virtual state evolution trajectories, eliminates trajectories that violate physical constraints, constructs a parameter deviation cost function, calculates the deviation between the compliant trajectory and the theoretical optimal target, selects the parameters corresponding to the trajectory with the smallest cumulative parameter deviation cost, and obtains the set of control parameters for the verification target. This includes the following steps:

[0092] Step S31: Physical compliance verification, specifically, constructing the Hamiltonian energy residual equation based on the energy conservation law of the multi-domain system, calculating the energy violation index of each trajectory in the set of virtual state evolution trajectories, and determining the trajectories whose index exceeds the preset physical tolerance threshold as non-physical trajectories and removing them, thus obtaining a set of valid trajectories that are physically compliant;

[0093] Preferably, for any controlled physical subsystem, such as the heating element of an injection molding machine or a servo motor, a discretized residual is constructed, and the rate of change of the Hamiltonian function of the controlled physical subsystem is calculated using the following formula:

[0094] ;

[0095] ;

[0096] In the formula, It is the instantaneous energy residual at time k. It is the value of the Hamiltonian function. It is the discrete time step. It is the active power injected into the system at time k. It is the power dissipation of the system at time k. It is the energy violation index of the trajectory. It is the forward prediction of the horizon length;

[0097] Step S32: Construct the parameter deviation potential energy and performance cost function. Specifically, construct a comprehensive cost function that includes a trajectory tracking error term and a parameter deviation potential energy term. The parameter deviation potential energy term uses a risk weighting matrix to calculate the Mahalanobis distance of the candidate parameter relative to the current operating parameter, thereby characterizing the potential risk cost of parameter adjustment and obtaining the comprehensive evaluation value corresponding to each effective trajectory.

[0098] The formula for calculating the comprehensive cost function is as follows:

[0099] ;

[0100] In the formula, It is a comprehensive evaluation value of the candidate parameters. It is a performance weighting coefficient. It is the target value set at time k. It is a virtual predicted output value. It is a risk weighting coefficient. It is a candidate control parameter vector. It is the current running parameter vector. It is a risk-weighted matrix, specifically the inverse of the covariance matrix;

[0101] Step S33: Strategy optimization, specifically, minimizing the comprehensive evaluation value in the set of effective trajectories, selecting the control parameter combination corresponding to the trajectory with the lowest cumulative cost, and obtaining the theoretically optimal parameter candidate;

[0102] The formula for calculating the theoretically optimal candidate parameters is as follows:

[0103] ;

[0104] In the formula, These are theoretically optimal parameter candidates. It is traversing variables. It is a valid set of parameters for physical compliance;

[0105] Step S34: Verify the generation of the target control parameter set. Specifically, verify the non-emptiness of the effective trajectory set. When it is non-empty, encapsulate the theoretically optimal parameter candidates. When it is empty, directly encapsulate the current running parameters and trigger an anomaly flag to obtain the final verification target control parameter set.

[0106] By performing the above operations, this solution addresses the technical problem in existing parameter optimization verification methods that rely solely on single control errors or short-term performance indicators for parameter selection, neglecting system physical constraints. This leads to theoretically optimal parameters causing energy anomalies, actuator overloads, or control instability in actual systems. The solution creatively adopts a parameter optimization method based on residual screening and risk potential energy fields. This method introduces a physical residual verification mechanism when evaluating the virtual state evolution trajectory of control parameters, eliminating "non-physical trajectories" that violate energy balance or power constraints. Furthermore, it characterizes the adjustment risk of candidate parameters relative to the current operating parameters by deviating from the potential energy term.

[0107] For example, in the pressure control scenario of a hydraulic pump station, even if a certain set of parameters shows a small pressure tracking error in the simulation prediction, if the power residual of its corresponding virtual trajectory is significantly amplified, or if the parameter deviation is too large and the potential impact risk increases, the parameter combination will be automatically eliminated, thereby achieving robust and low-risk selection of control parameters while ensuring physical feasibility.

[0108] Example 5, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S4, the frequency domain topology gating analysis is used to perform security and stability interception on the target control parameter set at the signal frequency domain level. Specifically, it generates a pre-simulated control signal based on the target control parameter set, uses an adaptive frequency domain security gating method based on spectral entropy topology features to perform frequency domain transformation on the pre-simulated control signal to extract spectral topology features, determines whether there are high-frequency resonances or unstable energy peaks, and generates a blocking signal if the safety threshold is exceeded; otherwise, it generates a conducting signal to obtain a frequency domain gating command with a security identifier. The steps include:

[0109] Step S41: Short-time virtual pre-simulation signal generation, specifically, taking the current system state as the starting point, using the verification target control parameter set to perform short-time closed-loop iterative simulation on the underlying control loop, applying a preset step excitation signal, generating a pre-simulation control signal sequence containing the dynamic response characteristics of the new parameters, and obtaining a short-time time-domain pre-simulation sequence;

[0110] Step S42: Power spectral density function construction, specifically, windowing and fast Fourier transform are performed on the pre-simulated control signal sequence to calculate the power spectral density, and the power spectral density is normalized to satisfy the mathematical properties of the probability density function, thereby constructing a spectral probability density function that reflects the shape of the signal energy distribution in the frequency domain, and obtaining normalized frequency domain topological distribution data;

[0111] Preferably, the power spectral density function is constructed by first truncating the time-domain sequence using a Hamming window, and the calculation formula is as follows:

[0112] ;

[0113] In the formula, It is the preview signal sequence after windowing and truncation. It is the original short-time temporal pre-simulation sequence, where n is the discrete-time sampling point index and N is the sampling window length;

[0114] Next, a Fast Fourier Transform is performed on the windowed sequence, calculated using the following formula:

[0115] ;

[0116] In the formula, It is the power spectral density amplitude. These are the complex coefficients of the Fast Fourier Transform. It is a frequency index;

[0117] Next, the power spectral density is normalized across the entire frequency band to construct the spectral probability density function, calculated as follows:

[0118] ;

[0119] In the formula, It is the normalized spectral probability density. It is a normalized frequency index. It is the power spectral density amplitude;

[0120] Step S43: Spectral entropy topological feature extraction, specifically, based on the spectral probability density function, the spectral entropy feature value is calculated using the information entropy formula to characterize the frequency domain order of the control signal; at the same time, the spectral geometric centroid feature value is calculated using the weighted average method to characterize the center position of the frequency domain distribution of the control energy, thus obtaining a frequency domain security feature vector composed of spectral entropy and spectral centroid.

[0121] Preferably, the spectral entropy topological feature extraction specifically involves first calculating the spectral entropy feature value, using the following formula:

[0122] ;

[0123] In the formula, H se It is the spectral entropy eigenvalue;

[0124] Next, the geometric centroid eigenvalues ​​are calculated using the following formula:

[0125] ;

[0126] In the formula, F sc These are spectral geometric barycentric eigenvalues. It is the physical frequency;

[0127] Subsequently, based on the spectral entropy eigenvalues ​​and the geometric centroid eigenvalues,

[0128] ;

[0129] In the formula, It is a frequency domain security feature vector

[0130] Step S44: Dual-threshold safety topology gating, specifically, comparing the spectral entropy feature value with a preset system stability entropy threshold, and comparing the spectral geometric centroid feature value with a preset system bandwidth cutoff frequency. If either feature value exceeds the threshold, an instability risk is determined and a blocking signal is generated; if both feature values ​​are within the threshold range, a conduction signal is generated, resulting in a frequency domain gating instruction with a safety identifier. Based on the frequency domain gating instruction, a control parameter update or retention mechanism is triggered, so that the control parameters are written to the controller only when the frequency domain stability condition is met.

[0131] The formula for calculating the logical condition for generating the blocking signal is as follows:

[0132] ;

[0133] In the formula, This is the system stability entropy threshold, which can preferably be taken as 60% of the maximum theoretical entropy value. It is the system bandwidth cutoff frequency threshold, which can preferably be taken as 1.5 times the system closed-loop bandwidth;

[0134] When either excessively low spectral entropy or excessively high centroid frequency is detected, the system immediately determines that there is a risk of instability and generates a blocking signal; countermeasures are taken if and only if... When both conditions are met, a conduction signal is generated, allowing parameters to be written to the controller.

[0135] By performing the above operations, this solution addresses the technical problems of existing frequency domain topology gating analysis methods, which rely solely on a single spectral amplitude threshold or fixed bandwidth limitation, making it difficult to comprehensively reflect changes in the frequency domain structure of the control signal and effectively identify high-frequency resonances or potential instability modes introduced by parameter adjustments. This solution creatively employs an adaptive frequency domain security gating method based on spectral entropy topology features. By simultaneously introducing spectral entropy features and spectral geometric centroid features, the method jointly determines the orderliness of the frequency domain energy distribution of the control signal and the position of its center frequency.

[0136] Taking the feed axis control of CNC machine tools as an example, when candidate parameters cause the control signal energy to concentrate in the high-frequency band and cause abnormal changes in spectral entropy, even if the time domain response has not yet become obviously unstable, this scheme can still identify potential resonance risks in advance and generate blocking commands to prevent unsafe parameters from being applied to the control loop, thereby achieving forward-looking frequency domain safety constraints and stability assurance for the control parameter update process.

[0137] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0139] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An adaptive control parameter optimization method for multi-domain industrial control systems, characterized in that: The method includes the following steps: Step S1: Heterogeneous data manifold feature reconstruction. Collect multi-source heterogeneous time-series data from multiple domain industrial control systems, construct system manifold features, and obtain a holographic manifold feature vector representing the current operating state of the system. This includes the following steps: data acquisition and alignment, energy distribution construction, and manifold feature calculation. Step S2: Controlling the virtual state evolution of parameters. Using the holographic manifold feature vector as the initial state, a time-series prediction model is constructed, and multiple sets of candidate control parameters are loaded. Multi-step forward virtual iterative calculations are performed to obtain a set of virtual state evolution trajectories corresponding to different candidate parameters, including the following steps: Step S21: Parameter space sampling, specifically, setting a preset disturbance range centered on the control parameters of the current system operation, and using the Latin hypercube sampling method to generate multiple uniformly distributed candidate control parameter combinations within this range to obtain the candidate parameter set to be evaluated; Step S22: Heterogeneous feature tensor fusion, specifically, the holographic manifold feature vector is used as a dynamic channel, each set of static parameters in the candidate parameter set is expanded in the time dimension, and the two are spliced ​​and fused in the feature dimension to obtain a composite input feature tensor corresponding to each set of candidate parameters; Step S23: Recursive temporal reasoning, specifically, calling a preset long short-term memory network model, inputting the composite input feature tensor into the model, and performing rolling forward prediction according to a preset time step to obtain the original prediction sequence containing the changing trends of key physical quantities of the system; Step S24: Trajectory reconstruction, specifically, the original prediction sequence is denormalized and the physical dimensions are restored, and the predicted values ​​of the controlled variable and the predicted values ​​of the actuator output are encapsulated into a time-series trajectory object to obtain the final set of virtual state evolution trajectories for subsequent verification. Step S3: Parameter optimization verification based on physical constraints and parameter deviation cost modeling. A parameter optimization method based on residual screening and risk potential field is adopted. The physical conservation equation is used to perform compliance verification on the set of virtual state evolution trajectories, eliminating trajectories that violate physical constraints, constructing a parameter deviation cost function, calculating the deviation between the compliant trajectory and the theoretical optimal target, and selecting the parameters corresponding to the trajectory with the minimum cumulative parameter deviation cost to obtain the set of verification target control parameters. This includes the following steps: physical compliance verification, construction of parameter deviation potential and performance cost function, strategy optimization, and generation of the set of verification target control parameters. Step S4: Frequency domain topology gating analysis. Based on the target control parameter set, a pre-simulation control signal is generated. An adaptive frequency domain safety gating method based on spectral entropy topology features is used to transform the pre-simulation control signal in the frequency domain to extract spectral topology features. It is determined whether there are high-frequency resonances or unstable energy peaks. If the peaks exceed the safety threshold, a blocking signal is generated; otherwise, a conducting signal is generated, resulting in a frequency domain gating command with a safety identifier.

2. The adaptive control parameter optimization method for multi-domain industrial control systems according to claim 1, characterized in that: In step S3, the physical compliance verification specifically involves constructing a Hamiltonian energy residual equation based on the energy conservation law of a multi-domain system, calculating the energy violation index of each trajectory in the set of virtual state evolution trajectories, and identifying trajectories whose index exceeds a preset physical tolerance threshold as non-physical trajectories and removing them, thereby obtaining a set of valid trajectories that are physically compliant.

3. The adaptive control parameter optimization method for multi-domain industrial control systems according to claim 2, characterized in that: In step S3, the construction of the parameter deviation potential energy and performance cost function specifically involves constructing a comprehensive cost function that includes a trajectory tracking error term and a parameter deviation potential energy term. The parameter deviation potential energy term uses a risk weighting matrix to calculate the Mahalanobis distance of the candidate parameter relative to the current operating parameter, thereby characterizing the potential risk cost of parameter adjustment and obtaining the comprehensive evaluation value corresponding to each valid trajectory.

4. The adaptive control parameter optimization method for multi-domain industrial control systems according to claim 3, characterized in that: In step S3, the preferred strategy is to minimize the comprehensive evaluation value in the set of effective trajectories, select the control parameter combination corresponding to the trajectory with the lowest cumulative cost, and obtain the theoretically optimal parameter candidate. The generation of the verification target control parameter set specifically involves verifying the non-emptiness of the valid trajectory set. When the set is non-empty, the theoretically optimal parameter candidates are encapsulated; when the set is empty, the current running parameters are directly encapsulated and an anomaly flag is triggered, thus obtaining the final verification target control parameter set.

5. The adaptive control parameter optimization method for multi-domain industrial control systems according to claim 4, characterized in that: In step S4, the frequency domain topology gating analysis includes the following steps: Step S41: short-time virtual pre-simulation signal generation; Step S42: power spectral density function construction; Step S43: spectral entropy topology feature extraction; Step S44: dual-threshold secure topology gating.

6. The adaptive control parameter optimization method for multi-domain industrial control systems according to claim 5, characterized in that: In step S41, the short-time virtual pre-simulation signal generation specifically involves taking the current system state as the starting point, using the verification target control parameter set to perform short-time closed-loop iterative simulation on the underlying control loop, applying a preset step excitation signal, generating a pre-simulation control signal sequence containing the dynamic response characteristics of the new parameters, and obtaining a short-time time-domain pre-simulation sequence. In step S42, the construction of the power spectral density function specifically involves windowing and performing a fast Fourier transform on the pre-simulation control signal sequence, calculating the power spectral density, and normalizing the power spectral density to satisfy the mathematical properties of the probability density function, thereby constructing a spectral probability density function that reflects the shape of the signal energy distribution in the frequency domain, and obtaining normalized frequency domain topological distribution data.

7. The adaptive control parameter optimization method for multi-domain industrial control systems according to claim 6, characterized in that: In step S43, the spectral entropy topological feature extraction specifically involves calculating the spectral entropy feature value based on the spectral probability density function using the information entropy formula to characterize the frequency domain orderliness of the control signal; simultaneously, calculating the spectral geometric centroid feature value using the weighted average method to characterize the center position of the frequency domain distribution of the control energy, thereby obtaining a frequency domain security feature vector composed of spectral entropy and spectral centroid. In step S44, the dual-threshold security topology gating specifically involves comparing the spectral entropy feature value with a preset system stability entropy threshold and comparing the spectral geometric centroid feature value with a preset system bandwidth cutoff frequency. If either feature value exceeds the threshold, an instability risk is identified and a blocking signal is generated. If both feature values ​​are within the threshold range, a conduction signal is generated, resulting in a frequency domain gating instruction with a security identifier. Based on the frequency domain gating instruction, a control parameter update or retention mechanism is triggered, ensuring that the control parameters are written to the controller only when the frequency domain stability condition is met.

Images