A data-driven reference signal security conditioning method and system

Abstract

This invention provides a data-driven reference signal safety regulation method and system. The method includes: acquiring operational data of a closed-loop system during the controlled process at a fixed sampling period, preprocessing it to obtain a high-quality standardized dataset; defining physical constraints based on the physical limits of the controlled object and the safety requirements of the production process; establishing an explicit discrete-time state-space linear model describing the implicit dynamics of the system based on the high-quality standardized dataset and a subspace identification algorithm, and solving for the linear model parameters of the system; calculating the maximum allowable output set by combining the physical constraints and the linear model parameters, running a scalar reference regulator online, and solving for the optimal safety reference signal in real time; the underlying controller calculating specific actuator actuation signals based on the optimal safety reference signal to drive the controlled object; and introducing a rolling time-domain mechanism to drive the system time axis stepping, realizing an anti-limit-over-limit control closed loop. This invention achieves end-to-end anti-limit-over-limit safety control.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of industrial automation control and software engineering, and in particular to a data-driven method and system for safe adjustment of reference signals. Background Technology

[0002] In modern industrial process control, aerospace, automotive electronics, and robotics, closed-loop systems must strictly adhere to physical limitations, safety regulations, and operational boundaries during operation—that is, state constraints and control input constraints. Traditional proportional-integral-derivative (PID) control or state feedback control primarily focuses on stability and dynamic tracking performance during design, often making it difficult to directly handle hard constraint problems.

[0003] The Reference Governor (RG), a typical "add-on" control strategy, is based on the idea of ​​adding a nonlinear regulating element to the reference input of an existing closed-loop control system. This element ensures that the system's evolution trajectory in the prediction time domain remains within the safe constraint region by online truncation or modification of the external reference signal.

[0004] However, traditional reference regulators suffer from the following core technical challenges in practical industrial applications: (1) High dependence on accurate explicit mathematical models. Traditional reference controller design process requires obtaining accurate linear state-space models of the controlled object and the underlying controller. (Matrix). In complex real-world industrial scenarios, due to the influence of nonlinear factors, time-varying parameters, and environmental disturbances, establishing a high-precision global linear model is extremely costly and cumbersome, and model mismatch often directly leads to constraint violations.

[0005] (2) The offline computation threshold for system security domains and infeasible sets is extremely high. The core of traditional RG theory lies in the computation of the maximum admissible set (MAS) or infeasible set. The derivation of these sets involves high-dimensional polyhedral projection, Minkowski summation, and finiteness processing of infinite-dimensional constraints. This requires algorithm implementers to have a very deep background in control theory and mathematical optimization, which makes it difficult for this technology to be popularized among front-line engineers who lack professional theoretical support, greatly limiting the pace of its industrialization.

[0006] (3) Lack of end-to-end integrated automatic generation tools. Currently, most industrial control software on the market focuses on local parameter tuning or system identification, and there is a lack of an integrated tool that can start directly from the underlying raw operating data and run through the entire process from model extraction, boundary calculation, to regulator parameter design. Frontline engineers often have to deal with complex algorithm logic and perform fragmented secondary development between different mathematical tools, which greatly increases the difficulty of engineering deployment and makes it impossible to achieve end-to-end automated delivery of "raw data input - safety control algorithm output". Summary of the Invention

[0007] In view of this, the present invention provides a data-driven reference signal security adjustment method and system to solve the above problems.

[0008] This invention provides a data-driven reference signal safety regulation method, comprising: acquiring operational data of a closed-loop system during the controlled process at a fixed sampling period and performing data preprocessing to obtain a high-quality standardized dataset; defining physical constraints based on the physical limits of the controlled object and the safety requirements of the production process; establishing an explicit discrete-time state-space linear model describing the implicit dynamics of the system based on the high-quality standardized dataset and a subspace identification algorithm, and solving for the linear model parameters of the system; calculating the maximum allowable output set by combining the physical constraints and the linear model parameters, and running a scalar reference regulator online to obtain the optimal safety reference signal in real time; the underlying controller calculating specific actuator actuation signals based on the optimal safety reference signal to drive the controlled object; and introducing a rolling time-domain mechanism to drive the system time axis stepping to achieve a closed-loop control system to prevent exceeding limits.

[0009] In another implementation of the present invention, the running data includes an input sequence and an output sequence; the data preprocessing includes: calculating the mean of the input sequence and the output sequence respectively, and performing mean-reduction processing to obtain a zero-mean fluctuation sequence; and using a low-pass digital filter to filter the zero-mean fluctuation sequence to eliminate high-frequency measurement noise.

[0010] In another implementation of the present invention, the physical constraints include control input constraints and system output constraints, wherein the control input constraints are expressed as follows:

[0011] in, and These represent the lower and upper limit vectors of the actuator, respectively; For the input vector, express A real vector space of dimensionless numbers, i.e. for 3D real vector This represents the total number of input channels.

[0012] The system output constraints are expressed as follows:

[0013] in, and This is a preset safety boundary vector; For the output vector, express A real vector space of dimensionless numbers, i.e. for 3D real vector This represents the total number of constrained output channels.

[0014] In another implementation of the present invention, the dynamic equations of the explicit discrete-time state-space linear model are expressed as:

[0015]

[0016] in, Let be the internal state vector of the system. The system order; These are the system matrix to be identified, the input matrix, the output matrix, and the direct transfer matrix, respectively.

[0017] In another implementation of the present invention, the maximum allowable output set is represented as:

[0018] in, and Reference instructions after constant modification Predicting future evolutionary trajectories under the influence of these factors; These are physical constraints.

[0019] In another implementation of the present invention, the optimal safety reference signal is represented as:

[0020] in, It is the globally optimal scalar factor; The instruction from the previous moment; This is the original expected reference instruction.

[0021] In another implementation of the present invention, the globally optimal scalar factor is expressed as:

[0022] in, Let be the total number of rows in the polyhedral inequalities. For any of the inequalities in the system of inequalities Row index; The first element in the system's real-time dynamic gain vector at the current moment. Each element component; Let be the th element in the system's real-time safety margin vector at the current moment. Each element component.

[0023] In another aspect, the present invention provides a data-driven reference signal safety control system, comprising: a data acquisition module: acquiring operational data of the controlled process of the closed-loop system at a fixed sampling period and performing data preprocessing to obtain a high-quality standardized dataset; a constraint construction module: defining physical constraints based on the physical limits of the controlled object and the safety requirements of the production process; a parameter solving module: establishing an explicit discrete-time state-space linear model describing the implicit dynamics of the system based on the high-quality standardized dataset and a subspace identification algorithm, and solving for the linear model parameters of the system; an optimal signal calculation module: calculating the maximum allowable output set by combining the physical constraints and the linear model parameters, and running a scalar reference controller online to obtain the optimal safety reference signal in real time; a driving module: the underlying controller calculates the specific actuator actuation signal based on the optimal safety reference signal to drive the controlled object; and a rolling optimization module: introducing a rolling time-domain mechanism to drive the system time axis stepping, realizing a closed-loop control system to prevent exceeding limits.

[0024] In another aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a data-driven reference signal security adjustment method as described in any of the preceding claims. In another aspect, the present invention provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of a data-driven reference signal security adjustment method as described in any of the preceding claims.

[0025] The data-driven reference signal security modulation method of the present invention has the following beneficial effects: (1) Pure data-driven and precise error quantification, completely freeing us from the constraints of physical mechanism modeling. Addressing the pain point that traditional algorithms heavily rely on precise explicit mathematical models and are prone to constraint violations due to model mismatch, this invention innovatively introduces Subspace Identification (SID) technology. This allows for the fully automatic extraction of the state space matrix using only the raw input and output data from the closed-loop system operation. Simultaneously, the system automatically traces and extracts the maximum single-step posterior residual as the extremum boundary of the perturbation at the underlying level, accurately quantifying unmodeled dynamics and environmental disturbances. This fundamentally eliminates the tedious derivation of physical equations and the bias in mechanistic assumptions, significantly reducing the modeling cost in complex industrial scenarios.

[0026] (2) Implicitly resolves high-dimensional set operations, overcoming the dual barriers of secure domain computation and online deployment. Addressing the pain point that traditional reference regulators (RGs) involve complex polyhedral projections and Pontryagin differences in calculating the maximum allowable set (MAS), resulting in a high barrier to engineering applications, this invention ingeniously transforms abstract topological set operations and infinite time-domain predictions into constant matrix iterations and series summations that are easily processed by low-level computers during the offline phase. During the online operation phase, a scalar reference regulator (SRG) is employed to drastically reduce the multidimensional optimization problem to analytical ratio calculations of one-dimensional scalar factors, completely eliminating the iterative optimization process. This design not only eliminates the requirement for a deep mathematical background for algorithm implementers but also compresses the single-step limit-breaking computation overhead to the microsecond level, perfectly supporting high-frequency real-time deployment on low-computing-power PLCs / DSPs.

[0027] (3) Construct robust boundary protection to achieve integrated industrial software delivery throughout the entire process. Addressing the current pain point of lacking end-to-end integrated generation tools, this invention deeply integrates quantized identification errors and unknown perturbations into set construction. By implementing rigorous "Robust Constraint Tightening," it generates a robust maximum output tolerance set (Robust MOAS) capable of withstanding the most severe perturbations, providing the system with mathematically proven "absolute anti-boundary protection." The overall architecture establishes an automated closed-loop chain from "offline data import and model generation" to "robust boundary automatic reconstruction" and then to "online security command rapid interception." Frontline engineers can achieve one-stop secure delivery without having to perform fragmented secondary code development across different platforms, completely filling the gap in the practical application of advanced constraint control algorithms in industrial automation. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention. In the accompanying drawings: Figure 1 This is a schematic flowchart of a data-driven reference signal security adjustment method according to an embodiment of the present invention.

[0029] Figure 2 This is an end-to-end architecture diagram of a data-driven reference signal security conditioning system according to an embodiment of the present invention.

[0030] Figure 3This is a schematic diagram illustrating the automated determination of system order based on SVD, according to an embodiment of the present invention.

[0031] Figure 4 This is a schematic diagram of robust constraint tightening geometry according to an embodiment of the present invention.

[0032] Figure 5 This is a geometrical schematic diagram of a scalar reference regulator for one-dimensional dimensionality reduction optimization according to an embodiment of the present invention.

[0033] Figure 6 This is a diagram of a rolling time-domain closed-loop control structure with a state observer, according to an embodiment of the present invention.

[0034] Figure 7 This is a closed-loop response curve of a system using the method of the present invention, according to an embodiment of the present invention.

[0035] Figure 8 This is a closed-loop response curve of a control group system that does not use the method of the present invention, according to an embodiment of the present invention.

[0036] Figure 9 This is a pseudocode diagram of a subspace identification algorithm according to an embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly 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 should fall within the protection scope of the present invention.

[0038] Figure 1 This is a schematic flowchart of a data-driven reference signal security modulation method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this embodiment mainly includes: S1. Collect the operating data of the closed-loop system during the controlled process at a fixed sampling period, and perform data preprocessing to obtain a high-quality standardized dataset.

[0039] For example, during the normal operation or engineering commissioning phase of the closed-loop system, a fixed sampling period is used. Discretized data sampling is performed on the controlled object to collect input and output data during the controlled process of the system. To ensure that the collected data can fully excite all dynamic modes of the system (i.e., satisfy the rank condition for subspace identification), a persistently exciting (PE) probe signal (such as a pseudo-random binary sequence PRBS or a multi-sine wave signal) is superimposed on the external reference command.

[0040] Define the actual input data sequence of the system as The actual output data sequence is ,in , Let be the total length of the collected data. To eliminate the impact of deterministic trend terms and high-frequency measurement noise in the industrial field data on the accuracy of subsequent model identification, preprocessing operations are performed on the original data sequence to obtain a high-quality standardized dataset, denoted as . and .

[0041] S2. Define physical constraints based on the physical limits of the controlled object and the safety requirements of the production process.

[0042] For example, in the software's interactive configuration interface, the operator defines the hard constraints during system operation based on the physical limits of the controlled object and the safety requirements of the production process.

[0043] S3. Based on high-quality standardized datasets and subspace identification algorithms, establish an explicit discrete-time state-space linear model describing the implicit dynamics of the system, and solve for the linear model parameters of the system.

[0044] For example, utilizing the obtained high-quality standardized dataset and Based on the Subspace Identification (SID) method, an explicit discrete-time state-space linear model describing the implicit dynamics of the system is automatically established.

[0045] S4. Combining physical constraints and linear model parameters, calculate the maximum allowable output set, and run the scalar reference regulator online to obtain the optimal safety reference signal in real time.

[0046] For example, a safe reference law is automatically generated based on the identified linear model parameters. and the input physical constraints The system calculates the Maximum Output Admissible Set (MOAS) offline or online. ).

[0047] S5. The underlying controller calculates the specific actuator actuation signal based on the optimal safety reference signal to drive the controlled object.

[0048] For example, the optimal safety reference signal obtained from real-time solution is extracted. This is then used as the actual execution instruction and sent to the underlying closed-loop controller. The underlying controller then... Calculate the specific actuator actuation signal to drive the physical controlled object.

[0049] S6. Introduce a rolling time domain mechanism to drive the system time axis stepping and realize the closed loop of anti-limit overrun control.

[0050] For example, at the end of the control cycle, the system's measured input and output data repository is updated, with the time step changed from [previous step]. Advance to The state observer absorbs the latest measurement data to complete the state vector. The system updates its time and measurements, then re-enters steps S2 to S6 to perform a new round of online optimization evaluation. This RecedingHorizon mechanism ensures the system's continuous adaptability to dynamic disturbances and provides overall security.

[0051] like Figure 2 As shown, the system architecture is mainly divided into two parts: the offline generation layer and the online execution layer. In the offline generation layer, the system treats it as a "black box" encapsulated module, directly receiving actual running data and automatically executing data preprocessing (S1), subspace identification (S3), and the construction of the offline robust maximum output tolerance set (MOAS) (S4) in sequence, finally outputting high-fidelity system parameters. and robust static boundary At the online execution layer, the system distributes the offline-generated parameters and boundaries, and intercepts the original expected instructions through a pre-reference regulator (supervisory filter). The closed-loop system state is fed back by a state observer. Real-time online optimization and truncation to generate absolutely safe correction instructions Then, the data is sent to the closed-loop controller to drive the controlled object.

[0052] In another implementation of the present invention, the running data includes an input sequence and an output sequence; the data preprocessing includes: calculating the mean of the input sequence and the output sequence respectively, and performing mean-centering processing to obtain a zero-mean fluctuation sequence; and using a low-pass digital filter to filter the zero-mean fluctuation sequence to eliminate high-frequency measurement noise.

[0053] For example, a discrete time step is defined. The system in the The actual multidimensional control input vector of the step is denoted as The actual multidimensional sensor measurement output vector is denoted as... The software system constructs the raw input data matrix from continuously acquired time series data. Compared with the original output data matrix :

[0054]

[0055] Among them, the total sampling length The continuous excitation order must satisfy the requirement that the corresponding input Hankel matrix has full rank.

[0056] Raw data collected in industrial settings typically includes a DC bias representing the system's operating point. To extract fluctuations that reflect the system's true linear incremental dynamics, the software automatically performs mean-reduction on the raw matrix.

[0057] First, calculate the entire data window. Time-averaged vectors of internal input and output signals and :

[0058] Subsequently, the mean vector is subtracted from each sampling point in the original sequence, and the sequence is then translated into a state-space coordinate system with zero as the equilibrium point, resulting in a zero-mean fluctuation sequence. and :

[0059]

[0060] Industrial sensor measurement output High-frequency measurement noise is often inevitably mixed in, which will affect the identified system matrix. This results in biased estimates. To address this, the system incorporates a low-pass digital filter to smooth the fluctuation sequence.

[0061] To avoid the phase shift introduced by traditional unidirectional digital filtering (such as standard IIR or FIR filters), which disrupts the crucial input-output causal relationship and time-phase mapping in the control system, this system employs zero-phase digital filtering technology. The difference equation for the low-pass filter is defined as follows:

[0062] in These are the filter coefficients. The zero-phase-shift filtering algorithm first processes the sequence along the positive time direction. The above filtering process yields the intermediate sequence. Then Reverse the time axis, pass the signal through the same filter again, and finally reverse the result again to output the final filtered signal. This process mathematically achieves strictly zero-phase distortion:

[0063] For the input sequence The same zero-phase-shift low-pass filtering process is also performed synchronously to obtain the denoised input signal. .

[0064] After mean elimination and zero-phase-shift filtering, high-quality time-series data that is pure and accurately characterizes the linear dynamics of the system near the equilibrium point is extracted. The column vectors from each time step are concatenated in chronological order and finally packaged to generate a standardized dataset matrix for subspace identification. and The aforementioned standard dataset will be directly transmitted to the memory-resident identification module as the data basis for model identification.

[0065] In another implementation of the invention, the physical constraints include control input constraints and system output constraints, which are standardized to the mathematical form of a convex polyhedron.

[0066] For the physical limits of actuators (such as valve opening, motor torque, pump speed, etc.), define an input vector. Permissible range The control input constraints are expressed as:

[0067] in, and These represent the lower and upper limit vectors of the actuator, respectively; For the input vector, express A real vector space of dimensionless numbers, i.e. for 3D real vector This represents the total number of input channels. During the calculation, this constraint is further generalized to a linear inequality matrix form: .

[0068] Define the system output vector for various safety indicators of the controlled process (such as pressure upper limit, liquid level safety range, temperature change range, etc.). Permissible set The system output constraints are expressed as:

[0069] in, and This is a preset safety boundary vector; For the output vector, express A real vector space of dimensionless numbers, i.e. for 3D real vector This represents the total number of constrained output channels. In the underlying software algorithm, this set is represented by a linear mapping as: .

[0070] For example, the physical limit parameters of the underlying actuators and the safety process thresholds of the controlled process can be read by the operator through the software's human-machine interface (HMI) or configuration file.

[0071] Define the system control input vector as ,in Let be the total number of input channels; define the system output vector as . ,in This represents the total number of constrained output channels. The read bilateral physical constraints are initialized to constant boundary vectors: Obtain the absolute lower limit vector of the control input With absolute upper limit vector .

[0072] Obtain the safe lower bound vector output by the system. With safety upper bound vector .

[0073] The aforementioned boundary vectors constitute the original operational box-constraints of the system in the multidimensional Cartesian coordinate system.

[0074] In order to transform the original upper and lower bound intervals into the form of half-space intersections that can be directly processed by the subsequent solver, a matrix reconstruction is performed on the control input boundary.

[0075] definition identity matrix of order Construct the weight coefficient matrix of the control input constraints. With the corresponding constraint boundary vector Its algebraic structure is configured as follows:

[0076] Therefore, control input vector Permissible range It is rigorously mathematically formalized and generalized into the following linear inequality space:

[0077] Similarly, for various output safety indicators of the controlled process, the software maps them as the internal region of a closed convex polyhedron enclosed by multidimensional hyperplanes.

[0078] definition identity matrix of order Construct the weight coefficient matrix of the system output constraints. With the corresponding constraint boundary vector Its algebraic structure is configured as follows:

[0079] Therefore, the system output vector Permissible set It is rigorously mathematically formalized and generalized into the following linear inequality space:

[0080] To achieve end-to-end automated computation, the software will define independently independent input constraints. With output constraints By performing normalization and splicing, a generalized joint constraint space encompassing all physical constraints is constructed. .

[0081] Define joint constraint variables Construct the joint constraint weight matrix for diagonal partitioning. Joint boundary vector with splicing :

[0082] Ultimately, the system's global security requirements are encapsulated into a unified, compact inequality form:

[0083] The result of this normalization encapsulation will be used as an absolute criterion and directly input into the Maximum Output Allowable Set (MOAS) automated calculation engine to ensure that the generated scalar reference regulator (SRG) achieves a strict isomorphic mapping with the actual physical limit in mathematical logic.

[0084] Through the above parameterized configuration, the software automatically abstracts complex physical conditions into unified mathematical constraints. This standardized configuration process provides clear boundary criteria for the subsequent automated iterative solution of the Maximum Allowable Output Set (MOAS), ensuring that the generated reference instruction law is strictly aligned with the actual physical constraints in mathematical logic.

[0085] In another implementation of the present invention, the dynamic equations of the explicit discrete-time state-space linear model are expressed as:

[0086]

[0087] in, Let be the internal state vector of the system. The system order; These are the system matrix to be identified, the input matrix, the output matrix, and the direct transfer matrix, respectively.

[0088] For example, such as Figure 9 As shown, the subspace identification algorithm automatically constructs an explicit linear state-space prediction model of the system without presetting physical parameters.

[0089] The specific calculation process is as follows: First, set the time window length. Construct the past input Hankel matrix Future input Hankel matrix Past output Hankel matrix And the future output Hankel matrix .

[0090] Secondly, define the instrument variable matrix. The projection matrix of the future output along the future input space from the past data space is calculated using oblique projection. :

[0091] For projection matrix Singular Value Decomposition (SVD) using a weighted matrix:

[0092] Based on the truncated principal singular value Determine the order of the system Based on this, the extended observability matrix is ​​extracted. Finally, using The shift invariance property is solved using the least squares method to obtain the linear state-space evolution matrix of the system. This step enables a seamless, automated transformation from raw data to an explicit predictive model.

[0093] To meet the spatial geometric mapping requirements of subsequent subspace double oblique projection operations, the underlying algorithm uses the one-dimensional discrete time series dataset (normalized input) obtained in step S1. With output The structure is reorganized into a high-dimensional block Hankel matrix, and its translation feature matrix for state temporal recursion is extracted simultaneously.

[0094] First, set the time sliding window parameters for subspace identification. This parameter is configured by the system to be strictly greater than the maximum physical order estimated for the controlled object. (Based on the total sampling length of the dataset.) Calculate and determine the uniform number of effective columns for all constructed matrices. Its algebraic constraint relationship is:

[0095] For dimensions System input data sequence Construction time steps cover to Generalized past input matrix and time steps covered to Generalized future input matrix Its precise multidimensional algebraic arrangement rules are as follows:

[0096]

[0097] Similarly, for dimensions... System output data sequence Using a completely consistent temporal arrangement rule, a generalized past output matrix is ​​constructed. With the generalized future output matrix .

[0098]

[0099]

[0100] Based on the aforementioned past data blocks, the system assembles the basic instrumental variable matrix in memory. :

[0101] This matrix is ​​fully encapsulated in algebraic topology. Historical excitation information of the initial state of the system at any given time.

[0102] In order to solve the state-space evolution matrix jointly using the least squares method at the end of the identification algorithm, the system must simultaneously construct a system capable of representing... A time-series shift matrix set for the state at each time step. Based on the shift property of the Hankel matrix, a recursive feature matrix shifted backward by one block unit is extracted on the time axis.

[0103] First, define the future input matrix after translation and reduction. With the future output matrix after translation and reduction Its structure is achieved by stripping away the original future matrix. and The first row is obtained in blocks, and the time steps cover to :

[0104] Corresponding They are generated using the same index structure.

[0105]

[0106] Secondly, define the past input matrix after translation expansion. Compared with the past output matrix after translation and expansion Its structure is achieved by combining the original... and The first row of blocks stripped out (i.e. Single-step data at any given time Appended to the bottom of the original past matrix to obtain:

[0107] Corresponding By adding generate.

[0108]

[0109] Based on the expanded past data blocks described above, the system constructs an updated instrument variable matrix. :

[0110] This completes the mapping from the original time-series data to a static high-dimensional geometric space. The resulting fundamental matrix set... Corresponding to the current moment The dynamic characteristics, and the translation characteristic matrix group Corresponding to the next moment The dynamic characteristics of the above two sets of orthogonal and complete matrix spaces are locked in the underlying memory architecture, serving as the legitimate input source for the dual oblique projection operation core.

[0111] In order to decouple the free response components driven purely by the internal state of the system from the mixed input-output data collected in closed-loop operation, and to completely eliminate the forced response of future control inputs on future output trajectories, the underlying computing core calls the high-dimensional geometric projection engine to perform oblique projection algebraic operations on the constructed basic matrix group and the translation feature matrix group.

[0112] The specific algebraic operation logic is as follows: First oblique projection: Extract the current time. dynamic characteristic subspace To isolate only the system The future output components, determined by the initial state at time step, are first defined with respect to the basic future input matrix. orthogonal projection operator This operator is used to project the entire data space to... In the absolutely orthogonal complement of the row space:

[0113] in, for An identity matrix of order 1. This represents the matrix transpose operation. Represents the Moore-Penrose pseudo-inverse operation of a matrix.

[0114] Subsequently, the fundamental future output matrix is ​​calculated. In the basic instrument variable matrix In the row space, along the fundamental future input matrix Oblique projection matrix of the row space Its precise algebraic expression is:

[0115] In the principle of system dynamics space mapping, this projection matrix It eliminates interference from external input excitation. Mathematically, it is strictly equivalent to the controlled system. Generalized observability matrix With the implicit state sequence at the current time The linear product of the system was thus successfully extracted, thus successfully extracting the characterization of the system in... Information on the evolution of the pure state at any given moment.

[0116] Second oblique projection: Extracting the next moment dynamic characteristic subspace To meet the algorithm's requirement for joint temporal recursive solution of the state-space evolution equations at the end, the software engine performs a second geometric projection operation on the temporal translation matrix group in parallel.

[0117] Similarly, define a method for translating the future input matrix. orthogonal projection operator :

[0118] Calculate the translation of the future output matrix In the updated instrument variable matrix In the row space, along the translation of the future input matrix Oblique projection matrix of the row space Its algebraic analytical expression is extended as follows:

[0119] The translation projection matrix Implicitly encapsulating the features of the system after a backward time step is strictly equivalent to Generalized observability matrix With the implicit state sequence at the next time step The linear product of .

[0120] Through the aforementioned dual geometric oblique projection operation, the software precisely extracts two key controlled subspace projection matrices from the high-dimensional hybrid data space, which are entirely determined by the inherent dynamic characteristics of the controlled object itself. and These two projection matrices are temporarily stored in the memory stack as an absolutely pure carrier of dynamic features, and are directly and exclusively fed into the subsequent weighted singular value decomposition decision module.

[0121] To obtain the controlled subspace projection matrix at the current moment. Subsequently, in order to separate the effective state subspace that governs the actual dynamic evolution of the system from the high-dimensional projection space, and to automatically filter out redundant noise dimensions caused by numerical calculation truncation or minor environmental perturbations, the underlying software modules... Perform weighted singular value decomposition (WSVD) and automated model order determination algorithms.

[0122] Introducing the left weighting matrix With right weighted matrix Under the standard deterministic subspace identification algorithm configuration, the software defaults to... and It is the identity matrix of the corresponding dimension; when faced with complex industrial noise distribution, the underlying estimation algorithm of the software can adaptively switch to the statistical optimal weight of the non-identity matrix to achieve pre-filtering.

[0123] Perform full-dimensional singular value decomposition on the weighted projection matrix:

[0124] in, An orthogonal matrix composed of left singular vectors (satisfying) ), An orthogonal matrix composed of right singular vectors (satisfying) ). For a diagonal matrix containing non-negative singular values, the singular values ​​on the main diagonal are arranged in monotonically decreasing order of magnitude: .

[0125] like Figure 3 As shown, traditional modeling relies heavily on control engineers manually setting the order of the physical system a priori. This software establishes an automated order determination criterion by analyzing the energy distribution characteristics of singular value spectra online.

[0126] Software for extracting singular value sequences Calculate the set of logarithmic energy decay gradients between adjacent singular values. :

[0127] The system automatically searches for the global maximum abrupt change point (i.e., the inflection point where the singular value spectral line experiences a "cliff-like" drop) in the gradient decay sequence. The index value corresponding to this abrupt change point is strictly defined by the system as the effective physical dynamic order of the controlled object. :

[0128] The above-mentioned decision logic mathematically completely separates the high-energy principal singular values, which represent the true internal state of the system, from the low-energy secondary singular values, which represent numerical noise, thus achieving zero-human intervention in the generation of the model order.

[0129] In the automatic locking system order Subsequently, the software engine... Using this as a boundary, matrix partitioning and dimensionality reduction are performed on the full-dimensional singular value decomposition results. The matrix... The space is divided into principal component blocks (subscript 1) corresponding to the effective state subspace and truncated blocks (subscript 2) corresponding to the noise subspace:

[0130] in: Including the front A left singular vector.

[0131] Including the front A right singular vector.

[0132] It is a core singular value diagonal matrix containing the dominant dynamic energy of the system.

[0133] After the above-described automated truncation process based on energy gradients, the high-dimensional projected space is precisely reduced to the lowest-dimensional closure space that can fully describe the system dynamics. The truncated principal singular value matrix set... It will serve as the sole algebraic basis for constructing the generalized observability matrix and the implicit state sequence.

[0134] After weighted singular value decomposition and automated dimensionality reduction truncation, the core singular value matrix set representing the dominant dynamic energy of the system was extracted. This step aims to use the dimensionality reduction result to algebraically reconstruct the extended observability matrix of the system without a prior physical model, and directly derive the implicit internal state sequence between the current time step and the next time step.

[0135] According to the subspace identification theorem, the projection matrix The column space of the system is consistent with the column space of the system's generalized observability matrix. This is achieved using the extracted left singular vector principal component blocks. Principal singular value diagonal matrix And the inverse of the defined left-weighted matrix The software's underlying structure directly constructs dimensions as The system's generalized observability matrix Its precise analytical expression is:

[0136] Built at this time The system is fully mapped in continuous The observation topology within each time step.

[0137] For linear time-invariant (LTI) systems, their generalized observability matrix possesses strict spatial translation invariance. To match the obtained reduced-order translation projection matrix... The corresponding time step must be constructed. Generalized observability matrix .

[0138] By performing a matrix truncation operation, the original generalized observability matrix is ​​directly removed. The end The rows (corresponding to the output observation dimension of the last step) are used to automatically extract the recursive observability matrix after the backward shift. :

[0139] This algebraic truncation operation achieves temporal recursive order reduction of the observation dimension without changing the system's basis coordinate system.

[0140] After establishing the spatial mapping relationship of the observation, the system can utilize the obtained high-purity double oblique projection matrix. and This allows for the direct reverse derivation of the actual internal state trajectory of the controlled object under the corresponding time series.

[0141] For the current moment implicit state sequence ,use Moore-Penrose pseudoinverse matrix For projection matrix Perform left multiplication:

[0142] The equation calculates the optimal state sequence in the least squares sense, which is in high agreement with the current observation history.

[0143] Similarly, for the next moment after recursively pushing forward one time step... implicit state sequence Using the translation observability matrix pseudo-inverse translation projection matrix Perform left multiplication:

[0144] Through the rigorous algebraic reconstruction process described above, the causal paradox in traditional control engineering, which necessitates prior knowledge of the state equations to estimate the state, is completely broken. The system, operating entirely in a "black box" state, directly generates a continuous two-step implicit internal state sequence with uniform dimensions and a consistent coordinate system using only external input and output data. and This array of state sequences with strict temporal causal relationships will be temporarily stored in a cache as a deterministic target variable for solving linear regression.

[0145] The implicit state sequence of the controlled system at the current moment has been successfully reconstructed. With the implicit state sequence of the next time step Next, the underlying algorithm will enter the final parameter estimation stage of subspace identification. To minimize the need for traditional step-by-step solutions (first calculating...),... Seeking again To address the accumulated error caused by [the error], the system constructs a global residual minimization objective function. Through a one-time inverse mapping of the overdetermined augmented matrix, the complete set of explicit discrete-time state-space matrices of the system is jointly solved. .

[0146] To establish a rigorous state-space mapping equation, the system first extracts external stimuli and observation responses that match the implicit state sequence from the constructed matrix space.

[0147] Locating to the basic future input matrix With the basic future output matrix The first row of blocks directly extracts the corresponding... Single-step input data sequence at time step With single-step output data sequence Its internal column vector element structure is as follows:

[0148]

[0149] At this point, the extracted input / output data blocks and the system's internal state sequence , In column dimension Strict physical temporal alignment was achieved in terms of (i.e., the number of samples).

[0150] Based on the standard dynamic equations of linear discrete time-invariant systems: as well as The software underlying layer will integrate the above two sets of equations into all... The samples are arrayed and expanded along the time dimension, and then concatenated into a high-dimensional augmented matrix equation.

[0151] Define the global system parameter matrix to be solved. Its algebraic construction and the corresponding augmented data equation are expressed as:

[0152] in, , Represents the model fitting residuals (Process Residuals) during the state evolution process. This represents the measurement residuals in the output observation equation. Due to the length of the effective observation data... The dimensionality is much larger than the parameters to be estimated. This augmented matrix equation mathematically constitutes a standard overdetermined system of linear equations.

[0153] To minimize the prediction error globally, the system directly establishes a joint least squares optimization criterion based on the Frobenius Norm for the aforementioned overdetermined equations:

[0154] The underlying computational core calls a high-precision linear algebra solver, utilizing the Moore-Penrose pseudo-inverse operator of the right-hand side explanatory variable matrix. We can directly obtain the globally unique optimal analytical solution to this optimization problem:

[0155] After completing the analytical multiplication operation, the software proceeds according to the preset system order. Input dimensions and output dimensions For the solved optimal parameter matrix Perform algebraic block extraction and output synchronously: State evolution matrix Input gain matrix Output observation matrix and direct feedforward matrix .

[0156] After obtaining the globally optimal system parameter matrix, in order to completely avoid subsequent safety control failures caused by unmodeled physical dynamics and identification truncation errors, the software does not ignore the fitting residual terms in the formula, but performs strict perturbation tracing and boundary quantization.

[0157] First, substitute the solved optimal parameter matrix back into the augmented state equations to calculate the actual posterior residual matrix under the historical trajectory. and :

[0158] Subsequently, the system follows the sample time series dimension ( The residual matrix is ​​traversed in an array to extract the maximum absolute error in each dimension, thereby constructing the upper bound vector of the single-step state perturbation. With the upper bound vector of single-step output perturbation Its any number The state component and the first The analytical formulas for each output component are as follows:

[0159]

[0160] After completing parameter identification and residual boundary extraction, the system performs a final algebraic reconstruction and summary of the constructed explicit discrete-time prediction model. Due to the unavoidable identification errors and physical disturbances in actual industrial environments, the system establishes a robust state-space model with additive bounded perturbations as follows:

[0161]

[0162] Based on the quantified historical extreme value boundary, the above disturbance at any given time All strictly satisfy the following hard boundary conditions based on the absolute values ​​of vector elements:

[0163]

[0164] Includes nominal identification parameters With uncertainty extreme boundary The complete mathematical description will serve as a high-confidence underlying dynamic foundation, providing extremely rigorous algebraic input for the subsequent construction of data-driven "ConstraintTightening" and "Robust MOAS".

[0165] In another implementation of the present invention, the maximum allowable output set is represented as:

[0166] in, and Reference instructions after constant modification Predicting future evolutionary trajectories under the influence of these factors; These are physical constraints.

[0167] For example, the global explicit discrete state-space parameter matrix of the controlled system is obtained. Subsequently, the underlying algorithm combines it with a standardized set of joint operating condition constraints. By constructing the polyhedral boundary of the Maximum Output Allowable Set (MOAS) offline and running the Scalar Reference Governor (SRG) online, the system achieves strict compliance correction for the original external reference signal.

[0168] This set The safety space is defined as the joint safety space of all initial states and reference inputs that ensure the system does not violate constraints over an infinitely long line-of-sight period. The infinite-dimensional constraints are truncated to a finite number of predefined steps using a finite-step decidability algorithm. This allows for the precise reconstruction of the polyhedral form of MOAS.

[0169] In the system's real-time closed-loop control loop, the generated anti-limit-crossing logic is connected in series as a supervisory filter before the actual closed-loop controller. This supervisory filter intercepts the original expected reference signal sent from the upper layer in real time. In each control cycle, the supervisory filter solves the following rigorous convex quadratic programming (QP) optimization problem with linear inequality constraints online:

[0170]

[0171] Based on the operational logic of the Scalar Reference Governor, the optimization algorithm evaluates the current state estimate in real time. With the original instructions Combinatorial safety, dynamically adjusting scalar factors If the system detects that directly executing the original instructions may lead to exceeding the limit, it will solve for the maximum value online. The value makes the generated safety reference signal While ensuring that the system state always remains within the MOAS boundary, it approaches the original target signal with optimal weights. This pre-monitoring filter mechanism, through this scalar truncation strategy, achieves real-time safety and compliance correction of the original reference signal, thereby achieving the overall goal of absolute safety control to prevent boundary violations.

[0172] Specifically, in order to seamlessly integrate with and fully reuse the rigorous geometric computation framework of the classic Reference Governor (RG) theory at the software level, while greatly simplifying the computational architecture of the online optimization solver, the system first addresses the independently defined physical control input constraints ( ) and system output constraints ( Perform a unified reconstruction of the algebraic dimensions and synchronously map the established external perturbation vectors into the new augmented space.

[0173] The software system extracts the actual control input vector of the underlying actuator. The actual measurement output vector of the controlled object In the underlying memory, the system opens a new continuous data channel and vertically concatenates the two physical vectors with different dimensions along the column dimension to define a completely new generalized restricted output vector. :

[0174] This generalized vector physically encompasses all restricted variables that may touch the safety boundary during the operation of the controlled system.

[0175] In an architecture where a scalar reference controller (SRG) intervenes in the system control loop, the safety reference command generated by the controller... This will serve as the actual control input for the underlying actuators, meaning it satisfies a strict equation in the mathematical mapping: .

[0176] Combined with the established robust state-space observation equations with additive bounded perturbations For generalized restricted output vectors The prediction equation can be rigorously analyzed and expanded as follows:

[0177] in, for OK A matrix of all zeros in the columns. for An identity matrix of order 1.

[0178] Based on this, the software automatically constructs a completely new generalized output matrix at the underlying level. With generalized feedforward matrix and generalized measurement perturbation vector :

[0179] Through the above augmentation operations, the system can express the generalized restricted output equation in a compact and standardized manner without adding any additional model identification burden:

[0180] Furthermore, based on the obtained extreme boundary, this generalized perturbation vector is strictly bounded by... .

[0181] Based on the above variable reconstruction, the input boundary polyhedra are established respectively. With output boundary polyhedron They will be merged in the underlying algorithm.

[0182] Define the generalized constraint weight matrix for diagonal partitioning. With augmented constraint boundary vector :

[0183] Ultimately, the various previously disparate physical constraints are uniformly transformed by the software into the most standard and fundamental compact linear polyhedral space inequality form in classical RG theory:

[0184] Through structured reorganization, the software not only completely eliminates the mathematical differences between "state / output constraints" and "input constraints" in its underlying code logic, but also exquisitely addresses the single-step measurement identification errors of the model. Synchronous isolation and encapsulation into a generalized perturbation vector This standardized and lossless algebraic channel reconstruction allows for the direct application of perturbation extremum boundaries to the static constraint set in subsequent steps. Performing "Robust Constraint Tightening" and constructing an absolutely safe Robust Maximum Output Allowable Set (Robust MOAS) provides the perfect structural preparatory work.

[0185] A generalized constrained output equation containing the worst global perturbation term was established. and generalized fundamental polyhedral constraints Following this, the system enters the offline safety geometric boundary reconstruction phase. This step aims to utilize the asymptotic stability characteristics of the closed-loop system to handle uncertain disturbances. and The security assessment in the infinite-dimensional time domain of the future is analytically transformed into a set of finite-dimensional convex polyhedral inequalities with absolute robustness.

[0186] To ensure absolute system safety under all possible perturbations, the system must evolve in the worst-case scenario. According to classical robust reference regulator theory, the system state evolution is subject to process perturbations. The continued impact, in the future Step output prediction Geometrically, this manifests as an ever-expanding set-valued prediction:

[0187] in, For the nominal trajectory prediction without disturbance, For Minkowski space and operators, For disturbance in The set of maximum cumulative uncertainties caused to the output after each step. To ensure Strictly located in the safe set Inside, nominal trajectory It must be further constrained to the Pontryaginal difference between the original set and the perturbation set, i.e. .

[0188] In the underlying matrix operations of the software code, this topological difference operation is precisely transformed into a one-dimensional constraint boundary vector. Recursive tightening. Define the first... The robust constraint boundary vector of the step is .

[0189] for At any given time, affected only by observational perturbations, the software calculates the initial tightening boundary:

[0190] For any number of prediction steps The disturbance along the system dynamic matrix Propagation, software based on extreme value boundaries Perform recursive algebra tightening:

[0191] in, This represents an operator that takes the absolute value of all elements within a matrix / vector. Through this recursive formula, the system avoids complex high-dimensional set operations, obtaining the absolutely safe and robust boundary required for each step using only basic vector subtraction. .

[0192] To construct the constraints, the nominal output prediction needs to be expanded. Traditional multi-step prediction formulas contain complex state gain accumulation terms (i.e., convolution sums). This will greatly increase the matrix splicing dimension during online configuration.

[0193] Therefore, the underlying software algorithm utilizes a closed-loop state evolution matrix. Schur stability characteristics lead to the introduction of a closed algebraic identity for the summation of infinite matrix series:

[0194] Based on this, the future... nominal generalized restricted output of step It can be parsed and folded into an extremely compact form as follows:

[0195] Accordingly, the system automatically generates and stores data at the underlying level for future use. State projection constant matrix for step constraint determination Projection constant matrix with reference instruction :

[0196]

[0197] At this time, the The robust absolute safety constraint of the step is perfectly linearized as follows: .

[0198] To ensure the instructions To ensure the system can stably reside within a safe space over infinite time, additional steady-state constraints are required. .

[0199] When predicting steps At that time, due to Stable attenuation characteristics , , It will converge to the steady-state gain matrix. To address residual jitter in steady state and maintain sufficient safety margin, the software introduces a configurable steady-state positive tolerance factor. Define steady-state robust boundary conditions:

[0200] like Figure 4 As shown, based on the Finite Determinedness theorem, since the nominal matrix of the system is strictly asymptotically stable, and a steady-state safety margin has been introduced, In computing the infinite time-domain secure geometric space At that time, there must exist a finite integer number of truncation steps. .

[0201] Execute iterative search decision algorithm: Initialization .

[0202] In each subsequent step, the system performs intersection expansion:

[0203] In each iteration, the algorithm uses linear programming (LP) to determine the newly added... Whether the step constraint set is completely enveloped by the constructed constraint set. Internally, once it is detected that a new constraint does not produce any cutting effect on the current polyhedral space (i.e., spatial redundancy occurs), the truncation step is confirmed to have been reached. The iteration terminates.

[0204] Ultimately, the system converges the infinite-dimensional time series into a set of ultra-high-dimensional exact finite row matrix inequalities with definite physical meaning, encompassing all time-domain dynamics and unknown perturbation boundaries:

[0205] The entire structure is encapsulated into a highly streamlined final robust polyhedron form:

[0206] Thus, the complex infinite-time-domain security prediction problem containing bounded unknown perturbations has been successfully transformed into a set of static constant coefficient matrices through extremely rigorous matrix algebraic transformations and spatial difference dimensionality reduction. This robust constraint datum array provides static boundary calls for the next stage of online one-dimensional scalar rapid optimization.

[0207] In another implementation of the present invention, the optimal safety reference signal is represented as:

[0208] in, It is the globally optimal scalar factor; The instruction from the previous moment; This is the original expected reference instruction.

[0209] In another implementation of the present invention, the globally optimal scalar factor is expressed as:

[0210] in, Let be the total number of rows in the polyhedral inequalities. For any of the inequalities in the system of inequalities Row index; The first element in the system's real-time dynamic gain vector at the current moment. Each element component; Let be the th element in the system's real-time safety margin vector at the current moment. Each element component.

[0211] For example, such as Figure 5 As shown, the polyhedral space reconstruction of the maximum allowable set of outputs (MOAS) is completed in the offline stage. Following this, the system will enter a real-time online closed-loop control process. This step aims to utilize the Scalar Reference Governor (SRG) architecture to perform real-time safety assessment and compliance truncation of the original expected instructions issued externally within each control cycle.

[0212] In any real-time online control cycle Internally, the system obtains the optimal internal state estimation vector recursively derived from the underlying state observer. And the original expected reference instructions issued by the planning algorithm or operator at the next higher level. Simultaneously, the system retrieves the absolute safety command calculated and issued to the actuator in the previous control cycle. .

[0213] To completely avoid the high computational cost of solving multidimensional nonlinear or quadratic programming (QP) problems online, the system establishes a scalar factor-based filter in the pre-supervisory filter. One-dimensional instruction evolution law:

[0214] in, Let be the scalar interpolation weights to be solved. This evolutionary law, in geometric topology, represents: the actual instructions issued in the current cycle. Must be strictly located in Starting point On a multidimensional line segment with the endpoint as the endpoint.

[0215] The goal of the software's online limit-breaking control is transformed into: ensuring the current state and instruction combination... Strictly subordinate to the safety boundary of finite-dimensional polyhedra constructed offline Under the premise of finding the largest This makes the revised instructions Approximate the original desired instruction as closely as possible along the multidimensional line segment. .

[0216] Substituting the aforementioned one-dimensional instruction evolution law directly into the established set of MOAS polyhedral inequalities In the middle, we get:

[0217] By algebraically rearranging terms and merging like terms, the software's underlying implementation precisely reduces this multidimensional search problem to a set of terms relating only to a single scalar. Linear inequality system:

[0218] Here, the real-time dynamic gain vector of the system at the current moment is defined. With the real-time safety margin vector as follows( (Total number of rows in polyhedral inequalities)

[0219]

[0220] Based on the principle of positive invariance of the MOAS set, due to the instructions at the previous time step... For the current prediction status It still maintains theoretical absolute security, therefore the real-time security margin vector mathematically always satisfies the non-negativity property, that is... Heng was established.

[0221] After performing algebraic dimensionality reduction, find the maximum scalar weight. The optimization problem is simplified to a very basic ratio-test method. The solver does not need to call any interior point method or active set method that requires repeated iterations, but directly outputs the optimal solution through analytical expression.

[0222] For any i in the system of inequalities OK( ): like This indicates that the instruction evolves along the desired direction without touching the constraint boundary of that dimension, at which point the row... This does not constitute a restriction.

[0223] like This indicates that instruction evolution is approaching a safety boundary, and the upper limit of the scalar factor must be limited, i.e. .

[0224] Therefore, the system directly obtains the globally optimal scalar factor by performing a single traversal of the hardware vector register. :

[0225] Calculation obtained Then, the system substitutes it back into the one-dimensional instruction evolution law:

[0226] The software then immediately implemented the absolutely safe correction command. It was issued to the lower-level implementing agencies.

[0227] like Figure 6 As shown, the mathematical safety commands calculated online by the scalar reference regulator are transformed into underlying physical control behaviors, and the system time axis is driven to step by introducing a rolling time domain (Receding Horizon) mechanism to achieve a closed loop of anti-limit overrun control.

[0228] In the current control cycle After the online optimization solution is completed, the system extracts the optimal scalar factor for preventing boundary breaches obtained from the real-time solution. And synthesize an absolutely safe reference signal. .

[0229] The software system will use virtual mathematical instructions The actual tracking setpoint is sent to the pre-configured, stable underlying closed-loop controller. The underlying closed-loop controller receives... Then, based on the current physical measurement state of the controlled object and its internally fixed feedback control law, the specific actuator actuation signal is calculated. This actuation signal directly drives the physical controlled object to produce a controlled response. Because... Strictly belongs to the maximum allowable output set The physical response is ensured by a mathematical mechanism to never exceed the preset physical hard constraint boundary.

[0230] After the underlying controller completes the action on the physically controlled object, the current control cycle begins. Once the execution phase ends, the state of the controlled object undergoes a physical evolution.

[0231] The system synchronously acquires the latest measured controlled output data and updates the system's measured input and output data repository. In the memory data stack, the system-maintained data time window slides forward, saving the currently issued control commands and measurement data. Subsequently, the system's underlying global clock triggers the recursive evolution of the time step, shifting the control time step from the current moment... Proceed to the next moment .

[0232] Time step advances to Then, in order to prepare for the next round of prediction optimization, the system must obtain the latest internal state.

[0233] The system's underlying layer calls the state observer to absorb the latest external measurement data and complete the time update and measurement update of the state vector. The specific recursive calculation process includes two steps: Time update (prior prediction): using the identified system parameter matrix And the instructions issued in the previous cycle Predict the system state and output at the next moment:

[0234]

[0235] Measurement update (posterior correction): Incorporating the latest actual measurement output data. Combined with the observer gain matrix The residuals of the prior state are corrected to obtain the current state. High-precision state vector at any given time :

[0236] Obtain the updated state vector Afterwards, the software logic jumps back and enters steps S2 to S4. The system will receive a new round of original reference instructions for the new time window. And a new round of online optimization evaluation is performed under the constraints of an infinite time-domain polyhedral space.

[0237] The aforementioned "measurement-update-optimization-deployment" process repeats infinitely on the timeline. This Receding Horizon mechanism ensures that the system can correct deviations caused by unmodeled dynamics and external disturbances in real time, thereby giving the controlled system continuous adaptability to dynamic disturbances and absolute global security.

[0238] Example 1 In a specific embodiment for verifying the control effect of the present invention, the system preset test condition parameters are as follows: the original expected command issued externally. A step signal with an amplitude of 1.0; the physical safety hard constraint of the controlled object is set as: actuation signal. The upper and lower bounds are limited by Within the interval, state variables The upper and lower bounds are limited by Within the range.

[0239] It should be noted that the specific command amplitudes and constraint boundary values ​​mentioned above are merely exemplary test conditions selected to intuitively demonstrate the anti-limit-over-limit effect of the present invention. In actual industrial applications, the above parameters and boundary conditions can be automatically and regularly configured by the system of the present invention within any range according to the actual physical characteristics of different controlled objects (such as the maximum valve opening, the motor's limit torque, the container's safe liquid level, etc.). The specific values ​​mentioned above should not constitute any limitation on the universality and scope of protection of the algorithm of the present invention.

[0240] like Figure 7 The figure shows the closed-loop response curve of the system after adopting the data-driven reference signal security adjustment method provided in this embodiment of the invention. The upper part of the figure is the command tracking channel, and the lower part is the physical constraint channel (including state variables). With actuation signal (The upper and lower bounds). It can be seen that when the original expected instruction... When a step occurs, the scalar reference regulator of this invention generates a smooth, absolutely safe command through online optimization. (Dashed line). In the instruction Driven by [the system's] requirements, the system approximates the reference value as closely as possible while satisfying the constraints. Throughout the entire transient and steady-state evolution cycle, although the underlying physical state... With actuation signal Dramatic fluctuations occurred, but all were strictly limited to the preset safety boundaries (i.e., ,and No limit-crossing behavior occurred within the system, verifying the effectiveness of the absolute limit-crossing control of this invention.

[0241] like Figure 8 The figure shows the closed-loop response curve of the control group system that does not employ the safety regulation mechanism of this invention. In this control group, the system directly sends the original expected command to the underlying controller (i.e., As can be seen from the curve at the top of the graph, although the system output... It was able to respond quickly and eventually reach the target setpoint, but as can be seen from the physical constraint channel below, in the initial stage of the system's transient response, the state variables... and actuation signal Extremely severe overshoot occurred in all cases, significantly exceeding the safety hard constraint boundaries marked in the figure (e.g. In the initial stage, it far exceeded its upper limit of 0.1. It also severely fell below the lower limit of -0.09. In actual industrial settings, such severe over-limit behavior during transient processes often directly leads to irreversible damage to equipment or triggers system safety shutdown protection. Figure 7 and Figure 8 The direct comparison fully demonstrates the significant technical advantages of this invention in ensuring the safe operation of complex dynamic systems.

[0242] In another aspect, the present invention provides a data-driven reference signal security conditioning system, comprising: Data acquisition module: Collects operational data of the closed-loop system during the controlled process at a fixed sampling period, and performs data preprocessing to obtain a high-quality standardized dataset.

[0243] Constraint Construction Module: Defines physical constraints based on the physical limits of the controlled object and the safety requirements of the production process.

[0244] Parameter solving module: Based on high-quality standardized datasets and subspace identification algorithms, an explicit discrete-time state-space linear model describing the implicit dynamics of the system is established, and the linear model parameters of the system are solved.

[0245] Optimal signal calculation module: Combining physical constraints and linear model parameters, it calculates the maximum allowable output set and runs the scalar reference regulator online to obtain the optimal safety reference signal in real time.

[0246] Drive module: The underlying controller calculates the specific actuator actuation signal based on the optimal safety reference signal to drive the controlled object.

[0247] Rolling optimization module: Introduces a rolling time domain mechanism to drive the system time axis stepping and realize a closed loop of anti-limit overrun control.

[0248] Another aspect of the present invention provides an electronic device, which includes a processor, a memory, a communication bus, and a communication interface.

[0249] in: The processor, memory, and communication interface communicate with each other via a communication bus.

[0250] A communication interface is used to communicate with other electronic devices or servers.

[0251] The processor is used to execute programs, specifically the steps of any of the data-driven reference signal security adjustment methods described in the above embodiments.

[0252] Specifically, the program may include program code, which includes computer operation instructions.

[0253] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0254] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device.

[0255] Specifically, the program can be used to cause the processor to execute the steps of any of the data-driven reference signal security conditioning methods described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed by any of the data-driven reference signal security conditioning methods described above, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0256] An exemplary embodiment of this application also provides a computer storage medium, wherein a computer program is stored on the computer storage medium, and when the computer program is executed by a processor, it implements the methods of the various embodiments of this application.

[0257] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0258] Specific embodiments of the present invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result.

[0259] It should be noted that all directional indications (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between the components in a certain order (as shown in the figure). If the specific order changes, the directional indication will also change accordingly.

[0260] In the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include at least one of that feature.

[0261] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0262] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0263] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0264] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data-driven reference signal security modulation method, characterized in that, include: The system collects operational data during the controlled process of the closed-loop system at a fixed sampling period and performs data preprocessing to obtain a high-quality standardized dataset. Define physical constraints based on the physical limits of the controlled object and the safety requirements of the production process; Based on high-quality standardized datasets and subspace identification algorithms, an explicit discrete-time state-space linear model describing the implicit dynamics of the system is established, and the linear model parameters of the system are solved. By combining physical constraints and linear model parameters, the maximum allowable set of outputs is calculated, and the scalar reference regulator is run online to obtain the optimal safety reference signal in real time. The underlying controller calculates specific actuator actuation signals based on the optimal safety reference signal to drive the controlled object; A rolling time-domain mechanism is introduced to drive the system's time axis stepping, thereby achieving a closed-loop control system to prevent exceeding limits.

2. The method according to claim 1, characterized in that, The running data includes the input sequence and the output sequence; Data preprocessing includes: Calculate the mean of the input sequence and the output sequence respectively, and perform mean-reduction processing to obtain a volatility sequence with zero mean; A low-pass digital filter is used to filter the zero-mean fluctuation sequence to eliminate high-frequency measurement noise.

3. The method according to claim 1, characterized in that, Physical constraints include control input constraints and system output constraints; Control input constraints are represented as follows: in, and These represent the lower and upper limit vectors of the actuator, respectively; For the input vector, express A real vector space of dimensionless numbers, i.e. for 3D real vector This represents the total number of input channels. The system output constraints are expressed as follows: in, and This is a preset safety boundary vector; For the output vector, express A real vector space of dimensionless numbers, i.e. for 3D real vector This represents the total number of constrained output channels.

4. The method according to claim 3, characterized in that, The dynamic equations of the explicit discrete-time state-space linear model are expressed as follows: in, Let be the internal state vector of the system. The system order; These are the system matrix to be identified, the input matrix, the output matrix, and the direct transfer matrix, respectively.

5. The method according to claim 1, characterized in that, The maximum allowable output set is represented as: in, and Reference instructions after constant modification Predicting future evolutionary trajectories under the influence of these factors; These are physical constraints.

6. The method according to claim 1, characterized in that, The optimal safety reference signal is represented as: in, It is the globally optimal scalar factor; The instruction from the previous moment; This is the original expected reference instruction.

7. The method according to claim 6, characterized in that, The globally optimal scalar factor is expressed as: in, Let be the total number of rows in the polyhedral inequalities. For any of the inequalities in the system of inequalities Row index; The first element in the system's real-time dynamic gain vector at the current moment. Each element component; Let be the th element in the system's real-time safety margin vector at the current moment. Each element component.

8. A data-driven reference signal security conditioning system, characterized in that, include: Data acquisition module: Acquires operational data of the controlled process of the closed-loop system at a fixed sampling period, and performs data preprocessing to obtain a high-quality standardized dataset; Constraint construction module: Defines physical constraints based on the physical limits of the controlled object and the safety requirements of the production process; Parameter solving module: Based on high-quality standardized datasets and subspace identification algorithms, an explicit discrete-time state-space linear model describing the implicit dynamics of the system is established, and the linear model parameters of the system are solved. Optimal signal calculation module: Combining physical constraints and linear model parameters, it calculates the maximum allowable set of outputs and runs the scalar reference regulator online to obtain the optimal safety reference signal in real time; Drive module: The underlying controller calculates the specific actuator actuation signal based on the optimal safety reference signal to drive the controlled object; Rolling optimization module: Introduces a rolling time domain mechanism to drive the system time axis stepping and realize a closed loop of anti-limit overrun control.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the data-driven reference signal security conditioning method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, A computer program is stored on a computer storage medium, and when the computer program is executed by a processor, it implements the steps of the data-driven reference signal security adjustment method as described in any one of claims 1 to 7.

Images