Analytical system for complex multi-source computations

Abstract

The application relates to the technical field of computer system data analysis based on specific calculation models, and discloses an analysis system for complex multi-source calculation, which comprises a data input interface, a feature decoupling operation module, a calculation logic consistency checking module and a hidden state sensing module; original input data flow is projected to a feature subspace through a plurality of orthogonal basis vectors, nonlinear interference among feature components is eliminated, and a decoupled state feature vector is extracted; meanwhile, a physical boundary is constructed according to real-time power consumption parameters of a controlled object power source, a determined energy envelope is used as an audit threshold of the state feature vector, and forced convergence correction is executed when the audit threshold is exceeded; the application establishes an associated closed loop of calculation logic and a physical boundary, effectively solves the logic divergence problem in multi-source calculation, realizes adaptive compensation for the quasi-static drift deviation of a controlled object, and enhances system operation reliability.

Description

Technical Field

[0001] This invention relates to an analysis system for complex multi-source computation, belonging to the field of computer system data analysis technology based on a specific computation model. Background Technology

[0002] Currently, in the automated control process of metal rolling, high-precision thickness control systems integrate multi-source data, utilizing heterogeneous data generated by online thickness measurement, material tracking, and torque sensing, along with a pre-set deformation resistance model to calculate roll gap compensation to maintain strip thickness consistency. Under high-speed continuous rolling conditions, the physical variables collected by sensors exhibit frequency heterogeneity. The sampling period of fast variables such as rolling force and roll speed is in the millisecond range, while the update frequency of process variables such as material prediction and initial thickness is lower than that of fast variables. This leads to the technical constraint of sampling frequency mismatch when the calculation system processes parameter fusion. Existing technologies use zero-order hold or linear interpolation to align asynchronous data streams, but this results in a step jump in calculation results at points of sudden changes in material properties or switching of operating conditions. Furthermore, due to the nonlinear coupling between physical quantities, the serial calculation model cannot achieve deep decoupling of features while maintaining high-frequency response. This causes the compensation output to lag behind the physical deformation process, restricting the improvement of control accuracy. Moreover, due to the amplification effect of physical inertia, high-frequency oscillations are induced in the actuator.

[0003] Although existing technologies attempt to alleviate the aforementioned contradictions through hardware means such as optimizing roll body shape or reinforcing roll system stiffness, practice has shown that architectural limitations at the control algorithm level are the deeper cause of accuracy bottlenecks. For example, Chinese invention patent CN114120401A discloses a face anti-fraud method based on cross-domain feature alignment networks. By constructing a domain adapter and a multi-scale attention fusion module, it alleviates the distribution differences between heterogeneous data at the algorithm level to enhance feature generalization. However, such alignment logic, which is purely based on pattern recognition, is essentially still in the category of pure data-driven operation detached from physical entity constraints. In industrial control scenarios such as metal rolling, where real-time determinism and security requirements are extremely high, this approach is not feasible. In practice, relying solely on logical alignment of the feature space often fails to identify logically closed but physically divergent abnormal compensation commands caused by sensor interference. Furthermore, it is difficult to correlate the computational output with the energy conservation boundary of the mill's main drive in real time, making the computational model prone to instability under extreme conditions. To address these challenges, researchers have attempted to introduce high-order nonlinear iterative algorithms or deep learning models to capture dynamic features. However, such solutions incur significant computational overhead when processing high-dimensional heterogeneous data, making it difficult to meet the microsecond-level real-time closed-loop control requirements of hydraulic servo systems. Moreover, when sensors generate abnormal data due to environmental interference, there is a lack of a physical-level rationality audit mechanism, which can easily lead to logically correct but physically dangerous execution commands.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a specific computing architecture and corresponding feature space mapping computing model that adapts to the asynchronous characteristics of frequency and achieves deep decoupling of parameters while ensuring the real-time requirements of industrial controllers, thereby eliminating the computing steps caused by data frequency mismatch and solving the convergence problem caused by the disconnect between logic and physical reality of the specific computing model. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: An analysis system for complex multi-source computation, the system comprising a data input interface, a feature decoupling operation module, a computational logic consistency verification module, and an implicit state perception module:

[0006] The data input interface is used to acquire the raw input data stream that characterizes the multi-source coupled operating conditions of the controlled object;

[0007] The feature decoupling operation module, connected to the data input interface, is configured to establish feature subspace mapping logic. It projects the original input data stream onto a non-interfering feature subspace through multiple pre-calibrated orthogonal basis vectors to eliminate nonlinear interference between feature components of different dimensions and extract the decoupled state feature vector.

[0008] The computational logic consistency verification module, connected to the feature decoupling operation module, is configured to construct a physical boundary model based on the real-time power consumption parameters of the power source of the controlled object, and use the energy envelope determined by the physical boundary model as the audit threshold of the state feature vector. When the calculated output of the state feature vector exceeds the audit threshold, forced convergence correction is performed.

[0009] The implicit state perception module is connected to the feature decoupling operation module and the computational logic consistency verification module, respectively. It is configured to actively capture the residual signal generated in the projection operation, extract the associated perception features by analyzing the distribution characteristics of the residual signal, and generate a depth adjustment instruction for the feature subspace mapping logic based on the associated perception features, so as to realize the adaptive compensation of the quasi-static drift deviation of the controlled object by the computational model.

[0010] Preferably, the feature decoupling operation module follows the following calculation rules when extracting state feature vectors: ,in, For state feature vectors, The temporal feature vector in the original input data stream, Let i be the i-th orthogonal basis vector in the feature subspace. The feature decoupling operation module dynamically adjusts the projection weights of the orthogonal basis vectors based on the cross-correlation coefficients between physical variables in the original input data stream, thereby deeply decoupling the transient load fluctuation data and matrix resistance characteristic data in the calculation dimension, and stabilizing the state feature vector within the preset calculation convergence range.

[0011] Preferably, the computational logic consistency verification module includes an energy conservation determination unit; the energy conservation determination unit is configured to construct a physical boundary model based on the principle of energy conservation based on the real-time current and voltage parameters of the power source of the controlled object, and calculate the upper limit of the energy change rate corresponding to the physical boundary model. The computational logic consistency verification module substitutes the predicted compensation amount corresponding to the state feature vector into the physical boundary model for safety simulation. When the simulation yields the predicted energy change rate... satisfy At that time, the computational logic consistency verification module determines that the current computational logic has entered a divergent state.

[0012] Preferably, the implicit state perception module includes a residual information mining unit; the residual information mining unit is configured to perform frequency domain decomposition of the residual signal based on wavelet transform, and extract high-frequency pulse features and low-frequency drift features in the residual signal; the implicit state perception module identifies the quasi-static displacement deviation of the controlled object caused by the thermal response of the execution component based on the monotonic change trend of the low-frequency drift features, and identifies the transient disturbance inside the controlled object caused by the non-uniformity of the input matrix based on the amplitude distribution of the high-frequency pulse features.

[0013] Preferably, the system further includes an active micro-excitation detection unit; the active micro-excitation detection unit is used to inject a sinusoidal disturbance signal with a frequency of 50Hz to 100Hz and an amplitude of less than 2μm into the response execution loop of the controlled object, and simultaneously extract the cross-correlation coefficient between the response signal and the sinusoidal disturbance signal in the original input data stream; the feature decoupling operation module inverts the instantaneous dynamic stiffness characteristics of the controlled object based on the cross-correlation coefficient, and inputs the instantaneous dynamic stiffness characteristics as a correction factor into the feature subspace mapping logic.

[0014] Preferably, the data input interface is configured to perform data translation processing, converting the physical quantity measurement signals of the controlled object into discrete data packets; the data input interface also includes a synchronization arbitrator for time-stamp alignment processing of input signals from multiple sampling sources, ensuring that the original input data stream entering the feature decoupling operation module has a unified logical reference in the time dimension.

[0015] Preferably, the orthogonal basis vectors are generated by a combination of offline calibration and online updating; the feature decoupling operation module is configured to periodically correct the orthogonal basis vectors using the Gram-Schmidt orthogonalization algorithm based on the accompanying sensing features output by the latent state perception module, in order to compensate for the mismatch of the computational model caused by the evolution of the external environment.

[0016] Preferably, the state feature vector includes feature quantities used to characterize the system load distribution and feature quantities used to characterize the geometric position of the controlled object; the feature decoupling operation module realizes the orthogonality of the feature quantities of the system load distribution and the feature quantities of the geometric position of the controlled object in the calculation dimension through feature subspace projection, thereby eliminating the reaction force influence of the position adjustment action on the stability of the load distribution.

[0017] Preferably, the computational logic consistency verification module further includes a redundant decision unit; the redundant decision unit is configured to identify computational anomalies caused by algorithm overflow or data singularities by comparing the output consistency of multiple parallel computation paths after taking into account the disturbance components injected by the active micro-excitation detection unit; when the output deviation of multiple parallel computation paths exceeds a preset tolerance threshold, the redundant decision unit forcibly calls a safety preset value based on historical steady-state data.

[0018] Preferably, the system adopts a distributed computing architecture, in which the feature decoupling operation module is deployed on the real-time processing plane and the implicit state perception module is deployed on the strategy analysis plane. The strategy analysis plane generates system operation status evaluation data based on the long-term statistical characteristics of the residual signal and sends depth adjustment instructions to the real-time processing plane to dynamically adjust the projection depth parameters of the orthogonal basis vectors so that the running trajectory of the execution component of the controlled object is stabilized within the preset range determined by the geometric shape and position features of the target.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In the analysis of complex multi-source computations, by constructing an asynchronous feature projection space, the problem of computational logic discontinuity under sampling frequency mismatch of multi-source heterogeneous data is solved. Existing technologies often produce computational steps or cause compensation instructions to lag due to waiting for slow variable updates when processing input variables with different periods. This invention uses an orthogonal subspace mapping mechanism to decouple fast sampling variables and slow process variables to different logic planes. Within the sampling gap of slow variables, predictive linear extrapolation compensation is achieved by using the incremental change rate of fast variables. This transforms the system from passively waiting for data updates to actively deducing logic based on changing trends, eliminating computational dead zones caused by asynchronous data, and ensuring the continuity and smoothness of compensation instructions on the time axis.

[0021] 2. By orthogonalizing the feature plane, the deep decoupling of strongly coupled physical quantities in the computational dimension is achieved. In complex multi-source environments, different types of physical variables often have nonlinear interference, causing the disturbance of a single variable to cause drastic fluctuations in the global output. The state space mapping logic established in this invention projects heterogeneous inputs to mutually non-interfering feature subspaces through pre-calibrated orthogonal basis vectors, cutting off the interference path between tension fluctuations and material hardness changes from the computational level. This allows the system to still output physically meaningful deterministic compensation values ​​at the moment of sudden changes in material properties or switching of working conditions, avoiding high-frequency oscillations in the control loop caused by parametric coupling.

[0022] 3. By introducing an arbitration mechanism based on the energy conservation boundary, a safe closed loop is established between computational logic and physical reality. This invention utilizes the real-time electrical parameters of the rolling mill main motor to map the current physical energy envelope, and uses this as the physical intuition audit threshold for the output results of the computational analysis system. When a logically correct but physically dangerous compensation amount is generated due to sensor failure or algorithm overflow, the system forces convergence and executes instructions through the boundary defined by the law of energy conservation. This solves the problem of logical divergence that is prone to occur in complex analysis systems under extreme interference. Based on a computational model, a first-principles-based low-level defense capability is provided to ensure the safety of production equipment when encountering sudden data anomalies from the perspective of computer system algorithm architecture. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the data flow path and logic control closed loop of each functional module of the present invention;

[0024] Figure 2 This is a schematic diagram of the hardware deployment architecture and physical connection of the present invention in the hot-rolled strip steel production line scenario. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides an analysis system for complex multi-source computing, comprising a data input interface, a feature decoupling operation module, a computational logic consistency verification module, and an implicit state perception module. The data input interface receives and aligns heterogeneous data streams from a sensor array. The feature decoupling operation module establishes decoupling mapping logic between heterogeneous variables through orthogonal basis vector projection. The computational logic consistency verification module constructs physical boundary thresholds based on the power consumption parameters of the controlled object's power source. The implicit state perception module achieves adaptive compensation for quasi-static drift of the system by mining projection residual signals. All modules operate collaboratively through high and low processing planes to jointly achieve accurate analysis of the state of the controlled object under complex operating conditions. In high-precision automated control scenarios such as metal rolling, the system... Systems often need to process real-time data from multiple sources simultaneously, such as hydraulic servos, thickness gauges, and load sensors. Due to differences in the hardware characteristics and communication protocols of various sensors, the collected physical variables exhibit asynchronous sampling frequencies. For example, the sampling period for fast variables such as rolling force and roll speed is typically 1ms, while the update frequency for process variables such as material prediction and initial thickness is much lower. This frequency mismatch makes traditional synchronous calculation models prone to computational steps or response lags when handling parameter fusion. To address this challenge, the data input interface in this invention is configured to perform data translation processing. A built-in synchronous arbitrator is used to perform time-scale alignment processing on input signals from multiple sampling sources. This synchronous arbitrator receives the first type of fast variables... With the second type of slow variable Then, based on the sampling period of the fast variable, the slow variable is subjected to discretization data packet encapsulation based on the zero-order hold, so as to ensure that the original input data stream entering the feature decoupling operation module has a unified logical benchmark in the time dimension.

[0027] After the heterogeneous data streams are time-scaled, due to the strong nonlinear coupling between tension fluctuations and material hardness parameters during metal rolling, even small disturbances in a single physical quantity can cause drastic fluctuations in the global calculation output through the system stiffness chain. This makes it difficult for traditional serial calculation models to accurately extract independent controlled features. The feature decoupling operation module establishes a feature subspace mapping logic, using multiple pre-calibrated orthogonal basis vectors to project the original input data stream into mutually independent feature subspaces. When extracting state feature vectors, the feature decoupling operation module follows the following calculation rules: ,in, For state feature vectors, The temporal feature vector in the original input data stream, Let be the i-th orthogonal basis vector in the feature subspace. During the execution of this procedure, the feature decoupling operation module dynamically adjusts the projection weights of the orthogonal basis vectors by calculating the cross-correlation coefficients between physical variables in the original input data stream. For example, after the rolling mill enters the stable operating range, the system determines the initial image matrix A according to the preset offline calibration process. When the cross-correlation coefficients fed back by the sensors in real time deviate from the reference value by more than 5%, the Gram-Schmidt orthogonalization algorithm is triggered to dynamically correct the orthogonal basis vectors. The system performs deep decoupling of transient load fluctuation data and matrix resistance characteristic data in the computational dimension, so that the final output state feature vector can accurately characterize the geometric shape and load distribution of the controlled object, avoiding high-frequency oscillation of control commands caused by parametric coupling. In the system initialization phase, the covariance matrix C is constructed by collecting N sets of synchronous original data streams under the steady-state operating conditions of the controlled object. Singular value decomposition is performed on the covariance matrix C to extract the feature value sequence. Based on the eigenvectors, the top k eigenvectors are selected as the initial orthogonal basis vectors according to the cumulative contribution rate of the eigenvalues. This ensures that the feature subspace covers more than 99% of the energy distribution characteristics of the controlled object, establishing a linear mapping relationship from the physical observation space to the feature projection space. During debugging, when the cross-correlation coefficient of the physical variables in the original input data stream deviates from the initial calibration value by 5%, the feature decoupling operation module updates the covariance matrix C using real-time data and performs orthogonalization operation again to correct the orthogonal basis vectors. It compensates for the transfer function drift caused by mechanical wear of the controlled object or environmental changes.

[0028] In complex multi-source computing processes, if sensors are blocked by cooling water or experience abnormal fluctuations due to transient electromagnetic interference, the purely logical analysis plane may output divergent results that are mathematically valid but physically infeasible. Directly issuing such instructions to the actuators could easily cause roll collisions or material breakage. To establish a safe closed loop between computational logic and physical reality, the computational logic consistency verification module constructs a physical boundary model based on the real-time power consumption parameters of the controlled object's power source. Its energy conservation judgment unit acquires the current and voltage parameters of the mill's main drive motor in real time and calculates the upper limit of the energy change rate corresponding to the physical boundary model. The physical boundary model specifically refers to a dynamic power constraint function established based on the principle of energy conservation. It establishes a linear mapping relationship between the real-time input power of the main drive motor and the mechanical energy dissipation of the actuator, and incorporates a pre-stored system static loss benchmark. It transforms the product of real-time voltage and current parameters into the allowable instantaneous energy output extremum of the controlled object in physical space, thereby determining the geometric envelope boundary of the state feature vector projected in three-dimensional space. This ensures that the compensation commands at the computational level are always controlled by the physical energy efficiency limitations of the power source. Within each computation cycle, the computational logic consistency verification module verifies the pre-defined power constraint function generated by the feature decoupling operation module. The compensation value is substituted into the physical boundary model for safety simulation. Specifically, the Euclidean norm of the state eigenvector is calculated using the 32-bit floating-point processing unit inside the actuator as the strength benchmark. This benchmark is then multiplied by a power conversion scaling factor predefined in the memory of a specific hardware platform. The strength is obtained by measuring the ratio of the actual active power of the rolling mill main motor at a rated current of 450A (1200kW) to the magnitude of the state eigenvector at that time (960 units), i.e., 1.25kW per unit of characteristic strength. This linearly maps the dimensionless eigenvector to a physically meaningful predicted energy change rate. When the predicted energy change rate obtained from the simulation is... satisfy When the system determines that the current computational logic has entered a divergent state, it triggers a forced convergence correction procedure to restrict the execution instructions to the edge of the current energy envelope. This underlying audit mechanism based on the principle of energy conservation is based on the system's physical layer security intuition and ensures the operational reliability of the controlled object under extreme data anomaly conditions.

[0029] The computational logic consistency verification module acquires the real-time voltage U and real-time current I of the drive motor during the sampling period, and calculates the upper limit of the energy change rate corresponding to the physical boundary model by combining the no-load loss benchmark provided by the offline computing unit. The safe operating envelope of the actuator under the constraints of physical laws is determined, and the state feature vector output by the feature decoupling operation module corresponds to the predicted energy change rate. Exceeding the upper limit of the rate of change of energy At that time, the state feature vector is orthogonally projected onto the edge of the physical boundary model to correct the execution command limit to the upper limit of the energy change rate. Within a defined numerical range, algorithm divergence is suppressed and control continuity is maintained to avoid overshoot in the hydraulic servo system caused by command step changes. As rolling production continues, the thermal expansion and mechanical wear of the rolls cause a slow quasi-static drift in the system stiffness characteristics. This drift, due to its extremely low rate of change, is often filtered out as background noise by the main control path. However, it accumulates into thickness deviation after long-term operation. To address the need for sensing this type of latent state, the latent state sensing module is configured to actively capture the residual signal generated during projection calculation. The system extracts the high-frequency residuals that are removed during the feature subspace projection process, and uses the residual information mining unit to perform frequency domain decomposition of the residual signal based on wavelet transform, decomposing it into high-frequency pulse features and low-frequency drift features. When the low-frequency drift features show a continuous monotonically increasing or decreasing trend and the slope remains within the preset threshold range, the system identifies the displacement deviation of the controlled object due to thermal response and generates a depth adjustment command for the feature subspace mapping logic. By correcting the offset of the mapping matrix A, the system achieves adaptive compensation for alignment static drift, enabling the system to continuously maintain the fidelity of the physical environment evolution by utilizing the information overflow during the calculation process without adding additional temperature measurement hardware.

[0030] To further eliminate the interference of implicit hardness gradients within the material on prediction accuracy, the system also integrates an active micro-excitation detection unit. This unit injects a sinusoidal perturbation signal into the response execution loop of the controlled object. The frequency of this perturbation signal is set between 50Hz and 100Hz, and the amplitude is limited to less than 2. Within a small range to ensure that normal production accuracy is not affected, the feature decoupling calculation module synchronously extracts the cross-correlation coefficient between the response signal and the sinusoidal disturbance signal in the original input data stream, thereby inverting the instantaneous dynamic stiffness characteristics of the controlled object. The specific inversion logic is as follows: the reference dynamic stiffness constant is preset to 1000 kN / mm in the controller program. When the real-time extracted cross-correlation coefficient is within the range of 0.92 to 1.0, the stiffness correction is determined to be 0; when the cross-correlation coefficient decreases by 0.05 units, an attenuation compensation of 45 kN / mm is subtracted from the reference dynamic stiffness constant. Through this stepwise decreasing calculation based on the deviation of the cross-correlation coefficient, the statistical correlation is transformed into an instantaneous dynamic stiffness characteristic value reflecting the change in the hardness of the physical structure, and this characteristic is used as a correction factor and input into the feature subspace mapping logic in real time. Through this active detection mechanism, the asynchronous compensation calculation unit can identify the hard points of the material that are about to enter the deformation zone, and Based on the abrupt change characteristics of the cross-correlation coefficient, the prediction step size of the estimated roll gap compensation value is reduced in advance, thereby effectively suppressing transient disturbances caused by material non-uniformity and improving the thickness uniformity of the finished strip steel at the surface scale. The active micro-excitation detection unit injects a sinusoidal disturbance signal with a frequency of 50Hz to 100Hz and an amplitude of less than 2μm into the response execution loop of the controlled object. The selected frequency avoids the inherent mechanical resonance frequency of the frame of the controlled object. Without causing mechanical resonance, the instantaneous dynamic stiffness characteristics of the material are extracted through the response signal. The system synchronously obtains the cross-correlation coefficient γ of the original input data stream response signal and the sinusoidal disturbance signal. The hardness gradient of the material entering the deformation zone is identified based on the instantaneous change slope of the cross-correlation coefficient γ. When the deviation of the cross-correlation coefficient γ exceeds the preset threshold, the asynchronous compensation calculation unit reduces the prediction step size from 5ms to 1ms, so that the compensation increment update frequency is adapted to the closed-loop control bandwidth of the hydraulic servo system, thereby improving the accuracy of material non-uniformity disturbance suppression.

[0031] Example 1: In the cold continuous rolling production of ultra-thin strip steel with a thickness of 0.2mm, the system faces a rapid pressure signal with a sampling period of 1ms and an exit thickness lag signal with a sampling update period of 100ms. Due to hardness fluctuations caused by component segregation within the material, and because these fluctuations cannot be obtained by thickness measurement before entering the deformation zone of the rolling mill, the control model generates an 80ms control blind zone, inducing periodic thickness deviation ripples on the plate surface; the data input interface acquires a first-type rapid variable with a sampling period of 1ms. With a second type of slow variable with an update period of 100ms Synchronous arbitrator with first-class fast variables Based on the time series, for the second type of slow variables The discretization encapsulation based on the zero-order hold is performed. The feature decoupling operation module receives the processed time-series feature vector and uses pre-calibrated orthogonal basis vectors. Execution based on formula The projection operation separates the transient pressure fluctuation component from the composite dimension that includes mechanical stiffness and tension interference; among which, For state feature vectors, For time series feature vectors, These are orthogonal basis vectors.

[0032] The active micro-excitation detection unit injects an amplitude of 1.5 at a frequency of 75Hz into the hydraulic servo actuator. sinusoidal disturbance signal The feature decoupling operation module extracts the response signal. With sinusoidal disturbance signal The cross-correlation coefficient γ is used by the asynchronous compensation calculation unit to invert the instantaneous dynamic stiffness of the material based on the change of the cross-correlation coefficient γ. When the deviation from the baseline value is 8%, the prediction step size of the estimated compensation value is compressed from 5ms to 1ms. The logic consistency verification module synchronously obtains the real-time voltage U and real-time current I of the drive motor and calculates the upper limit of the energy change rate. The energy envelope is used as the audit threshold constraint compensation instruction; where U is the real-time voltage in V and I is the real-time current in A. The upper limit of the energy change rate; the discretized data packet provided by the synchronous arbitrator provides a unified timing reference for the feature decoupling operation module, enabling the projection operation to accurately separate the transient pressure fluctuation component reflecting the material characteristics based on the elimination of timing misalignment. The separation of this component further improves the signal-to-noise ratio of the cross-correlation coefficient γ, making the instantaneous dynamic stiffness change captured by the active micro-excitation detection unit physically deterministic. By changing the calculation boundary of the compensation logic, the system transforms the originally uncontrollable material jump into a predictable logical correction. The hard point characteristics inside the material are identified before entering the deformation zone, and the compensation command achieves high-frequency correction within the energy conservation boundary. The thickness tolerance of the finished strip steel is stabilized within the specified range. Within.

[0033] Example 2: The experimental platform includes a physical simulation device simulating a 2050mm hot strip mill. The response frequency of the hydraulic servo system is set to 500Hz, and the sampling accuracy of the pressure transmitter is set to 0.05%. The raw input data stream is acquired from the real-time signal generated by the sensor array of the simulation device, with a sampling period of 1ms. The technical consideration for this parameter value is to achieve a balance between the accuracy of capturing the transient step response of the hydraulic system and the computational load. To avoid signal aliasing when analyzing vibration modes above 100Hz, the sampling frequency is determined to be 1000Hz according to the sampling law. The raw input data stream is superimposed with Gaussian white noise with a signal-to-noise ratio of 20dB and power frequency interference harmonics with a frequency of 50Hz. The experiment is divided into a control group using a linear summation calculation model and a sample group using the technical solution of this invention. The control group handles the first type of fast variables with asynchronous sampling frequencies. With the second type of slow variable During the slow variable update interval, a computational step is generated. The measured output roll gap compensation increment exhibits a stepped distribution, with a root mean square error of 5.2 μm. The sample group undergoes time-scale alignment via the data input interface, and orthogonal basis vectors are applied in the feature decoupling operation module. Perform projection operation on the time series feature vector, the calculation formula is as follows: ,in, For state feature vectors, The temporal feature vector in the original input data stream, For the first in the characteristic subspace Based on orthogonal basis vectors, experimental data shows that, in an environment containing 20dB of noise, the transient components separated by projection operations effectively suppressed the interference signal, reducing the root mean square value of the compensation error to 1.25. To verify the adaptability to changes in material hardness gradients, the experiment constructed a strength gradient control system by replacing billets with different materials. Three groups of billets with material uniformity of 95%, 85%, and 70% were selected for continuous rolling tests. Under the condition of 95% material uniformity, the thickness deviation of the sample group remained stable at 0.8 mm. Nearby, when the material uniformity drops to 85%, the active micro-excitation detection unit injects an amplitude of 1.5 at a frequency of 75Hz into the actuator. sinusoidal disturbance signal The feature decoupling operation module extracts the response signal in real time. With sinusoidal disturbance signal The cross-correlation coefficient γ is calculated using the following formula: ,in, To respond to the measured value of the signal, The input value is a sinusoidal disturbance signal. To adjust the mean of the corresponding signal, monitoring revealed that the cross-correlation coefficient γ decreased from the initial 0.92 to 0.78. The asynchronous compensation calculation unit then automatically compressed the prediction step size from 5ms to 1ms. The thickness fluctuation caused by the sudden change in material hardness recovered to a steady state within 15.2ms, and the deviation of the control group under the same operating conditions increased to 8.6. .

[0034] The computational logic consistency verification module operates under different efficiency coefficients η. When the efficiency coefficient η is set to 0.85, the system maintains convergence under heavy load jump conditions. When the drive motor voltage U is 600V and the current I instantaneously rises to 450A, the calculated predicted energy change rate... It is 225kW, triggered based on the formula The safety audit mechanism restricts output commands to within the motor overload curve, where U is the real-time voltage in V, I is the real-time current in A, and η is the efficiency coefficient. To prevent hardware damage due to logic overcompensation, the energy change rate was set to an upper limit. However, when the efficiency coefficient η was set to 0.99, the system exhibited calculation divergence after 30.5ms of operation. Ten sets of key state comparison data collected in the experiment showed that the thickness tolerance of the sample remained within a certain range as the input load fluctuation range increased from 100kN to 500kN. Within this range, its response latency is reduced by more than 65% compared to the linear processing method, and it exhibits consistent convergence characteristics in repeated experiments. The coordinated operation of asynchronous feature space projection and physical energy boundary verification eliminates logical breaks caused by time misalignment of heterogeneous data.

[0035] Example 3: This example combines Figures 1 to 2 Description of analysis systems used for complex multi-source computing, such as Figure 1As shown, the analysis system for complex multi-source computation mainly consists of a data input interface, a feature decoupling operation module, an implicit state perception module, and a computational logic consistency verification module. These modules form a closed-loop control logic through specific data flow directions. The data input interface acquires the raw input data stream characterizing the multi-source coupled operation of the controlled object and transmits it to the feature decoupling operation module. This module is configured to establish feature subspace mapping logic. After eliminating nonlinear interference and extracting state feature vectors, it sends the state feature vectors to the computational logic consistency verification module and simultaneously transmits the residual signals generated during computation to the implicit state perception module. The implicit state perception module captures the residual signals and extracts associated sensing features to achieve adaptive compensation for alignment static drift deviation. Based on this, it generates a depth adjustment command and feeds it back to the feature decoupling operation module to adjust its mapping parameters. The computational logic consistency verification module, based on the constructed physical boundary model and energy envelope, performs forced convergence correction on the received state feature vectors and finally outputs the corrected execution command. This module also maintains a connection with the implicit state perception module to obtain auxiliary correction information.

[0036] like Figure 2 As shown, the system is integrated into a specific hardware platform and physically connected to a 1580mm hot-rolled strip steel production line, which serves as the controlled environment. The controlled environment contains a sensor array signal source and an actuator. The real-time physical quantities collected by the sensor array signal source form a raw input data stream, which is transmitted to the data input interface of the specific hardware platform. The specific hardware platform is equipped with a data input interface, a feature decoupling operation module, a latent state perception module, and a computational logic consistency verification module. After these modules work together, the computational logic consistency verification module outputs a corrected execution instruction and feeds it back to the actuator in the controlled environment to adjust the production process. In addition, the architecture also includes an auxiliary computing environment, which contains an offline computing unit. This offline computing unit is responsible for calculating and providing a static loss benchmark to the computational logic consistency verification module in the specific hardware platform, thereby ensuring the system's computational accuracy and physical boundary safety under complex operating conditions.

[0037] Example 4: During the cold start phase, the system acquires the reference input vector of the rolling mill under no-load idling conditions, which consists of the main motor speed signal with a sampling period of 1ms and the laser velocity measurement signal at the stand exit with a sampling period of 100ms. The system constructs a covariance matrix C using N sets of pre-stored sample data in the memory, and applies the singular value decomposition algorithm to extract the first n eigenvectors of the covariance matrix C as the initial orthogonal basis vectors. An initial mapping matrix A is established to cover more than 99% of the energy distribution characteristics of the controlled object, where N is the number of sample groups, C is the covariance matrix, and n is the dimension of the feature space. Let A be the i-th orthogonal basis vector and A be the initial mapping matrix; the data input interface receives the analog voltage signal from the pressure sensor and uses the analog-to-digital converter to convert the signal into a 16-bit wide digital fast variable. The synchronous arbitrator determines the second type of slow variable within the timer interrupt period. The update flag state is adjusted. When the flag is 0, the synchronous arbiter reads the buffered data from the data register at the previous sampling time and performs linear hold based on the zero-order hold. When the flag is 1, the synchronous arbiter triggers a data push instruction to push the newly arrived slow variable onto the stack. Replace the old cache and include it as a component of the current temporal feature vector, where, For fast variables, For slow variables, the ordered execution steps of the synchronous arbiter provide vector inputs with defined physical definitions, eliminating computational dead zones caused by data sampling frequency mismatch.

[0038] When the asynchronous compensation calculation unit detects the transient impact caused by the billet head entering the deformation zone, it calculates the impact based on the response frequency of the hydraulic servo system. Determine the prediction step size of the compensation command Predicting step size The calculation follows the formula below: ,in, The prediction step size is in seconds (s). The closed-loop control bandwidth of the hydraulic servo system is expressed in Hz; when the closed-loop control bandwidth of the unit is 500Hz, the prediction step size is calculated. The upper limit is 1ms. The system reduces the prediction compensation logic step from the normal 5ms step to 1ms step, so that the compensation increment is within the effective frequency response range of the physical actuator, suppressing the high-frequency oscillation of the hydraulic system caused by computational overshoot. After the system runs continuously for 24 hours, it generates quasi-static drift. The implicit state perception module monitors the residual signal generated by the projection operation. The trend of change, when the residual signal When the mean value of the low-frequency component deviates from the preset zero threshold by 12%, the computational logic consistency verification module triggers online recalibration of the physical boundary model. The computational logic consistency verification module corrects the efficiency coefficient η based on the no-load loss current of the main drive motor at the current speed, ensuring that the energy envelope determined by the physical boundary model adapts to the aging state of the hardware. When the sensor is subjected to electromagnetic group pulse interference, the system uses the energy envelope determined by the physical boundary model to constrain the predicted compensation amount, controlling the deviation of the compensation command within the elastic deformation threshold of the controlled object. The residual signal is η, and the efficiency coefficient is η. The system achieves stable operation of the analysis system under complex working conditions through the forced constraint of the physical energy boundary.

[0039] Example 5: When the system is deployed on different specifications of unit hardware platforms, the system performs an on-site calibration process by collecting the no-load current vector of the motor of the controlled object under zero load. With bus voltage vector A no-load loss benchmark is established. N sets of operating condition data samples are recorded using an offline computing unit under a preset speed gradient. The correlation curve between no-load power consumption and speed is fitted using the least squares method to determine the efficiency coefficient η in the physical boundary model. The system corrects the efficiency coefficient η based on the calculated real-time no-load power consumption residual, so that the value of the efficiency coefficient η matches the mechanical transmission loss and electromagnetic conversion efficiency of a specific hardware platform. This procedure provides an accurate static loss benchmark for subsequent energy change rate calculations.

[0040] In application scenarios where impedance mismatch of the signal source is caused by sensor replacement, the data input interface executes a signal normalization procedure, and the synchronous arbiter detects the first type of fast variable within a 1ms sampling period. The voltage change slope is used to identify signal jumps, mapping the input level to a dimensionless characteristic range of 0 to 1, and based on the hysteresis time constant of the controlled object's actuator. Adjust the reconstruction window size of the zero-order hold to make the second type of slow variable... The timing characteristics are matched with the physical response speed of the controlled object, ensuring that the heterogeneous signals entering the feature decoupling operation module are in a unified dimensional scale and dynamic response range. This process eliminates the prediction calculation offset caused by hardware link differences.

[0041] Example 6: In the deployment procedure of a 1580mm hot-rolled strip steel production line, the system collects 2000 sets of raw input data streams from the controlled object under no-load operation. A parametric benchmark is established, and the synchronous arbiter extracts a physical vector containing the pressing position signal and the main motor speed signal within each 1ms sampling period. The feature decoupling operation module uses principal component analysis to calculate the covariance matrix C of this dataset, and obtains the eigenvalue sequence by solving the eigenvalue equation. The dimension k of the feature space is determined according to the cumulative contribution rate criterion, and the corresponding k feature vectors are selected as the initial orthogonal basis vectors. This allows the initial mapping matrix A to cover more than 99% of the energy distribution characteristics of the controlled object. This procedure establishes a quantization mapping path from the physical space to the orthogonal feature subspace.

[0042] When determining the working boundary of the computational logic consistency verification module, the system executes a load response calibration program. This involves applying a 500kN pulse load signal to the hydraulic servo system and recording the real-time power consumption changes of the power source. The computational logic consistency verification module then statistically predicts the energy change rate within a 0.5s observation period. Compared with the measured rate of energy change The difference distribution is such that when the mean deviation of the difference is less than 3.5%, the current conversion coefficient is determined as the working reference for the efficiency coefficient η. The active micro-excitation detection unit is based on the mechanical resonant frequency of the frame. Adjust the frequency of the sinusoidal disturbance signal so that the frequency of the disturbance signal is at the mechanical resonant frequency of the frame. Within the range of 0.8 to 0.9 times, in order to avoid the mechanical resonance zone and capture the instantaneous dynamic stiffness of the material, this parameter calibration method enables the asynchronous compensation calculation unit to make prediction step corrections based on the inherent physical response characteristics of the frame under the switching conditions of billets of different thicknesses, and limit the overshoot of the system response to sudden load changes to within 1.5%.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An analysis system for complex multi-source computing, characterized in that, The system includes a data input interface, a feature decoupling operation module, a computational logic consistency verification module, and a implicit state awareness module. The data input interface is used to acquire the raw input data stream characterizing the multi-source coupled operation of the controlled object. This raw input data stream receives and aligns real-time data from the hydraulic servo, thickness gauge and load sensor, and includes rolling force, roll speed, material prediction and initial thickness. The feature decoupling operation module, connected to the data input interface, is configured to establish feature subspace mapping logic. It projects the original input data stream onto a non-interfering feature subspace through multiple pre-calibrated orthogonal basis vectors to eliminate nonlinear interference between feature components of different dimensions and extract the decoupled state feature vector. The computational logic consistency verification module, connected to the feature decoupling operation module, is configured to construct a physical boundary model based on the real-time power consumption parameters of the power source of the controlled object, and use the energy envelope determined by the physical boundary model as the audit threshold of the state feature vector. When the calculated output of the state feature vector exceeds the audit threshold, forced convergence correction is performed. The implicit state perception module is connected to the feature decoupling operation module and the computation logic consistency verification module, respectively. It is configured to actively capture the residual signal generated in the projection operation, extract the associated perception features by analyzing the distribution characteristics of the residual signal, and generate a depth adjustment instruction for the feature subspace mapping logic based on the associated perception features, so as to realize the adaptive compensation of the quasi-static drift deviation of the controlled object by the computation model. When extracting state feature vectors, the feature decoupling module follows the following calculation rules: ,in, For state feature vectors, The temporal feature vector in the original input data stream, Let be the i-th orthogonal basis vector in the feature subspace; the feature decoupling operation module dynamically adjusts the projection weights of the orthogonal basis vectors based on the cross-correlation coefficients between physical variables in the original input data stream, thereby deeply decoupling the transient load fluctuation data and the matrix resistance characteristic data in the computational dimension; The computational logic consistency verification module includes an energy conservation determination unit. This unit is configured to construct a physical boundary model based on the energy conservation principle, using the real-time current and voltage parameters of the controlled object's power source, and to calculate the upper limit of the energy change rate corresponding to the physical boundary model. The computational logic consistency verification module substitutes the predicted compensation amount corresponding to the state feature vector into the physical boundary model for safety simulation. When the simulation yields the predicted energy change rate... satisfy At this time, the computational logic consistency verification module determines that the current computational logic has entered a divergent state; The implicit state perception module includes a residual information mining unit. The residual information mining unit is configured to perform frequency domain decomposition of the residual signal based on wavelet transform to extract high-frequency pulse features and low-frequency drift features from the residual signal. Based on the monotonic change trend of the low-frequency drift features, the implicit state perception module identifies the quasi-static displacement deviation of the controlled object caused by the thermal response of the execution component, and based on the amplitude distribution of the high-frequency pulse features, identifies the transient disturbances inside the controlled object caused by the non-uniformity of the input matrix.

2. The analysis system for complex multi-source computing according to claim 1, characterized in that, The system also includes an active micro-excitation detection unit; the active micro-excitation detection unit is used to inject a sinusoidal disturbance signal with a frequency of 50Hz to 100Hz and an amplitude of less than 2μm into the response execution loop of the controlled object, and simultaneously extract the cross-correlation coefficient between the response signal and the sinusoidal disturbance signal in the original input data stream; The feature decoupling operation module inverts the instantaneous dynamic stiffness characteristics of the controlled object based on the cross-correlation coefficient, and inputs the instantaneous dynamic stiffness characteristics as a correction factor into the feature subspace mapping logic.

3. The analysis system for complex multi-source computing according to claim 1, characterized in that, The data input interface is configured to perform data translation processing, converting the physical quantity measurement signals of the controlled object into discrete data packets; the data input interface also includes a synchronization arbiter for time-stamp alignment of input signals from multiple sampling sources, ensuring that the original input data stream entering the feature decoupling operation module has a unified logical reference in the time dimension.

4. The analysis system for complex multi-source computing according to claim 1, characterized in that, Orthogonal basis vectors are generated through a combination of offline calibration and online updating. The feature decoupling operation module is configured to periodically correct the orthogonal basis vectors using the Gram-Schmidt orthogonalization algorithm based on the accompanying sensing features output by the latent state perception module, in order to compensate for the mismatch of the computational model caused by the evolution of the external environment.

5. The analysis system for complex multi-source computing according to claim 1, characterized in that, The state feature vector includes feature quantities used to characterize the load distribution of the system and feature quantities used to characterize the geometric position of the controlled object; The feature decoupling operation module achieves orthogonality between the feature quantities of the system load allocation and the feature quantities of the geometric position of the controlled object in the calculation dimension through feature subspace projection, thereby eliminating the reaction force influence of the position adjustment action on the stability of load allocation.

6. The analysis system for complex multi-source computing according to claim 1, characterized in that, The computational logic consistency verification module also includes redundant decision units; The redundant decision unit is configured to identify computational anomalies caused by algorithm overflow or data singularities by comparing the output consistency of multiple parallel computing paths after taking into account the perturbation components injected by the active micro-stimulation detection unit. When the output deviation of multiple parallel computing paths exceeds the preset tolerance threshold, the redundant decision unit forcibly calls the safety preset value based on historical steady-state data.

7. The analysis system for complex multi-source computing according to claim 1, characterized in that, The system adopts a distributed computing architecture, in which the feature decoupling operation module is deployed on the real-time processing plane and the implicit state perception module is deployed on the policy analysis plane. The policy analysis plane generates system operation status evaluation data based on the long-term statistical characteristics of the residual signal and sends depth adjustment instructions to the real-time processing plane to dynamically adjust the projection depth parameters of the orthogonal basis vectors.

Images