Feed processing process closed-loop control and state monitoring system based on digital twinning

Abstract

The application relates to the technical field of industrial control systems, and particularly discloses a feed processing process closed-loop control and state monitoring system based on digital twinning, which generates compensation state data by reconstructing the state of data loss caused by communication delay through adaptive unscented Kalman filtering; the compensation state data is input into a rolling horizon estimation to predict the future load change trend and output a predicted state sequence; the predicted state sequence is used to drive a digital twinning simulation process based on smooth particle dynamics to iteratively optimize and output an optimal feeding amount set value; residual analysis is performed on the measured data and the predicted state sequence to online correct filter parameters and output a prediction deviation; and a control instruction is triggered and issued through fuzzy reasoning according to the current change rate and the prediction deviation; the application realizes real-time dynamic suppression of the load fluctuation of a granulator, reduces ton electricity consumption and the risk of machine blockage, and improves the intelligent control level of the feed processing process.

Description

Technical Field

[0001] This invention relates to the field of industrial control system technology, specifically to a closed-loop control and status monitoring system for feed processing based on digital twins. Background Technology

[0002] In recent years, digital twin technology has been gradually applied in industrial control systems, simulating and optimizing physical processes by constructing virtual models. In feed processing industrial control systems, the pellet mill, as a core piece of equipment, directly affects production capacity and product quality due to load fluctuations. Existing industrial control systems typically employ programmable logic controllers combined with proportional-integral-derivative (PID) control strategies, adjusting the feed rate based on current feedback.

[0003] In the feed pelleting process, existing industrial control systems suffer from technical problems such as control commands lagging behind the dynamic changes of physical entities due to the nondeterministic delay of communication networks and data packet loss. At the same time, the time consumed by digital twin simulation calculations is mismatched with the real-time control requirements, and the lack of an effective online correction mechanism causes the model accuracy to decrease with raw material fluctuations and equipment wear. Ultimately, this results in technical problems such as insufficient ability to suppress sudden load changes, increased power consumption per ton, and increased risk of machine blockage. Summary of the Invention

[0004] The purpose of this invention is to provide a closed-loop control and status monitoring system for feed processing based on digital twins, in order to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A digital twin-based closed-loop control and status monitoring system for feed processing includes: The data acquisition and compensation module collects current and vibration data in real time during the operation of the pellet mill. It uses an adaptive unscented Kalman filter algorithm to reconstruct the state of the missing time series caused by communication delay or data packet loss, and generates compensated state data that matches the current moment of the physical entity. The state prediction module uses the compensation state data as initial conditions to perform rolling time-domain estimation, recursively predicts the load change trend within a preset time window, and outputs the predicted state sequence. The simulation optimization module uses the predicted state sequence as boundary conditions to drive the digital twin simulation process. It quickly iterates through lightweight numerical solutions to obtain the load response characteristics under different feed rate adjustment schemes, and outputs the optimal feed rate setting value with the goal of minimizing power consumption per ton and minimizing the risk of machine blockage. The residual correction module performs residual analysis on the actual feedback measured data and the predicted state sequence, and uses rolling time-domain estimation to correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter online, while outputting the prediction bias. The threshold control module monitors the current rate of change and prediction deviation. If the current exceeds the preset threshold, the optimal feed rate setting is sent to the feeding actuator through the real-time communication link. Otherwise, the current control parameters are maintained, and the corrected parameters are used for state reconstruction in subsequent cycles.

[0006] As a further aspect of the present invention: the generation of compensation state data that matches the current moment of the physical entity specifically includes: Variational mode decomposition is performed on current and vibration data from multiple consecutive sampling periods before the communication delay occurs, and multiple intrinsic mode components characterizing different frequency features are output. State-space equations are established for each intrinsic modal component, and adaptive unscented Kalman filtering is applied for recursive estimation. During the recursive process, the process noise covariance matrix is ​​adjusted in real time according to the instantaneous frequency of each modal component. The recursive estimation results of each intrinsic modal component are superimposed in the time domain to output compensated state data that matches the current moment of the physical entity.

[0007] As a further aspect of the present invention: the output predicted state sequence specifically includes: The compensation state data is decomposed into non-stationary components representing load trends and stationary components representing fluctuations by performing empirical wavelet transform. Different cost functions are constructed for non-stationary and stationary components respectively. The rolling optimization solution is performed using historical data within a fixed time window, and the predicted values ​​of each component within the future time window are output. The predicted values ​​of each component are reconstructed and superimposed, and a regularization term is introduced to constrain the rate of change of the amplitude of the control action, and the final predicted state sequence is output.

[0008] As a further aspect of the present invention: the output of the predicted values ​​of each component within the future time window specifically includes: For stationary components, a first cost function is constructed with the objective of minimizing the sum of squared prediction errors. The first parameter obtained by solving the first cost function is used as prior information, and a second cost function is constructed for non-stationary components with the objective of minimizing local fluctuation amplitude. The first cost function and the second cost function are solved iteratively and alternately. The output is the undetermined coefficients that make the two cost functions converge simultaneously. The undetermined coefficients are used to form the predicted values ​​of each component within the future time window.

[0009] As a further aspect of the present invention: the output process of the optimal feed rate setting value is as follows: The predicted state sequence is assigned as the initial boundary to the finite set of material flow points in the granulation channel. Each finite set of material flows carries the velocity and temperature attributes at the current moment. The relative positions of each particle are updated by adjusting the candidate values ​​based on the feed rate, and the momentum and energy exchange between particles are calculated using smooth particle dynamics, and the updated particle distribution state is output. The cumulative force of each mass point in the contact area of ​​the pellet mill roller is statistically analyzed and mapped to the load response characteristics at the current moment. Then, after traversing all candidate values ​​for feed rate adjustment, the optimal feed rate setting value is output.

[0010] As a further aspect of the present invention: the updated particle distribution state output specifically includes: Based on the current position of each particle and the preset smooth length, search all neighboring particles of each particle in the support domain, and calculate the kernel function value of each neighboring particle with respect to the current particle. The pressure gradient and viscous force between each neighboring particle and the current particle are calculated based on the kernel function value, and then the velocity and acceleration of the current particle are updated in the next moment. Calculate the frictional heat generated by the relative motion between the current particle and its neighboring particles based on the updated velocity and acceleration, and add the frictional heat to the internal energy property of the current particle. The updated velocity and internal energy properties of all particles are summarized, and particles that exceed the boundary are removed according to the granulation die hole constraints. The updated particle distribution is then output.

[0011] As a further aspect of the present invention: the method of using rolling time-domain estimation to online correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter, while simultaneously outputting the prediction bias, specifically includes: Subtract the measured data at the current moment from the corresponding predicted state sequence to output a residual sequence containing the estimation bias; Input the residual sequences from multiple recent time points into a sliding window, calculate the mean and variance of the residuals within the window, and determine whether there is a systematic shift at the current time point based on the mean. The calculated residual variance is compared with the theoretical innovation covariance, and the process noise covariance matrix of the adaptive unscented Kalman filter is adjusted according to the comparison results. The adjusted process noise covariance matrix and residual mean are used as corrections to the state transition parameters, while the residual value at the current time is used as the prediction bias output.

[0012] As a further aspect of the present invention: the monitoring of the current rate of change and prediction deviation specifically includes: The current value at the current moment is calculated by difference from the current value at the previous moment, and the current rate of change is output. At the same time, the prediction deviation is compared with the preset deviation tolerance range, and the comparison result is output. The current rate of change of current and the comparison result are input into the fuzzy inference unit, and the confidence level of the control action trigger at the current moment is calculated according to the predefined fuzzy rule base. The control action trigger confidence is compared with the preset action execution threshold. If the confidence exceeds the threshold, a trigger signal is generated, and the optimal feeding amount setting value is sent to the feeding actuator through the real-time communication link. If the confidence level does not exceed the threshold, the current control parameters are maintained, and the process noise covariance matrix and state transition parameters corrected at the current time are stored in the historical database for state reconstruction in subsequent cycles.

[0013] As a further aspect of the present invention: the step of calculating the control action trigger confidence level at the current moment based on a predefined fuzzy rule base specifically includes: The current rate of change is mapped to the membership values ​​of multiple fuzzy sets according to a preset membership function, and the comparison results are mapped to the membership values ​​of multiple deviation states according to a preset membership function. The membership values ​​of the current change rate and the deviation state are used as inputs to the antecedent of the fuzzy rule. All rules in the fuzzy rule base are traversed, and the rule strength of each rule is output as the consequent. Take the maximum value among all rule strengths as the comprehensive reasoning result, and use the centroid method to defuzzify the comprehensive reasoning result and the membership function of the rule consequent, and output the control action trigger confidence at the current moment.

[0014] The beneficial effects of this invention are: (1) This invention combines adaptive unscented Kalman filtering with variational mode decomposition to reconstruct the state of missing current and vibration data caused by communication delay or packet loss, and generates compensated state data synchronized with physical entities, which effectively solves the problem of control command lag caused by the nondeterministic delay of the industrial field network; combined with fuzzy inference to evaluate the confidence of current change rate and prediction deviation, the control action is only triggered when necessary, avoiding wear of actuators and system oscillation caused by frequent adjustment, and improving the dynamic response capability and operational reliability of the pellet mill under raw material fluctuation and load change conditions.

[0015] (2) This invention uses the predicted state sequence as the boundary condition to drive the digital twin simulation process based on smooth particle dynamics. It accurately simulates the flow, extrusion and frictional heat generation process of materials in the granulation channel through the momentum and energy exchange of a finite set of particles, and realizes the rapid iterative optimization of the load response characteristics under different feed rate adjustment schemes. At the same time, it outputs the optimal feed rate setting value with the goal of minimizing the power consumption per ton and minimizing the risk of machine blockage. It also corrects the filter parameters online through rolling time domain estimation, which can follow the changes in equipment wear and raw material characteristics, effectively reduce the power consumption per ton and reduce unplanned downtime while ensuring product quality. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a closed-loop control and status monitoring system for feed processing based on digital twins, comprising: The data acquisition and compensation module collects current and vibration data in real time during the operation of the pellet mill. It uses an adaptive unscented Kalman filter algorithm to reconstruct the state of the missing time series caused by communication delay or data packet loss, and generates compensated state data that matches the current moment of the physical entity. The state prediction module uses the compensation state data as initial conditions to perform rolling time-domain estimation, recursively predicts the load change trend within a preset time window, and outputs the predicted state sequence. The simulation optimization module uses the predicted state sequence as boundary conditions to drive the digital twin simulation process. It quickly iterates through lightweight numerical solutions to obtain the load response characteristics under different feed rate adjustment schemes, and outputs the optimal feed rate setting value with the goal of minimizing power consumption per ton and minimizing the risk of machine blockage. The residual correction module performs residual analysis on the actual feedback measured data and the predicted state sequence, and uses rolling time-domain estimation to correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter online, while outputting the prediction bias. The threshold control module monitors the current rate of change and prediction deviation. If the current exceeds the preset threshold, the optimal feed rate setting is sent to the feeding actuator through the real-time communication link. Otherwise, the current control parameters are maintained, and the corrected parameters are used for state reconstruction in subsequent cycles.

[0020] In the data acquisition and compensation module, current and vibration data during the pellet mill's operation are collected in real time. An adaptive unscented Kalman filter algorithm is used to reconstruct the state of the time series data missing due to communication delays or data loss, generating compensated state data that matches the current physical entity. Specifically, this includes: First, vibration and current signals during pellet mill operation are collected in real time at a fixed sampling frequency using a piezoelectric accelerometer installed on the pellet mill spindle bearing housing and a current transformer connected in series with the main motor power supply line. When communication delays or data packet loss occur in the industrial Ethernet, current and vibration data from multiple consecutive sampling cycles before the communication delay occurs are extracted as a base sequence. This base sequence is then input into a variational mode decomposition algorithm for processing. This algorithm iteratively solves a variational problem to decompose the original signal into a preset number of intrinsic mode components. Specifically, each intrinsic mode component represents the oscillation component of the original signal near different center frequencies. Its center frequency and bandwidth are adaptively determined by the variational mode decomposition algorithm during the iteration process, ultimately outputting multiple intrinsic mode components with different frequency characteristics.

[0021] Secondly, for each intrinsic mode component output by variational mode decomposition, a state-space equation characterizing its dynamic evolution is established. This state-space equation uses the amplitude of the current component as the state variable, and the linear relationship between the amplitude at the previous time step and the amplitude at the current time step as the state transition relationship. The observation noise is set to Gaussian white noise. An adaptive unscented Kalman filter algorithm is applied to recursively estimate each intrinsic mode component. During the recursive process, the fluctuation degree of each component relative to the historical average frequency is first calculated based on the instantaneous frequency of the current component. If the instantaneous frequency increases compared to the historical average frequency, the process noise covariance matrix is ​​increased proportionally; if the instantaneous frequency decreases, the process noise covariance matrix is ​​decreased proportionally. Through this adaptive adjustment mechanism, the filter can follow the changing characteristics of different frequency components, ultimately outputting the estimated value of each intrinsic mode component at the current time step.

[0022] Finally, the estimated values ​​of each intrinsic mode component obtained by the adaptive unscented Kalman filter recursive estimation at the current time are superimposed point-by-point in the time domain. This superposition process involves directly adding the amplitudes of all intrinsic mode components to obtain the reconstructed complete signal amplitude. This reconstructed complete signal amplitude is the compensated state data that matches the physical entity at the current time. This compensated state data is used to replace the measured data missing due to communication delays or packet loss, serving as the initial input value for subsequent rolling time-domain estimation. It is also stored in the buffer of the edge nodes for the filter recursive calculation at the next sampling time.

[0023] In the state prediction module, the compensated state data is used as the initial condition for rolling time-domain estimation, recursively predicting the load change trend within a preset time window, and outputting the predicted state sequence, specifically including: First, compensated state data reconstructed through adaptive unscented Kalman filtering is obtained. This compensated state data represents the complete time-series amplitude of the granulator current at the current moment. This compensated state data is then input into an empirical wavelet transform algorithm for processing. The empirical wavelet transform first calculates the Fourier spectrum of the compensated state data and detects local maxima on the spectrum. The spectrum is divided into multiple continuous intervals using the midpoint between two adjacent local minima as boundaries. Within each interval, an empirical scaling function and an empirical wavelet function are constructed. The scaling coefficients are obtained by performing an inner product of the original signal and the empirical scaling function, and the wavelet coefficients are obtained by performing an inner product of the original signal and the empirical wavelet function. Based on the scaling coefficients and wavelet coefficients, multiple components with different physical meanings are reconstructed. Components whose spectra are concentrated in the low-frequency region and change slowly over time are labeled as non-stationary components, while components whose spectra are distributed in the high-frequency region and fluctuate rapidly near zero are labeled as stationary components.

[0024] Secondly, for the stationary and non-stationary components output by the empirical wavelet transform decomposition, cost functions with different mathematical forms are constructed for rolling optimization. For the stationary components, a first cost function is constructed with the objective of minimizing the sum of squared prediction errors. This first cost function uses undetermined coefficients as independent variables, subtracts the linear combination of undetermined coefficients from the measured values ​​of the stationary components at each historical time within a fixed time window, and sums the squares of the differences within the time window. For the non-stationary components, a second cost function is constructed with the objective of minimizing local fluctuation amplitude. This second cost function uses undetermined coefficients as independent variables, accumulates the absolute values ​​of the differences between the predicted values ​​of the non-stationary components at adjacent time points within a fixed time window, and introduces a first-order difference operator to constrain the smoothness of the changes in the undetermined coefficients.

[0025] Then, the first cost function and the second cost function are solved iteratively and alternately. Specifically, firstly, the undetermined coefficients of the non-stationary component are fixed, and the updated values ​​of the undetermined coefficients of the stationary component are obtained by solving the first cost function as the objective. Then, the updated values ​​of the undetermined coefficients of the stationary component are used as known quantities, and the updated values ​​of the undetermined coefficients of the non-stationary component are obtained by solving the second cost function as the objective. The above alternating process is repeated until the change amplitude of the undetermined coefficients of both the stationary and non-stationary components in two adjacent iterations is less than the preset convergence threshold. The undetermined coefficients obtained after convergence are substituted into the corresponding basis functions to calculate the amplitude of the stationary component and the amplitude of the non-stationary component at each sampling time within the preset time window in the future, and the predicted value sequence of the stationary component and the predicted value sequence of the non-stationary component are output.

[0026] Finally, the predicted value sequences of the stationary components and the non-stationary components are reconstructed and superimposed point by point to obtain the predicted load current values ​​at each time point within a preset time window. During the reconstruction and superposition, a regularization term is introduced to constrain the rate of change of the control action's amplitude. Specifically, the absolute value of the difference between the optimal feed rate setpoints at two adjacent time points is calculated, and the square of this difference is multiplied by a preset regularization coefficient and then superimposed into the sum of the first and second cost functions. By adjusting the magnitude of the regularization coefficient, the drasticness of the action change is controlled, ensuring that the final output predicted state sequence not only meets the fitting accuracy requirements of the load change trend but also guarantees that the control commands generated by this predicted state sequence have a smooth execution trajectory in physical implementation.

[0027] In the simulation optimization module, the predicted state sequence is used as boundary conditions to drive the digital twin simulation process. Through lightweight numerical solutions, the load response characteristics under different feed rate adjustment schemes are obtained through rapid iteration. The optimal feed rate setpoint is output with the goals of minimizing power consumption per ton and minimizing the risk of machine blockage. Specifically, this includes: First, the current load current and temperature values ​​in the predicted state sequence are converted into the initial state of a finite set of particles representing the material flow within the granulation channel. Based on the material flow rate and temperature distribution represented by the predicted state sequence, the granulation channel is divided axially into several continuous control volumes, each mapped to a finite particle. Each finite particle is assigned a velocity attribute and a temperature attribute at the current moment. The velocity attribute is calculated based on the ratio of material flow rate to channel cross-sectional area, and the temperature attribute is directly taken from the temperature value in the predicted state sequence. All finite particles together constitute a discrete set of particles representing the material flow state. The total number of particles is preset according to the required calculation accuracy; in this embodiment, it is set to 2000.

[0028] Secondly, for each candidate value for feed rate adjustment to be evaluated, the relative positions of each particle are updated and the momentum and energy exchange between particles are calculated. The candidate values ​​for feed rate adjustment are a series of values ​​obtained by increasing or decreasing the current feed rate in fixed steps. For example, using the current feed rate as a baseline, 11 candidate values ​​are generated in steps of 0.1 tons per second, ranging from -0.5 tons per second to +0.5 tons per second. For each candidate value, the following sub-steps are performed to output the updated particle distribution: The first step is to search for all neighboring particles within the support domain for each particle based on its current position and a preset smoothing length. The smoothing length is a parameter characterizing the influence range of a particle; in this embodiment, a smoothing length is set. The value is 0.01 meters, and the support domain is defined as centered on the current mass point and extending outwards along a smooth length. A spherical region with a radius twice that of the given region. For each point mass... traverse all other particles Calculate the distance between the two. ,like Less than Then the particle Marked as a point mass The neighboring particles. Then, for each pair of neighboring particles, the kernel function value is calculated. The kernel function uses a cubic spline function, and its expression is: when hour, ; when hour, ; when hour, .

[0029] in, As a normalization constant, in this implementation, we take: This value corresponds to the one-dimensional flow approximation in three-dimensional space. The kernel function value reflects the particle... For point mass The influence weights are used for the weighted summation of subsequent physical quantities.

[0030] The second step involves calculating the pressure gradient and viscous force between each neighboring particle and the current particle based on the kernel function value, thereby updating the velocity and acceleration of the current particle at the next moment. First, the density of each particle is calculated using a summation method: ,in For point mass The mass of all particles is the same, and is taken as the total mass of the material divided by the total number of particles, which is 2000. Internal energy The pressure is obtained by multiplying the temperature attribute carried by the particle by the specific heat capacity. In this embodiment, the temperature attribute is directly used for subsequent calculations. Calculated using the ideal gas law: ,in The adiabatic index is taken as 1.4. Then, the particle density is calculated. The pressure gradient force: This force is contributed by all neighboring particles and is expressed as a negative summation term. Specifically, it is the sum of the mass of each neighboring particle multiplied by the ratio of the pressure of that particle to the square of the density of the current particle, multiplied by the kernel function gradient. The kernel function gradient is obtained by taking the derivative of the kernel function with respect to distance. For each current particle, the viscous force is contributed by all neighboring particles. First, the projection component of the relative velocity between the current particle and each neighboring particle in the direction of relative position is calculated. Then, the product of this projection component, the sound velocity of the current particle, and the average density of the current particle and its neighbors is calculated. The dynamic viscosity is taken as 0.001 Pa·s and added to the above product, then multiplied by a preset artificial viscosity coefficient. Finally, this result is multiplied by the product of the mass of the current particle and its neighbors and the kernel function gradient, and the contributions of all neighboring particles are summed to obtain the viscous force on the current particle. Adding the pressure gradient force and the viscous force gives the particle's viscous force. The net force acting on the body, divided by the mass, yields the acceleration. Therefore, the speed is updated, and the calculation expression is: Update the position, and calculate the expression as follows: The time step Take as seconds, to ensure numerical stability. Indicates update speed. Indicates the new location.

[0031] The third step is to calculate the frictional heat generated by the relative motion between the current particle and its neighboring particles based on the updated velocity and acceleration, and then add this frictional heat to the internal energy attribute of the current particle. The frictional heat originates from viscous dissipation, and its calculation formula is as follows: ; in, For point mass Within a time step, due to the interaction with neighboring particles The increase in internal energy caused by relative motion and They are point masses and dynamic viscosity, and For density, For a relative velocity vector, the squared term represents the square of the magnitude of the relative velocity. For kernel function values, The time step is the calculated value. Accumulated to a point mass Current internal energy The updated internal energy is obtained from the above, and the calculation expression is: ;in This represents the updated internal energy. The change in internal energy reflects the increase in material temperature, which will affect subsequent pressure calculations and material properties.

[0032] The fourth step involves summarizing the updated velocity and internal energy properties of all particles, while simultaneously removing particles that exceed the boundary based on the pelleting die constraints. The pelleting die is located at the end of the channel; only particles with position coordinates less than the channel length are considered to remain within the pelleting channel. For particles whose positions exceed the channel length, they are removed from the particle set, and their velocity and internal energy as they pass through the die are recorded as properties of the extruded material. After updating all particles, the current particle distribution state is obtained, including the position, velocity, internal energy, and density of each remaining particle. This state is used for iterative calculations in the next time step. For a candidate feed rate adjustment value, steps one through four are repeated until a preset simulation duration is reached (in this embodiment, the simulation duration is 2 seconds), thus obtaining the complete particle distribution evolution process under that candidate value.

[0033] Then, the cumulative force of each mass point within the contact area of ​​the pelletizing die roller is statistically analyzed. The contact area of ​​the pelletizing die roller is defined as the wedge-shaped region between the inner surface of the pellet mill ring die and the outer surface of the pressure roller, where material is squeezed through the die holes. For each simulation time step, the pressure value of all mass points within this region is statistically analyzed, and the pressure of each mass point is multiplied by the area of ​​the infinitesimal element represented by the mass point to obtain the force exerted by that mass point on the die roller. The forces exerted by all mass points are accumulated over the simulation duration to obtain the total force. The total force is divided by the simulation duration to obtain the average force, which is the load response characteristic at the current moment and is directly related to the current load of the pellet mill's main motor. For each candidate value for feed rate adjustment, a corresponding load response characteristic value can be calculated.

[0034] Finally, all candidate values ​​for feed rate adjustment are iterated to obtain a set of load response characteristic values. Based on these load response characteristic values, combined with the preset power consumption per ton objective function and the blockage risk judgment criterion, the optimal feed rate setting value is selected. The power consumption per ton objective function is defined as the ratio of pellet mill power to output, where power is calculated by multiplying the load response characteristic value by the rotational speed, and output is determined by the candidate values ​​for feed rate adjustment. The blockage risk judgment criterion is whether the load response characteristic value exceeds a preset current upper limit threshold. In this embodiment, among all candidate values ​​that do not exceed the current upper limit threshold, the candidate value for feed rate adjustment that minimizes power consumption per ton is selected as the optimal feed rate setting value. If all candidate values ​​exceed the current upper limit threshold, the candidate value with the smallest current is selected as the optimal feed rate setting value, and a blockage warning is triggered.

[0035] In the residual correction module, residual analysis is performed on the actual feedback measured data and the predicted state sequence. Rolling time-domain estimation is used to correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter online, while simultaneously outputting the prediction bias, specifically including: First, the measured data collected by the current transformer and vibration sensor at the current moment is acquired, along with the corresponding values ​​in the predicted state sequence output by the rolling time-domain estimation at the same moment. The measured data and the corresponding values ​​in the predicted state sequence are subtracted point by point, i.e., the measured data value is subtracted from the predicted state sequence value, to obtain the residual value at the current moment. The residual value at the current moment, together with the residual values ​​from multiple consecutive previous moments, are arranged in chronological order to form a residual sequence containing estimation bias information. The length of this residual sequence is preset to a fixed value; in this embodiment, it is taken as the residual values ​​of 10 consecutive sampling moments.

[0036] Next, the residual sequences from multiple recent time points are input into a pre-defined sliding window for statistical calculation. The length of the sliding window is consistent with the length of the residual sequence; in this embodiment, it is set to 10 sampling points. First, the arithmetic mean of all residual values ​​within the window is calculated, which is obtained by summing the 10 residual values ​​and dividing by 10. This residual mean reflects whether there is a systematic deviation in the current time-based filtering estimate. If the absolute value of the residual mean exceeds a preset mean threshold, such as 0.5 amperes, a systematic deviation is determined to exist. Simultaneously, the degree of fluctuation of each residual value within the window relative to the residual mean is calculated. Specifically, each residual value is subtracted from the residual mean, the square is calculated, the 10 squared values ​​are summed, and then divided by 9 to obtain the residual variance. This residual variance characterizes the dispersion of the estimation bias at the current time.

[0037] Then, the calculated residual variance is compared with the theoretical innovation covariance, and the process noise covariance matrix of the adaptive unscented Kalman filter is adjusted based on the comparison result. The theoretical innovation covariance is a value pre-calculated during the recursive process of the adaptive unscented Kalman filter, which reflects the theoretically expected fluctuation range of the deviation between the predicted and measured values ​​under the current filtering parameters. The specific comparison and adjustment methods are as follows: If the residual variance is greater than 1.2 times the theoretical innovation covariance, it indicates that the current process noise covariance matrix is ​​set too small, causing the filter to over-rely on the predicted value and under-respond to the measured data. In this case, each element of the process noise covariance matrix is ​​enlarged proportionally, with the enlargement factor being the ratio of the residual variance to the theoretical innovation covariance. If the residual variance is less than 0.8 times the theoretical innovation covariance, it indicates that the current process noise covariance matrix is ​​set too large, causing the filter to over-follow the measured data and introduce noise. In this case, each element of the process noise covariance matrix is ​​reduced proportionally, with the reduction factor being the ratio of the theoretical innovation covariance to the residual variance. If the residual variance is between 0.8 and 1.2 times the theoretical innovation covariance, the current process noise covariance matrix is ​​kept unchanged.

[0038] Finally, the adjusted process noise covariance matrix is ​​used as a new parameter for the adaptive unscented Kalman filter in the state recursion calculation for the next sampling period. Simultaneously, the mean residual calculated at the current time step is used as a correction factor for the state transition parameters. This correction factor compensates for the predicted state value at the next time step; that is, the predicted value is added to the mean residual value during state transition to eliminate the influence of systematic offset. Furthermore, the residual value at the current time step is output as the prediction bias, which is used in subsequent steps to determine whether a control action needs to be triggered. The corrected process noise covariance matrix and state transition parameters are stored in the buffer of the edge nodes. At the start of the next sampling period, the corrected parameters are read for new state reconstruction and prediction calculation.

[0039] In the threshold control module, the current rate of change of current and prediction deviation are monitored. If the preset threshold is exceeded, the optimal feed rate setting is sent to the feeding actuator via a real-time communication link. Otherwise, the current control parameters are maintained, and the corrected parameters are used for state reconstruction in subsequent cycles. Specifically, this includes: First, the current value of the granulator's main motor and the prediction deviation output from the rolling time domain estimation are acquired in real time. The current value at the current moment is then subtracted from the current value at the previous sampling moment to calculate the current rate of change. Simultaneously, the prediction deviation at the current moment is compared with a preset deviation tolerance range. This tolerance range is a numerical range with an upper and lower limit, set to -2 amperes to +2 amperes in this embodiment. If the prediction deviation falls within this range, the comparison result is considered normal; if the prediction deviation exceeds the upper limit, the comparison result is a positive deviation; if the prediction deviation is less than the lower limit, the comparison result is a negative deviation. The current rate of change and the comparison result are used as inputs for subsequent fuzzy inference.

[0040] Secondly, the current rate of change and the comparison result are input into the fuzzy inference unit for fuzzification processing. For the current rate of change, three fuzzy sets are pre-defined, named negative large, zero, and positive large. Each fuzzy set corresponds to a membership function. The membership function for negative large is a falling-edge linear function, with a membership degree of 1 when the current rate of change is less than negative 5 amps per second, and linearly decreasing to 0 between negative 5 amps per second and 0 amps per second. The membership function for zero is triangular, with its center point at 0 amps per second, and linearly changing within the range of negative 2.5 amps per second to positive 2.5 amps per second. The membership function for positive large is a rising-edge linear function, with a membership degree of 1 when the current rate of change is greater than positive 5 amps per second, and linearly increasing between 0 amps per second and positive 5 amps per second. Based on the value of the current rate of change, the membership degree values ​​of the three fuzzy sets (negative large, zero, and positive large) are calculated respectively.

[0041] Simultaneously, for the comparison results, three fuzzy sets are pre-defined, named negative deviation, normal, and positive deviation, respectively. Since the comparison results are discrete values, their membership functions adopt a single-point form: if the comparison result is negative deviation, the membership degree of the negative deviation fuzzy set is 1, and the other two are 0; if the comparison result is normal, the membership degree of the normal fuzzy set is 1, and the other two are 0; if the comparison result is positive deviation, the membership degree of the positive deviation fuzzy set is 1, and the other two are 0. The comparison results are then mapped to three membership values.

[0042] Then, the three membership values ​​of the current change rate and the three membership values ​​of the comparison result are used as inputs to the fuzzy rule antecedents, and the predefined fuzzy rule base is traversed. The fuzzy rule base consists of several fuzzy rules, each of which is in the form of "If the current change rate is A and the comparison result is B, then the control action trigger confidence is C", where A belongs to one of negative large, zero, and positive large, B belongs to one of negative deviation, normal, and positive deviation, and C belongs to one of the three fuzzy sets of low, medium, and high. In this embodiment, the fuzzy rule base contains 9 rules, specifically: if the current change rate is negatively large and the comparison result is negatively biased, the confidence level is low; if the current change rate is negatively large and the comparison result is normal, the confidence level is low; if the current change rate is negatively large and the comparison result is positively biased, the confidence level is medium; if the current change rate is zero and the comparison result is negatively biased, the confidence level is low; if the current change rate is zero and the comparison result is normal, the confidence level is low; if the current change rate is zero and the comparison result is positively biased, the confidence level is low; if the current change rate is positively large and the comparison result is negatively biased, the confidence level is medium; if the current change rate is positively large and the comparison result is normal, the confidence level is medium; if the current change rate is positively large and the comparison result is positively biased, the confidence level is high. For each rule, the smaller of the membership value of the current change rate and the membership value of the comparison result is taken as the rule strength of that rule. After traversing all 9 rules, 9 rule strengths are obtained, and each rule strength corresponds to a consequent fuzzy set.

[0043] Next, the maximum value among all rule strengths is taken as the comprehensive inference result. This comprehensive inference result corresponds to the membership value of a consequent fuzzy set. For example, if the maximum rule strength appears in a rule with a consequent of "high", then the comprehensive inference result is that the "high" fuzzy set has that rule strength value, while the membership degrees of the "low" and "medium" fuzzy sets are both 0. Then, the comprehensive inference result is defuzzified, and the centroid method is used to calculate the precise value of the control action trigger confidence. The centroid method is calculated as follows: the membership function of the consequent fuzzy set is discretized into several points on the horizontal axis, the horizontal coordinate value of each point is multiplied by the membership degree of that point, and then the sum is divided by the sum of the membership degrees of all points. In this embodiment, the confidence range corresponding to the consequent fuzzy set "low" is 0 to 0.3, the range corresponding to "medium" is 0.3 to 0.7, and the range corresponding to "high" is 0.7 to 1.0. Based on the rule strength value of the "high" fuzzy set in the comprehensive reasoning result, the centroid position is calculated, and a value between 0 and 1 is obtained, which is the confidence level of the control action trigger at the current moment.

[0044] Finally, the calculated control action trigger confidence score is compared with a preset action execution threshold. In this embodiment, the action execution threshold is set to 0.8. If the control action trigger confidence score is greater than or equal to 0.8, the control action trigger condition is met, and a trigger signal is generated. This trigger signal sends the optimal feed rate setting value to the feeding actuator via a real-time communication link, specifically by writing the setting value into the control register of the feeding frequency converter via industrial Ethernet. If the control action trigger confidence score is less than 0.8, the trigger condition is not met, and the current control parameters remain unchanged, i.e., no new setting value is sent to the feeding actuator. Simultaneously, the corrected process noise covariance matrix and state transition parameters at the current moment are stored in a historical database located in the non-volatile storage area of ​​the edge node. These stored parameters will be read at the beginning of the next sampling period for a new round of state reconstruction and prediction calculation.

[0045] The working principle of this invention: This invention belongs to the field of industrial control technology. It collects real-time current and vibration data of a granulator through a data acquisition and compensation module, and uses adaptive unscented Kalman filtering combined with variational mode decomposition to reconstruct the state of time series missing due to communication delays or packet loss, generating compensated state data synchronized with the physical entity. The state prediction module uses the compensated state data as initial conditions, decomposes it into non-stationary and stationary components through empirical wavelet transform, constructs different cost functions for rolling time-domain estimation, and outputs the predicted state sequence within the future time window. The simulation optimization module uses the predicted state sequence as boundary conditions to drive a digital twin simulation process based on smooth particle dynamics, using the momentum and energy of a finite set of particles... The quantity exchange calculation calculates the load response characteristics under different feed rate adjustment schemes, and outputs the optimal feed rate setpoint with the goal of minimizing power consumption per ton and minimizing the risk of machine blockage. The residual correction module performs residual analysis on the measured data and the predicted state sequence, and uses rolling time domain estimation to correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter online and output the prediction deviation. The threshold control module calculates the control action trigger confidence level through fuzzy inference based on the current current change rate and the prediction deviation. When the confidence level exceeds the preset threshold, the optimal feed rate setpoint is sent to the feeding actuator. Otherwise, the current control parameters are maintained and the corrected parameters are used in subsequent cycles, thereby realizing real-time dynamic suppression and closed-loop optimization control of the pellet mill load fluctuation.

[0046] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A closed-loop control and status monitoring system for feed processing based on digital twins, characterized in that, include: The data acquisition and compensation module collects current and vibration data in real time during the operation of the pellet mill. It uses an adaptive unscented Kalman filter algorithm to reconstruct the state of the missing time series caused by communication delay or data packet loss, and generates compensated state data that matches the current moment of the physical entity. The state prediction module uses the compensation state data as initial conditions to perform rolling time-domain estimation, recursively predicts the load change trend within a preset time window, and outputs the predicted state sequence. The simulation optimization module uses the predicted state sequence as boundary conditions to drive the digital twin simulation process. It quickly iterates through lightweight numerical solutions to obtain the load response characteristics under different feed rate adjustment schemes, and outputs the optimal feed rate setting value with the goal of minimizing power consumption per ton and minimizing the risk of machine blockage. The residual correction module performs residual analysis on the actual feedback measured data and the predicted state sequence, and uses rolling time-domain estimation to correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter online, while outputting the prediction bias. The threshold control module monitors the current rate of change and prediction deviation. If the current exceeds the preset threshold, the optimal feed rate setting is sent to the feeding actuator through the real-time communication link. Otherwise, the current control parameters are maintained, and the corrected parameters are used for state reconstruction in subsequent cycles.

2. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 1, characterized in that, The generation of compensated state data that matches the current moment of the physical entity specifically includes: Variational mode decomposition is performed on current and vibration data from multiple consecutive sampling periods before the communication delay occurs, and multiple intrinsic mode components characterizing different frequency features are output. State-space equations are established for each intrinsic modal component, and adaptive unscented Kalman filtering is applied for recursive estimation. During the recursive process, the process noise covariance matrix is ​​adjusted in real time according to the instantaneous frequency of each modal component. The recursive estimation results of each intrinsic modal component are superimposed in the time domain to output compensated state data that matches the current moment of the physical entity.

3. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 1, characterized in that, The output predicted state sequence specifically includes: The compensation state data is decomposed into non-stationary components representing load trends and stationary components representing fluctuations by performing empirical wavelet transform. Different cost functions are constructed for non-stationary and stationary components respectively. The rolling optimization solution is performed using historical data within a fixed time window, and the predicted values ​​of each component within the future time window are output. The predicted values ​​of each component are reconstructed and superimposed, and a regularization term is introduced to constrain the rate of change of the amplitude of the control action, and the final predicted state sequence is output.

4. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 3, characterized in that, The output of the predicted values ​​of each component within the future time window specifically includes: For stationary components, a first cost function is constructed with the objective of minimizing the sum of squared prediction errors. The first parameter obtained by solving the first cost function is used as prior information, and a second cost function is constructed for non-stationary components with the objective of minimizing local fluctuation amplitude. The first cost function and the second cost function are solved iteratively and alternately. The output is the undetermined coefficients that make the two cost functions converge simultaneously. The undetermined coefficients are used to form the predicted values ​​of each component within the future time window.

5. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 1, characterized in that, The process for outputting the optimal feed rate setting value is as follows: The predicted state sequence is assigned as the initial boundary to the finite set of material flow points in the granulation channel. Each finite set of material flows carries the velocity and temperature attributes at the current moment. The relative positions of each particle are updated by adjusting the candidate values ​​based on the feed rate, and the momentum and energy exchange between particles are calculated using smooth particle dynamics, and the updated particle distribution state is output. The cumulative force of each mass point in the contact area of ​​the pellet mill roller is statistically analyzed and mapped to the load response characteristics at the current moment. Then, after traversing all candidate values ​​for feed rate adjustment, the optimal feed rate setting value is output.

6. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 5, characterized in that, The updated particle distribution state of the output specifically includes: Based on the current position of each particle and the preset smooth length, search all neighboring particles of each particle in the support domain, and calculate the kernel function value of each neighboring particle with respect to the current particle. The pressure gradient and viscous force between each neighboring particle and the current particle are calculated based on the kernel function value, and then the velocity and acceleration of the current particle are updated in the next moment. Calculate the frictional heat generated by the relative motion between the current particle and its neighboring particles based on the updated velocity and acceleration, and add the frictional heat to the internal energy property of the current particle. The updated velocity and internal energy properties of all particles are summarized, and particles that exceed the boundary are removed according to the granulation die hole constraints. The updated particle distribution is then output.

7. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 1, characterized in that, The method of using rolling time-domain estimation to online correct the noise covariance matrix and state transition parameters of the adaptive unscented Kalman filter, while simultaneously outputting the prediction bias, specifically includes: Subtract the measured data at the current moment from the corresponding predicted state sequence to output a residual sequence containing the estimation bias; Input the residual sequences from multiple recent time points into a sliding window, calculate the mean and variance of the residuals within the window, and determine whether there is a systematic shift at the current time point based on the mean. The calculated residual variance is compared with the theoretical innovation covariance, and the process noise covariance matrix of the adaptive unscented Kalman filter is adjusted according to the comparison results. The adjusted process noise covariance matrix and residual mean are used as corrections to the state transition parameters, while the residual value at the current time is used as the prediction bias output.

8. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 1, characterized in that, The monitoring of the current rate of change and prediction deviation specifically includes: The current value at the current moment is calculated by difference from the current value at the previous moment, and the current rate of change is output. At the same time, the prediction deviation is compared with the preset deviation tolerance range, and the comparison result is output. The current rate of change of current and the comparison result are input into the fuzzy inference unit, and the confidence level of the control action trigger at the current moment is calculated according to the predefined fuzzy rule base. The control action trigger confidence is compared with the preset action execution threshold. If the confidence exceeds the threshold, a trigger signal is generated, and the optimal feeding amount setting value is sent to the feeding actuator through the real-time communication link. If the confidence level does not exceed the threshold, the current control parameters are maintained, and the process noise covariance matrix and state transition parameters corrected at the current time are stored in the historical database for state reconstruction in subsequent cycles.

9. The feed processing closed-loop control and status monitoring system based on digital twins according to claim 8, characterized in that, The calculation of the control action trigger confidence at the current moment based on a predefined fuzzy rule base specifically includes: The current rate of change is mapped to the membership values ​​of multiple fuzzy sets according to a preset membership function, and the comparison results are mapped to the membership values ​​of multiple deviation states according to a preset membership function. The membership values ​​of the current change rate and the deviation state are used as inputs to the antecedent of the fuzzy rule. All rules in the fuzzy rule base are traversed, and the rule strength of each rule is output as the consequent. Take the maximum value among all rule strengths as the comprehensive reasoning result, and use the centroid method to defuzzify the comprehensive reasoning result and the membership function of the rule consequent, and output the control action trigger confidence at the current moment.

Images