Intelligent control system and method for feed processing equipment

Abstract

This invention discloses an intelligent control system and method for feed processing equipment, belonging to the field of intelligent control technology. The system includes a data collector, a characteristic extractor, an energy efficiency health monitor, a predictive controller, and a frequency-modulated sampler. The data collector collects raw material humidity, feed flow rate, crusher current, mixer torque, pellet mill temperature, and target particle size of the finished product in real time. The characteristic extractor weights humidity and flow rate according to the fiber content of the raw material and uses a residence time distribution model convolution to obtain a dynamic processability index. The energy efficiency health monitor performs time-frequency decomposition of the crusher current. The predictive controller uses pseudo-partial derivatives for online estimation. The frequency-modulated sampler samples according to the information entropy change sampling frequency, predicts torque mutations through spectrum analysis, and increases the sampling frequency, shortens the prediction time domain, and increases the control energy weight before the mutation, thereby improving control accuracy, equipment reliability, and product consistency.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an intelligent control system and method for feed processing equipment. Background Technology

[0002] Feed processing typically involves multiple stages, including crushing, mixing, and pelleting, with strong coupling and nonlinear characteristics between these stages. In actual production, process parameters such as raw material moisture and feed flow rate fluctuate with batch variations, and the residence time of materials within the mixer exhibits a certain distribution. This results in a significant delayed effect of changes in raw material characteristics on downstream pelleting quality.

[0003] In existing control systems, some employ deep learning models to extract time-series features of temperature, humidity, flow rate, and rotational speed of the ring die pellet mill, and then output rotational speed increase / decrease commands through a classifier. While this method can achieve a certain degree of adaptive adjustment, it fails to model the residence time distribution of raw materials during the mixing process, resulting in a lag in response to changes in raw material characteristics and affecting the stability of pellet quality. Another approach uses a mixing uniformity comparison table and a case library to adjust control parameters. This method relies on the matching accuracy of historical cases and lacks online sensing capabilities for the equipment's own degradation state. When the crusher hammers wear or the mixer blades age, the control parameters are difficult to automatically compensate for, leading to a gradual accumulation of control errors.

[0004] Furthermore, the aforementioned existing technologies only target the single-objective adjustment of the ring die speed in the granulation process. They do not incorporate the energy consumption, product quality, equipment wear, and actuator stability of the entire crushing, mixing, and granulation process into a unified multi-objective optimization framework. They also cannot predict and proactively adjust transient anomalies such as sudden torque changes in the mixer before they occur. Usually, they can only take passive protection measures after the anomaly occurs, resulting in fluctuations in product quality or equipment damage. Summary of the Invention

[0005] In response to the above situation, the present invention can model the residence time distribution of raw materials in the mixing process to compensate for the delay effect, use the high-frequency component of the crusher current to sense the wear status of the equipment online and adaptively adjust the control, and realize multi-objective collaborative control and transient active response in the feed processing process through five-objective weighted rolling optimization and torque mutation spectrum prediction mechanism.

[0006] The technical solution is an intelligent control system for feed processing equipment, comprising: The data collector collects data in real time on the moisture content of raw materials, feed flow rate, current of the crusher, torque of the mixer, temperature of the pellet mill, and the target value of the finished product particle size during the feed processing process. The feature extractor, after weighting the raw material moisture and the feed flow rate according to the raw material fiber content, uses a residence time distribution model that reflects the material mixing delay to convolve the historical combination index to obtain the dynamic processability index. The energy efficiency health device performs time-frequency decomposition on the current of the crusher, extracts the equipment wear factor from the cumulative variance of the high-frequency component, extracts the energy efficiency correction term from the amplitude of the mid- and low-frequency components, and obtains the corrected predicted unit energy consumption by combining the dynamic processability index. The predictive controller employs a pseudo-partial derivative online estimation refresh control model to solve for the optimal feed rate setpoint and optimal granulation temperature setpoint that minimize the rolling time domain objective function within the adaptive predictive time domain. The rolling time domain objective function consists of a weighted sum of squares of the following five terms: the first term is the square of the difference between the corrected predicted unit energy consumption and the target energy consumption multiplied by a first weight; the second term is the square of the difference between the estimated finished particle size and the target finished particle size multiplied by a second weight; the third term is the square of the difference between the mixer torque and the optimal torque multiplied by a third weight; the fourth term is the equipment wear factor multiplied by a fourth weight; and the fifth term is the sum of the squares of the feed rate increment and the squares of the granulation temperature increment multiplied by a fifth weight. The estimated finished particle size is calculated based on the granulator temperature and the raw material moisture content. The fourth weight increases linearly with the increase of the equipment wear factor; and the fifth weight automatically increases with the increase of the sum of the squares of the absolute values ​​of the feed rate increment and the absolute values ​​of the granulation temperature increment. The frequency-modulated sampler adjusts the feeder and steam valve according to the optimal feeding speed setting and the optimal granulation temperature setting; simultaneously, it changes the sampling frequency in real time according to the information entropy change rate of each sensor signal, and predicts sudden torque changes in the mixer by analyzing the spectrum of the crusher current. The prediction criterion is: when the energy in the 2nd to 5th harmonic frequency band of the crusher current exceeds 2.5 times the average energy in the 2nd to 5th harmonic frequency band of the mixer blades during steady-state operation, it is judged that a sudden change is about to occur; when a sudden change is predicted to occur, the sampling frequency is temporarily increased to twice the maximum value, the prediction time domain is forcibly shortened to 1 second, and the fifth weight is temporarily increased to twice the steady-state value.

[0007] Furthermore, the characteristic extractor linearly determines the weighting coefficients for humidity deviation and flow rate deviation according to the percentage of raw material fiber content. The sum of the two coefficients is 1. The humidity deviation rate and flow rate ratio are multiplied by their corresponding coefficients and then added together to obtain the static processability index. The static processability index is then convolved and integrally applied using a multi-reactor series residence time distribution model. The series number of the multi-reactor series residence time distribution model is identified and updated online during production by periodically injecting tracers at the feed inlet and detecting the response curve at the discharge outlet. The identified series number serves as the lower limit constraint for the prediction time domain length of the predictive controller. The initial value of the lower limit constraint is 2 seconds. During the online identification process, the characteristic extractor synchronously records the peak time and variance of the tracer response curve to determine whether dead zones or channeling occur in the material flow state within the mixer. When dead zone or channeling signs are detected, a mixer maintenance warning is automatically issued.

[0008] Furthermore, when the signal from any sensor exceeds the preset effective range corresponding to that sensor, the data collector automatically switches to the average value of historical data from the most recent 30 sampling periods as the current value and generates an abnormal alarm including the sensor number. During the first three dwell time periods after system startup, the feature extractor operates in an open-loop manner to initialize the historical queue of the convolution integral. The length of the historical queue is three times the time constant of the dwell time distribution model. After initialization, the feature extractor uses the first segment of open-loop data as a baseline for comparison with the subsequent online identification of cascaded series. When the absolute deviation between the identified cascaded series and the baseline cascaded series exceeds 1, the predictive controller is triggered to increase the lower limit of the prediction time domain by 0.5 seconds each time until the deviation is eliminated. Simultaneously, the data collector marks the data missing intervals during sensor failure and transmits them to the energy efficiency health device. The energy efficiency health device automatically removes the data within the marked intervals when calculating the high-frequency variance cumulative integral.

[0009] Furthermore, in the energy-efficient health device, the time-frequency decomposition of the pulverizer current employs wavelet packet decomposition. The number of decomposition layers is adaptively determined according to the pulverizer speed, ensuring that the lower limit frequency of the highest frequency band is always higher than twice the characteristic frequency of the hammerhead. The variance of the high-frequency component is first compared with the reference variance. This ratio is then mapped to the 0-1 interval using a hyperbolic tangent function and multiplied by a fusion coefficient that automatically switches according to the fiber content. The fusion coefficient is 0.8 when the fiber content is higher than 15% and 0.4 when it is lower than 15%. The equipment wear factor is defined as the ratio of the cumulative integral value of the high-frequency variance to the cumulative wear capacity. The upper limit of the ratio of the high-frequency variance cumulative integral value to the cumulative wear capacity is 1, where the cumulative wear capacity is the high-frequency variance cumulative integral value from the time the new equipment is put into use to the first major overhaul; when the ratio of the high-frequency variance cumulative integral value to the cumulative wear capacity reaches 0.8, it is marked as a high wear state and the predictive controller is forced to switch to the life priority mode. At the same time, the high wear mark is transmitted to the frequency modulation sampler, so that the 2.5 times threshold in the prediction standard is reduced to 1.5 times; when the equipment wear factor has not increased within 10 consecutive control cycles, the energy efficiency health device automatically reduces the cumulative integral value of the equipment wear factor by 5%.

[0010] Furthermore, the pseudo-partial derivatives in the predictive controller are estimated online using a recursive least squares method with a forgetting factor. Its input vectors are the feed rate increment and granulation temperature increment, and its output vectors are the actual unit energy consumption increment and particle size deviation increment. The learning rate factor in the estimation process is initially 0.5, and increases by 0.1 when the cumulative absolute value of the control error increases by more than 20% within 10 cycles. The robustness factor is initially 0.01, and increases by 0.005 when the peak-to-peak value of either the feed rate setpoint or the granulation temperature setpoint exceeds twice the average amplitude during steady-state operation within 5 consecutive cycles. The forgetting factor is initially 0.98, and decreases by 0.02 for every 0.1 increase in the equipment wear factor, with a minimum limit of 0.90. The length of the prediction time domain is adaptively changed according to the rate of change of the dynamic processability index: 2 seconds when the rate of change of the dynamic processability index is greater than 0.1 / second, 4 seconds when the rate of change of the dynamic processability index is between 0.05 and 0.1 / second, and 8 seconds when the rate of change of the dynamic processability index is less than 0.05 / second, and finally the larger value between the lower limit of the prediction time domain and the adaptively calculated value is taken; the prediction controller also includes a gradient suppression module for the sum of the square of the feed rate increment and the square of the granulation temperature increment. When the value of the sum of the square of the feed rate increment and the square of the granulation temperature increment in one control cycle exceeds 3 times the value of the previous cycle, the optimal feed rate setting value and the optimal granulation temperature setting value of the current cycle are automatically reduced by 50% towards the previous cycle.

[0011] Furthermore, in the frequency modulation sampler, the information entropy change rate is calculated using Shannon entropy with a sliding window width of 100 sampling points and an amplitude histogram divided into 20 bins. When the entropy change rate is positive and exceeds 0.2, the sampling frequency increases by 20%; when the entropy change rate is negative and below -0.2, the sampling frequency decreases by 20%, and the range of sampling frequency variation is 50% to 200% of the reference frequency. In the spectrum analysis predicting sudden torque changes in the mixer, the initial value of the multiple by which the frequency band energy exceeds the average value of the steady-state operating baseline is 2.5 times. When the wear factor of the equipment increases by 0.1, the frequency band energy exceeds the average value of the steady-state operating baseline by a factor of 0.3, with a minimum limit of 1.2 times. When a sudden change is predicted to occur, the sampling frequency is temporarily increased to twice the maximum value, the prediction time domain is forcibly shortened to 1 second, and the fifth weight is temporarily increased to twice the steady-state operating value. After the temporary adjustment of the sampling frequency, prediction time domain, and fifth weight is completed, the prediction time domain length is restored by increasing by 1 second per control cycle until it reaches the normal value under the current operating conditions.

[0012] A method for intelligent control of feed processing equipment includes the following steps: Step 1: The data collector collects a multi-source parameter set at a sampling frequency that changes in real time. The multi-source parameter set includes the raw material humidity, the feed flow rate, the crusher current, the mixer torque, the granulator temperature, and the target particle size of the finished product. Step 2: The characteristic extractor performs a weighted combination of the raw material moisture and the feed flow rate according to the raw material fiber content, and then performs convolution on the combined indicators of historical time moments through the residence time distribution model to obtain the dynamic processability index, and identifies the series stage and monitors the flow status of the mixer online; Step 3: The energy efficiency health device performs wavelet packet decomposition on the current of the crusher, obtains the wear factor of the equipment according to the cumulative variance of the high-frequency components and implements static relaxation attenuation, and obtains the corrected predicted unit energy consumption according to the amplitude of the mid- and low-frequency components and the dynamic processability index. Step 4: The predictive controller refreshes the pseudo-partial derivatives online and solves for the optimal feed rate setting and the optimal granulation temperature setting that minimize the rolling time domain objective function within the predictive time domain. The rolling time domain objective function consists of a weighted sum of five terms, and the weight of the equipment wear factor term increases adaptively as the equipment wear factor increases. Step 5: The frequency modulation sampler outputs the optimal feeding speed setting value and the optimal granulation temperature setting value to adjust the feeder and steam valve. At the same time, it changes the sampling frequency in real time according to the information entropy change rate, and predicts the sudden change in mixer torque through spectrum analysis. The prediction standard is: when the energy in the pulverizer current of the mixer blade passing frequency band 2 to 5 times the average energy of the mixer blade passing frequency band during steady-state operation exceeds 2.5 times the average energy of the mixer blade passing frequency band during steady-state operation, it is judged that a sudden change is about to occur. Before the sudden change occurs, the sampling frequency is increased to twice the maximum value, the prediction time domain is locked to 1 second, and the fifth weight is increased to twice the steady-state operation value. After the prediction is completed, the prediction time domain is restored step by step. Step Six: Feed back the deviation between the actual unit energy consumption and the corrected predicted unit energy consumption to Step Two for online correction of the residence time distribution model parameters and dynamic weights.

[0013] Furthermore, the feature is that, in step four, when the equipment wear factor exceeds 0.8, the weight of the equipment wear factor item is forcibly increased to above 0.5, and the energy-saving priority mode and quality priority mode are blocked, and the life priority mode is automatically switched; when the sum of the square of the feed speed increment and the square of the granulation temperature increment exceeds three times the average value of the sum of the square of the feed speed increment and the square of the granulation temperature increment during steady-state operation for three consecutive control cycles, the upper limit of the change step size of the feed speed setting value is automatically reduced from the rated maximum step size to 50% of the rated maximum step size.

[0014] Furthermore, the method includes a model monitoring step: calculating the ratio of the pseudo-partial derivative to the historical average of the most recent 100 control cycles; triggering a model reset when the ratio is less than 0.5 or greater than 2.0; requesting global model parameters from a remote parameter server as the initial value of the pseudo-partial derivative during reset, and simultaneously uploading the de-identified model parameters from the local server to the remote parameter server for updating the global model; the reset signal is also fed back to step two, temporarily freezing the online update of the dwell time distribution model parameters for 20 control cycles until the model reconverges.

[0015] Furthermore, when the equipment wear factor exceeds 0.95, the system automatically switches to maintenance mode: suspends the optimization algorithm, runs at a fixed low feed rate setting of 20% of the rated value and a rated pelletizing temperature setting of 90% of the rated value, and generates a maintenance warning signal; at the same time, in the maintenance mode, the weight of the equipment wear factor item in the rolling time domain objective function is reset to zero, the fifth weight is set to 5 times that of steady-state operation, and the upper limit of the feed rate setting is forcibly reduced to 30% of the rated value until the equipment wear factor falls back to below 0.9.

[0016] Due to the adoption of the above technical solutions, the present invention has the following advantages compared with the prior art; 1. On the one hand, a multi-reactor series residence time distribution model is introduced into the characteristic extractor. The static processability index at the current moment is convolved with the index at historical moments to obtain the dynamic processability index. This allows the predictive controller to sense the changing trend of raw material characteristics that are about to arrive at the pellet mill in advance, thereby adjusting the feeding speed and pelleting temperature ahead of time. On the other hand, this invention performs wavelet packet time-frequency decomposition on the crusher current in the energy efficiency health device. The cumulative variance of the high-frequency components is converted into the equipment wear factor after calibration, enabling the control system to determine the remaining life of the hammer online. The amplitude of the mid- and low-frequency components is combined with the dynamic processability index to correct the predicted unit energy consumption, solving the defect that conventional energy consumption models do not include equipment aging factors. Moreover, the frequency modulation sampler will correspondingly reduce the threshold for torque mutation prediction (from 2.5 times to 1.5 times), so that aging equipment can be more sensitively protected, and the control accuracy and equipment reliability are greatly improved.

[0017] 2. A rolling time-domain objective function with a weighted sum of squares was constructed in the predictive controller. Simultaneously, a torque mutation prediction technology based on current spectrum analysis was implemented in the frequency modulated sampler. By monitoring the energy in the frequency band of 2 to 5 times the frequency of the mixer blades in real time, an early warning can be issued more than 0.5 seconds before the actual torque mutation. Three operations are performed: the sampling frequency is increased to twice the maximum value to capture transient details, the prediction time domain is forcibly shortened to 1 second to speed up decision-making, and the fifth weight is temporarily increased to twice the steady-state value to limit the control amplitude. This upgrades the system from passive response to active defense. Through multi-objective dynamic balance and transient active prediction, the entire process of feed processing is achieved with multi-objective and proactive intelligent control, which significantly improves control accuracy, equipment life and product consistency. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention. Figure 2 This is a diagram illustrating the method steps of the present invention. Detailed Implementation The foregoing and other technical contents, features and effects of the present invention are described in conjunction with the appendix below. Figures 1 to 2 The detailed description of the embodiments will make this clear. All structural details mentioned in the following embodiments are based on the accompanying drawings.

[0019] Example 1: Based on existing technology, this example provides an intelligent control system for feed processing equipment. The system includes a data collector, a characteristic extractor, an energy efficiency health monitor, a predictive controller, and a frequency modulation sampler.

[0020] The data acquisition unit collects real-time data on raw material moisture, feed flow rate, grinder current, mixer torque, pellet mill temperature, and target particle size of the finished product during feed processing. When the signal from any sensor exceeds its preset effective range, the unit automatically switches to the average of the last 30 sampling periods as the current value. The unit also generates an alarm with the sensor number as its identifier. Furthermore, the unit marks the data loss intervals during sensor failures and transmits this information to the energy efficiency health monitor.

[0021] The feature extractor is connected to the data acquisition unit. The feature extractor performs a weighted combination of raw material moisture content and feed flow rate based on the raw material fiber content. The feature extractor uses a residence time distribution model reflecting material mixing delay to convolve historical combination indices, obtaining a dynamic processability index. Specifically, the feature extractor linearly determines the weighting coefficients for moisture deviation and flow rate deviation based on the percentage of raw material fiber content. The sum of these two coefficients is 1. The feature extractor multiplies the moisture deviation rate (the difference between standard moisture and real-time moisture divided by the standard moisture) and the flow rate ratio (the ratio of real-time flow rate to the maximum allowable feed flow rate) by their respective coefficients and then sums them to obtain a static processability index. The feature extractor then uses a multi-reactor series residence time distribution model to perform a convolution integral on this static processability index. The number of series stages in the multi-reactor series residence time distribution model is identified and updated online during production. The identification method involves periodically injecting a tracer (e.g., colored powder or salt solution) at the feed inlet and detecting the response curve at the discharge outlet. The identified number of series stages serves as a lower limit constraint on the prediction time domain length of the predictive controller. The initial value of the lower limit constraint is 2 seconds. During online identification, the feature extractor synchronously records the peak time and variance of the tracer response curve. Based on the peak time and variance, the feature extractor determines whether dead zones or channeling have occurred in the material flow within the mixer. When signs of dead zones or channeling are detected, the feature extractor automatically issues a mixer maintenance warning. During the first three residence time cycles after system startup, the feature extractor operates in an open-loop mode with a fixed feed rate and granulation temperature to initialize the historical queue of the convolution integral. The length of this historical queue is three times the time constant of the residence time distribution model. After initialization, the feature extractor uses the first segment of open-loop data as a baseline. The feature extractor uses this baseline to compare with the subsequent cascade levels identified online. When the absolute deviation between the identified cascade level and the baseline cascade level exceeds 1, the predictive controller is triggered to increase the prediction time domain lower limit by 0.5 seconds each time until the deviation is eliminated.

[0022] The energy efficiency health device is connected to a characteristic extractor. It performs time-frequency decomposition on the pulverizer current. The energy efficiency health device extracts the equipment wear factor from the cumulative variance of the high-frequency components. It extracts the energy efficiency correction term from the amplitude of the mid- and low-frequency components. The energy efficiency health device combines this with the dynamic processability index to obtain a corrected predicted unit energy consumption. The time-frequency decomposition of the pulverizer current in the energy efficiency health device uses wavelet packet decomposition. The number of decomposition layers is adaptively determined according to the pulverizer speed, ensuring that the lower limit frequency of the highest frequency band is always higher than twice the characteristic frequency of the hammerhead. The variance of the high-frequency components is first compared with the reference variance. The reference variance is the average variance collected after 10 hours of normal operation of the new equipment. This ratio of the high-frequency component variance to the reference variance is mapped to the 0-1 interval using a hyperbolic tangent function, and then multiplied by a fusion coefficient that automatically switches according to the fiber content. The fusion coefficient is 0.8 when the fiber content is higher than 15% and 0.4 when it is lower than 15%. The equipment wear factor is defined as the ratio of the cumulative integral of high-frequency variance to the cumulative wear capacity, with an upper limit of 1. The cumulative wear capacity is the cumulative integral of high-frequency variance from the time a new piece of equipment is put into use until its first major overhaul. When the ratio of this cumulative integral to the cumulative wear capacity reaches 0.8, the energy efficiency health monitor marks it as a high-wear state. The energy efficiency health monitor forces the predictive controller to switch to life-priority mode. Simultaneously, the energy efficiency health monitor transmits the high-wear marker to the frequency modulation sampler, reducing the 2.5 times threshold in the prediction standard to 1.5 times. When the equipment wear factor does not increase within 10 consecutive control cycles, the energy efficiency health monitor automatically reduces the cumulative integral value of the equipment wear factor by 5%. The data acquisition unit marks the data gaps during sensor failure periods and transmits them to the energy efficiency health monitor. The energy efficiency health monitor automatically discards data within the marked gaps when calculating the cumulative integral of high-frequency variance.

[0023] The predictive controller is connected to both the characteristic extractor and the energy efficiency health monitor. The predictive controller uses pseudo-partial derivatives for online estimation to refresh the control model. Within the adaptive predictive time domain, the predictive controller solves for the optimal feed rate setpoint and optimal pelleting temperature setpoint that minimize the rolling time domain objective function. The rolling time domain objective function consists of the following five weighted sums of squares: the first term is the square of the difference between the corrected predicted unit energy consumption and the target energy consumption multiplied by the first weight. The second term is the square of the difference between the estimated finished particle size and the target finished particle size multiplied by the second weight. The third term is the square of the difference between the mixer torque and the optimal torque multiplied by the third weight. The fourth term is the equipment wear factor multiplied by the fourth weight. The fifth term is the sum of the squares of the feed rate increment and the squares of the pelleting temperature increment multiplied by the fifth weight. The estimated finished particle size is calculated based on the pelleting temperature and raw material humidity. The specific calibration method is as follows: keeping other process parameters constant, the pelleting temperature and raw material humidity are changed separately, the rate of change of the finished particle size is measured, and empirical coefficients are determined through linear regression to establish a particle size estimation model. The fourth weight increases linearly with the increase of the equipment wear factor. The fifth weight automatically increases with the sum of the squares of the absolute values ​​of the feed rate increment and the granulation temperature increment. The pseudo-partial derivatives in the predictive controller are estimated online using a recursive least squares method with a forgetting factor. Its input vector consists of the feed rate increment and the granulation temperature increment. Its output vector consists of the actual unit energy consumption increment and the particle size deviation increment. The initial value of the learning rate factor during the estimation process is 0.5. When the cumulative absolute value of the control error increases by more than 20% within 10 cycles, the learning rate factor increases by 0.1. The initial value of the robustness factor is 0.01. When the peak-to-peak value of either the feed rate setpoint or the granulation temperature setpoint exceeds twice the average amplitude during steady-state operation within 5 consecutive cycles, the robustness factor increases by 0.005. The initial value of the forgetting factor is 0.98. For every 0.1 increase in the equipment wear factor, the forgetting factor decreases by 0.02, with a minimum limit of 0.90. The length of the prediction time domain is adaptively adjusted according to the rate of change of the dynamic processability index: 2 seconds when the rate of change is greater than 0.1 / second, 4 seconds when the rate of change is between 0.05 and 0.1 / second, and 8 seconds when the rate of change is less than 0.05 / second. The larger of the lower limit of the prediction time domain and the adaptively calculated value is ultimately used. The predictive controller also includes a gradient suppression module for the sum of the squares of the feed rate increment and the squares of the pelleting temperature increment. When the value of this sum of squares in a control cycle exceeds three times the value of the previous cycle, the gradient suppression module automatically backscales the optimal feed rate setpoint and optimal pelleting temperature setpoint of the current cycle by 50% towards the previous cycle.

[0024] The frequency modulation (FM) sampler is connected to the predictive controller. The FM sampler adjusts the feeder and steam valves according to the optimal feed rate and optimal granulation temperature setpoints. The FM sampler changes the sampling frequency in real time based on the information entropy change rate of each sensor signal. The information entropy change rate is calculated using Shannon entropy with a sliding window width of 100 sampling points and an amplitude histogram divided into 20 bins. When the entropy change rate is positive and exceeds 0.2, the sampling frequency increases by 20%. When the entropy change rate is negative and below -0.2, the sampling frequency decreases by 20%. The sampling frequency range is 50% to 200% of the reference frequency. The FM sampler predicts sudden torque changes in the mixer by analyzing the spectrum of the pulverizer current. The prediction criterion is: when the energy in the pulverizer current at frequencies 2 to 5 times the frequency passing through the mixer blades exceeds 2.5 times the average energy of that frequency band during steady-state operation, a sudden change is considered imminent. When a sudden change is predicted, the FM sampler temporarily increases the sampling frequency to twice the maximum value. The frequency-modulated sampler forcibly shortens the prediction time domain to 1 second. The frequency-modulated sampler temporarily increases the fifth weight to twice the steady-state value. After the temporary adjustments to the sampling frequency, prediction time domain, and fifth weight are completed, the prediction time domain length recovers by increasing by 1 second per control cycle until it reaches the normal value under the current operating conditions. In this embodiment, one control cycle is 0.5 seconds; those skilled in the art can adjust it according to the system response speed. In the spectrum analysis predicting sudden torque changes in the mixer, the initial value for the multiple by which the frequency band energy exceeds the average value of the steady-state operating baseline is 2.5 times. This multiple decreases by 0.3 for every 0.1 increase in the equipment wear factor, with a minimum limit of 1.2 times.

[0025] Example 2, based on Example 1, provides an intelligent control method for feed processing equipment using the above system. This method includes the following six steps.

[0026] Step 1: The data logger collects a multi-source parameter set at a real-time varying sampling frequency. This multi-source parameter set includes raw material moisture content, feed flow rate, crusher current, mixer torque, granulator temperature, and the target particle size of the finished product. The data logger simultaneously performs historical mean replacement and data missing interval marking for sensor failures.

[0027] Step Two: The characteristic extractor performs a weighted combination of raw material moisture and feed flow rate based on the raw material fiber content. The characteristic extractor then convolves the combined indices from historical time points using a residence time distribution model to obtain the dynamic processability index. Simultaneously, the characteristic extractor identifies the number of cascade stages and monitors the flow status of the mixer online. The specific operation method is the same as the detailed working process of the characteristic extractor in Example 1.

[0028] Step 3: The energy efficiency monitor performs wavelet packet decomposition on the crusher current. The monitor calculates the equipment wear factor based on the cumulative variance of the high-frequency components and implements static relaxation attenuation. Static relaxation attenuation refers to automatically reducing the cumulative integral value by 5% when the equipment wear factor has not increased for 10 consecutive control cycles. The monitor then calculates the corrected predicted unit energy consumption based on the amplitude of the mid-to-low frequency components and the dynamic processability index.

[0029] Step 4: The predictive controller refreshes the pseudo-partial derivatives online. The predictive controller solves for the optimal feed rate setpoint and optimal granulation temperature setpoint that minimize the rolling time-domain objective function within the predictive time domain. The rolling time-domain objective function consists of a weighted sum of squares of five terms. The weight of the equipment wear factor term adaptively increases as the equipment wear factor increases. When the equipment wear factor exceeds 0.8, the weight of this factor term is forcibly increased to above 0.5. The predictive controller disables the energy-saving priority mode and the quality priority mode, automatically switching to the lifespan priority mode. When the sum of the square of the feed rate increment and the square of the granulation temperature increment exceeds three times the average value of this sum during steady-state operation for three consecutive control cycles, the predictive controller automatically reduces the upper limit of the feed rate setpoint change step size from the rated maximum step size to 50% of the rated maximum step size.

[0030] Step 5: The frequency modulation (FM) sampler outputs the optimal feed rate setpoint and optimal granulation temperature setpoint to adjust the feeder and steam valve. The FM sampler changes the sampling frequency in real time according to the information entropy change rate. The FM sampler predicts sudden torque changes in the mixer through spectrum analysis. The prediction criterion is: when the energy in the pulverizer current at frequencies 2 to 5 times the frequency through which the mixer blades pass exceeds 2.5 times the average energy of that frequency band during steady-state operation, a sudden change is judged to be imminent. Before the sudden change occurs, the FM sampler increases the sampling frequency to twice the maximum value. The FM sampler locks the prediction time domain to 1 second. The FM sampler increases the fifth weight to twice the steady-state value. After the prediction is completed, the FM sampler gradually restores the prediction time domain (increasing by 1 second per control cycle until the normal value is reached).

[0031] Step Six: Feedback the deviation between the actual unit energy consumption and the corrected predicted unit energy consumption to Step Two. This feedback is used to correct the residence time distribution model parameters and dynamic weights online, forming a fully closed-loop self-learning mechanism. The specific method of feedback correction is as follows: when the actual energy consumption is consistently higher than the predicted value, the cascade number of the residence time distribution model is appropriately increased to simulate a worse mixing effect, thereby adopting a more conservative strategy in subsequent control.

[0032] The method also includes a model monitoring step. The ratio of the pseudo-partial derivatives to the historical average of the most recent 100 control cycles is calculated. When this ratio is less than 0.5 or greater than 2.0, a model reset is triggered. During reset, global model parameters are requested from a remote parameter server as initial values ​​for the pseudo-partial derivatives. Simultaneously, the model parameters before the local reset are anonymized and uploaded to the remote parameter server for updating the global model. Anonymization refers to removing non-parametric information such as production line identifiers, timestamps, and equipment numbers, retaining only the pseudo-partial derivatives and model parameter values. The reset signal is also fed back to step two, temporarily freezing the online update of the dwell time distribution model parameters for 20 control cycles until the model reconverges.

[0033] When the equipment wear factor exceeds 0.95, the system automatically switches to maintenance mode. In maintenance mode, the optimization algorithm is paused. The system operates at a fixed low feed rate setting of 20% of the rated value and a pelletizing temperature setting of 90% of the rated value. The system generates a maintenance warning signal. In maintenance mode, the weight of the equipment wear factor term in the rolling time-domain objective function is reset to zero. The fifth weight is set to 5 times the steady-state operating value. The upper limit of the feed rate setting is forcibly reduced to 30% of the rated value. The above maintenance mode continues to run until the equipment wear factor falls below 0.9.

[0034] In practical application, this invention, based on existing technology, firstly convolves historical data of parameters such as raw material moisture and feed flow rate using a residence time distribution model in the characteristic extractor, outputting a dynamic processability index in real time. This allows the predictive controller to anticipate changes in raw material characteristics approaching the granulation stage, thus avoiding the adjustment lag phenomenon that occurs in traditional control due to neglecting the mixing residence time distribution. Simultaneously, while existing technologies typically use pulverizer current only for load monitoring or energy efficiency estimation, this invention utilizes an energy efficiency health indicator to perform wavelet packet time-frequency decomposition. The cumulative variance of high-frequency components is used to quantify the equipment wear factor, and the mid-to-low-frequency components are used to correct unit energy consumption predictions, achieving dual utilization of the same signal and online perception of health status. When the wear factor exceeds a threshold, the system automatically increases the weight of the wear term in the objective function and switches to a lifespan-priority mode. Simultaneously, the frequency modulation sampler automatically lowers the prediction threshold for torque mutation based on the wear factor, providing more timely protection for aging equipment. To address the shortcomings of existing technologies that only adjust a single objective (such as the rotational speed of a ring mill), this invention constructs a five-objective weighted sum-of-squares rolling optimization framework using a predictive controller. This framework includes energy consumption, particle size, torque, wear factor, and a control action smoothness penalty term. The weight of the wear factor term adaptively increases with its value, and the weight of the control energy term automatically increases with the magnitude of the action, thereby achieving a dynamic balance between energy efficiency, quality, equipment lifespan, and actuator smoothness. Furthermore, while existing technologies typically only take passive measures such as shutdown or alarms after a torque surge occurs, this invention utilizes a frequency-modulated sampler to perform continuous spectrum analysis on the crusher current, monitoring the energy in the frequency band 2 to 5 times the frequency of the mixer blades. When the energy exceeds 2.5 times the steady-state average, a surge can be predicted more than 0.5 seconds in advance. The sampling frequency is increased, the prediction time domain is forcibly shortened to 1 second, and the control energy weight is temporarily increased, enabling the system to proactively adjust to a conservative strategy before the anomaly actually occurs. This minimizes equipment damage and product quality fluctuations, achieving full-process, multi-objective, proactive intelligent control of the feed processing process, significantly improving control accuracy, equipment lifespan, and product consistency.

[0035] The above description is a further detailed explanation of the invention in conjunction with specific embodiments, and it should not be considered that the specific embodiments of the invention are limited to this. For those skilled in the art to which this invention pertains and related fields, any extensions, operation methods, and data substitutions made based on the technical solution concept of this invention should fall within the protection scope of this invention.

Claims

1. An intelligent control system for feed processing equipment, characterized in that, include: The data collector collects real-time data on raw material moisture, feed flow rate, crusher current, mixer torque, granulator temperature, and target particle size of the finished product. The feature extractor weights the raw material moisture and the feed flow rate according to the raw material fiber content, and then uses the residence time distribution model to convolve the historical combination index to obtain the dynamic processability index. The energy efficiency health device performs time-frequency decomposition on the current of the crusher, extracts the equipment wear factor from the cumulative variance of the high-frequency component, extracts the energy efficiency correction term from the amplitude of the mid- and low-frequency components, and obtains the corrected predicted unit energy consumption by combining the dynamic processability index. The predictive controller employs a pseudo-partial derivative online estimation refresh control model to solve for the optimal feed rate setpoint and optimal granulation temperature setpoint that minimize the rolling time domain objective function within the adaptive predictive time domain. The rolling time domain objective function includes: the squared difference between the corrected predicted unit energy consumption and the target energy consumption, the squared difference between the estimated finished particle size and the target value, the squared difference between the mixer torque and the optimal torque, the equipment wear factor, and the sum of the squared increments of the feed rate and granulation temperature, each multiplied by its corresponding weight. The weight of the equipment wear factor increases linearly with its magnitude, and the weight of the sum of the squared increments of the feed rate and granulation temperature automatically increases with their magnitudes. The frequency-modulated sampler adjusts the feeder and steam valve according to the optimal feeding speed setpoint and the optimal granulation temperature setpoint; it changes the sampling frequency in real time according to the information entropy change rate of each sensor signal; it predicts sudden torque changes in the mixer by analyzing the spectrum of the crusher current. The prediction criterion is: when the energy in the 2 to 5 times the frequency of the mixer blades in the crusher current exceeds 2.5 times the average energy of that frequency band during steady-state operation, it is judged that a sudden change is about to occur; when a sudden change is predicted, the sampling frequency is temporarily increased to twice the maximum value, the prediction time domain is forcibly shortened to 1 second, and the weight of the sum of the square of the feeding speed increment and the square of the granulation temperature increment is temporarily increased to twice that during steady-state operation.

2. The intelligent control system for feed processing equipment according to claim 1, characterized in that, The characteristic extractor linearly determines the weighting coefficients for humidity deviation and flow rate deviation according to the percentage of raw material fiber content. The sum of the two coefficients is 1. The humidity deviation rate and flow rate ratio are multiplied by their corresponding coefficients and then added together to obtain the static processability index. The static processability index is then convolved and integrally applied using a multi-reactor series residence time distribution model. The series number of the multi-reactor series residence time distribution model is identified and updated online by periodically injecting tracers into the feed inlet and detecting the response curve at the discharge outlet. The identified series number serves as the lower limit constraint for the prediction time domain length of the predictive controller. During the online identification process, the characteristic extractor synchronously records the peak time and variance of the tracer response curve to determine whether dead zones or channeling occur in the material flow state within the mixer. When dead zone or channeling signs are detected, a mixer maintenance warning is automatically issued.

3. The intelligent control system for feed processing equipment according to claim 1, characterized in that, When the signal from any sensor exceeds the corresponding preset effective range, the data collector automatically switches to the average value of the historical data from the most recent 30 sampling periods as the current value and generates an abnormal alarm. During the first three dwell time periods after system startup, the feature extractor operates in an open-loop manner to initialize the historical queue of the convolution integral. The length of the historical queue is three times the time constant of the dwell time distribution model. After initialization, the feature extractor uses the first segment of open-loop data as a baseline for comparison with the subsequent cascaded levels of online identification. When the absolute deviation exceeds 1, the predictive controller is triggered to increase the lower limit of the prediction time domain by 0.5 seconds each time until the deviation is eliminated. The data collector marks the missing data intervals during sensor failure and transmits them to the energy efficiency health device. The energy efficiency health device automatically removes the data within the marked intervals when calculating the high-frequency variance cumulative integral.

4. The intelligent control system for feed processing equipment according to claim 1, characterized in that, In the energy efficiency health device, the time-frequency decomposition of the crusher current uses wavelet packet decomposition, and the number of decomposition layers is adaptively determined according to the crusher speed, so that the lower limit frequency of the highest frequency band is higher than twice the characteristic frequency of the hammer. The variance of the high-frequency component is first compared with the reference variance, and this ratio is mapped to the 0 to 1 interval by the hyperbolic tangent function, and then multiplied by the fusion coefficient that automatically switches with the fiber content. The equipment wear factor is defined as the ratio of the cumulative integral value of the high-frequency variance to the cumulative wear capacity, with an upper limit of 1. The cumulative wear capacity is the cumulative integral value of the high-frequency variance from the time the new equipment is put into use to the first major overhaul. When the ratio reaches 0.8, it is marked as a high wear state and the predictive controller is forced to switch to the life priority mode. At the same time, the high wear mark is transmitted to the frequency modulation sampler, so that the 2.5 times threshold in the prediction standard is reduced to 1.5 times. When the equipment wear factor does not increase within 10 consecutive control cycles, the energy efficiency health device automatically decays the cumulative integral value by 5%.

5. The intelligent control system for feed processing equipment according to claim 1, characterized in that, The pseudo-partial derivatives in the predictive controller are estimated online using a recursive least squares method with a forgetting factor. The input vector consists of the feed rate increment and the granulation temperature increment, and the output vector consists of the actual unit energy consumption increment and the particle size deviation increment. The length of the prediction time domain is adaptively changed according to the rate of change of the dynamic processability index, and the larger value between the lower limit of the prediction time domain and the adaptively calculated value is taken. The predictive controller also includes a gradient suppression module. When the sum of the square of the feed rate increment and the square of the granulation temperature increment exceeds three times the value of the previous cycle within a control cycle, the optimal feed rate setpoint and the optimal granulation temperature setpoint of the current cycle are pushed back by 50% towards the previous cycle.

6. The intelligent control system for feed processing equipment according to claim 1, characterized in that, In the frequency modulation sampler, when the entropy change rate is positive and exceeds 0.2, the sampling frequency increases by 20%; when it is negative and below -0.2, it decreases by 20%, with a range of 50% to 200% of the reference frequency. In the spectrum analysis to predict sudden torque changes in the mixer, the initial value of the frequency band energy exceeding the average value of the steady-state operating baseline is 2.5 times. When the equipment wear factor increases by 0.1, it decreases by 0.3, with a minimum limit of 1.2 times. When a sudden change is predicted to occur, the sampling frequency is temporarily increased to twice the maximum value, the prediction time domain is forcibly shortened to 1 second, and the weight of the sum of the square of the feed rate increment and the square of the granulation temperature increment is temporarily increased to twice that of steady-state operation. After the temporary adjustment, the prediction time domain length is restored by increasing by 1 second per control cycle until it reaches the normal value.

7. A method for intelligent control of feed processing equipment using the system described in any one of claims 1 to 6, characterized in that, Includes the following steps: Step 1: The data collector collects a set of multi-source parameters at a sampling frequency that changes in real time, including the raw material humidity, the feed flow rate, the crusher current, the mixer torque, the granulator temperature, and the target particle size of the finished product. Step 2: The characteristic extractor performs a weighted combination of the raw material moisture and the feed flow rate according to the raw material fiber content, and performs convolution on the combined indicators of historical time moments through the residence time distribution model to obtain the dynamic processability index, and identifies the series stage and monitors the flow status of the mixer online; Step 3: The energy efficiency health device performs wavelet packet decomposition on the current of the crusher, obtains the wear factor of the equipment according to the cumulative variance of the high-frequency components and implements static relaxation attenuation, and obtains the corrected predicted unit energy consumption according to the amplitude of the mid- and low-frequency components and the dynamic processability index. Step 4: The predictive controller refreshes the pseudo-partial derivatives online and solves for the optimal feed rate setting and the optimal granulation temperature setting that minimize the objective function in the rolling time domain within the predictive time domain, wherein the weight of the equipment wear factor term increases adaptively as the equipment wear factor increases. Step 5: The frequency modulation sampler outputs the optimal feeding speed setting value and the optimal granulation temperature setting value to adjust the feeder and steam valve. The sampling frequency is changed in real time according to the information entropy change rate, and the sudden change of mixer torque is predicted by spectrum analysis. The prediction standard is: when the energy of the frequency band 2 to 5 times the frequency of the mixer blades in the pulverizer current exceeds 2.5 times the average energy of this frequency band during steady-state operation, it is judged that a sudden change is about to occur. Before the sudden change occurs, the sampling frequency is increased to twice the maximum value, the prediction time domain is locked to 1 second, and the weight of the sum of the square of the feed speed increment and the square of the granulation temperature increment is increased to twice that during steady-state operation. After the prediction is completed, the prediction time domain is restored step by step. Step Six: Feed back the deviation between the actual unit energy consumption and the corrected predicted unit energy consumption to Step Two for online correction of the residence time distribution model parameters and dynamic weights.

8. The intelligent control method for feed processing equipment according to claim 7, characterized in that, In step four, when the wear factor of the equipment exceeds 0.8, the weight of the wear factor item is forcibly increased to 0.5 or above and the life priority mode is switched. When the sum of the square of the feed rate increment and the square of the granulation temperature increment exceeds three times the average value of steady-state operation in three consecutive control cycles, the upper limit of the change step size of the feed rate setting value is reduced to 50% of the rated maximum step size.

9. The intelligent control method for feed processing equipment according to claim 7, characterized in that, The model monitoring step also includes: calculating the ratio of the pseudo-partial derivative to the historical average of the most recent 100 control cycles; triggering a model reset when the ratio is less than 0.5 or greater than 2.0; requesting global model parameters from a remote parameter server as the initial value of the pseudo-partial derivative during the reset, and simultaneously uploading the desensitized model parameters from the local server to the remote parameter server; the reset signal is also fed back to step two, temporarily freezing the online update of the dwell time distribution model parameters for 20 control cycles until the model reconverges.

10. The intelligent control method for feed processing equipment according to claim 7, characterized in that, When the equipment wear factor exceeds 0.95, the system automatically switches to maintenance mode: suspends the optimization algorithm, runs at a fixed low feed rate setting of 20% of the rated value and a rated pelletizing temperature setting of 90% of the rated value, and generates a maintenance warning signal; at the same time, in maintenance mode, the weight of the equipment wear factor item in the rolling time domain objective function is reset to zero, the fifth weight is set to 5 times that of steady-state operation, and the upper limit of the feed rate setting is forcibly reduced to 30% of the rated value until the equipment wear factor falls back to below 0.9.

Images