Feed production ingredient automatic control system and control method thereof

Abstract

The present application relates to the technical field of feed production, and particularly relates to a feed production batching automatic control system and a control method thereof, comprising a predictive perturbation arch breaking module, which is used for identifying arching risk before discharging starts and applying a graded micro perturbation, adjusting the perturbation intensity according to material force chain relaxation feedback closed loop, and simultaneously adjusting the first-stage discharging speed according to the perturbation level; and a multi-dimensional compensation module, which is used for synchronously linking and correcting the batching parameters of the next batch when the perturbation influence exceeds the limited range. The present application calculates the arching risk index in real time through the predictive perturbation arch breaking module, actively intervenes in the micro perturbation which does not exceed one tenth of the normal arch breaking energy when the arching is still in the microscopic stress chain accumulation stage, adjusts the perturbation intensity according to the force chain relaxation feedback closed loop, and eliminates the arching in the budding state at a much lower energy cost, thereby significantly reducing the negative influence on the material compactness and the silo structure, and greatly reducing the arch breaking energy consumption.

Description

Technical Field

[0001] This invention relates to the field of feed production technology, and in particular to an automatic control system and control method for feed production ingredient batching. Background Technology

[0002] In the feed production process, the accuracy of ingredient proportioning is the core indicator that determines product quality and production cost. A typical automatic ingredient proportioning control system adopts a multi-stage feeding method: the first stage adds materials at a relatively high speed to about 90% of the target weight, and the second stage accurately replenishes materials to 100% of the target weight in a jogging manner. Under ideal working conditions, combined with iterative learning algorithms, high ingredient proportioning accuracy can be achieved.

[0003] Chinese Patent Publication No. CN107694469A discloses a direct-feed multi-component material batching method based on variable-rate learning. This method controls the feeding valve by iteratively predicting the amount of material in the feeding air based on single and cumulative feeding errors. During the iteration process, the learning factors for single and cumulative feeding errors are dynamically adjusted based on changes in the feeding error. Each feeding valve operates sequentially; after completing one batch of the formula feeding, the feeding valve opens, and the material is then uniformly mixed and discharged. This scheme uses distance sensors and vibration rods to detect and adjust the material accumulation in the feeding hopper, ensuring a stable feeding pattern. The learning factors can be automatically adjusted and optimized, eliminating the need for repeated trial and error adjustments of parameters and quickly obtaining an iterative formula with good convergence performance.

[0004] The existing technologies mentioned above also have the following problems: the arch breaking action occurs after the arch has been fully formed. At this time, the energy required to break the macroscopic arch is large. Forceful arch breaking will cause abnormal compaction of the material density and cumulative fatigue damage to the silo structure, thereby affecting the long-term stability and accuracy of production batching. At the same time, the arch breaking process will cause unsteady disturbances to the flow state of the material, which cannot be effectively suppressed by the existing drop compensation algorithm, resulting in large fluctuations in batch batching accuracy and poor product quality consistency. Summary of the Invention

[0005] To address this issue, the present invention provides an automatic control system and control method for feed production batching, which overcomes the problems in the prior art where strong arch breaking causes abnormal material compaction, directly affecting the stability of batching. At the same time, the non-steady-state disturbances generated by the material flow state after arch breaking cannot be effectively suppressed by conventional drop compensation algorithms, resulting in large fluctuations in batch batching accuracy and low product qualification rate.

[0006] To achieve the above objectives, the present invention provides an automatic control system for feed production and batching, including a predictive micro-perturbation arch-breaking module, which is used to identify the risk of arching before feeding begins and apply graded micro-perturbations, adjust the perturbation intensity according to the material force chain relaxation feedback closed loop, and adjust the first stage feeding speed according to the perturbation level. The disturbance propagation characteristic calculation module is used to determine the degree of impact of disturbance on the first stage feeding speed, shut-off advance amount, and second stage inching feeding after the batching is completed. The multi-dimensional compensation module is used to synchronously and in conjunction with the correction of the batching parameters for the next batch when the disturbance exceeds the limit. The graded data immunity module is used to grade and classify the disturbance batches and reuse the corresponding compensation strategies based on historical disturbance events. The feeding trend monitoring and deep fluidization processing module is used to monitor the stable state of feeding after arch breaking, and to perform overall fluidization reconstruction of the material inside the silo in case of abnormal over-limit conditions; The digital twin security verification module is used to perform offline simulation and security verification of all control parameters and execution strategies; The reinforcement learning training module is used to perform iterative training of control strategies in a virtual environment and then transfer the iterated strategies to the physical controller for execution.

[0007] Furthermore, the predictive micro-perturbation arch-breaking module includes a perturbation initiation subunit, a relaxation determination subunit, and a fluidization module; The disturbance initiation subunit activates the fluidization module to apply energy disturbance to the material based on the arching risk index being in the warning range, thereby physically disturbing the material. The relaxation determination subunit determines whether the internal force chain of the material has been successfully relaxed based on the acoustic emission signal. The relaxation determination subunit is also used to progressively increase the disturbance energy when relaxation fails, so as to apply disturbance to the material again until relaxation is determined to be effective or the disturbance energy reaches the safety limit.

[0008] Furthermore, the predictive perturbation arch-breaking module sets different levels of early warning intervals, and the fluidization module includes at least two different arch-breaking methods; The disturbance initiation subunit uses only one arch-breaking method when the arching risk index is in the low-risk range; and uses two arch-breaking methods in combination when the arching risk index is in the high-risk range. The disturbance energy increases gradually but does not exceed the material safety allowable range.

[0009] Furthermore, the predictive micro-perturbation arch-breaking module is equipped with a multi-dimensional sensor array to collect material characteristics, silo internal status, conveying equipment operating status, material level and quality correspondence, and environmental conditions. The arching risk index is calculated by comprehensively considering the above parameters. The disturbance propagation feature calculation module can extract three types of features from the batching process: weight deviation, number of jog feedings, and stability of feeding time. These features are then used to generate quantitative results that reflect the degree of influence of the disturbance on the first feeding speed, the turn-off advance, and the second jog feeding.

[0010] Furthermore, the multidimensional compensation module includes a disturbance compensation subunit; The disturbance compensation subunit, based on the disturbance impact exceeding the limited threshold, synchronously corrects the first stage feeding speed, shut-off advance time, and second stage inching feeding parameters of the next batch according to the historical disturbance response pattern of the silo.

[0011] Furthermore, the hierarchical data immunity module includes a hierarchical data processing subunit; The hierarchical data processing subunit classifies batch data into normal use, reduced weight, and direct removal based on the magnitude of the disturbance impact. The hierarchical data processing subunit establishes feature records for high-disturbance events, compares them with historical events in the cloud, and invokes historical compensation schemes based on the similarity level.

[0012] Furthermore, the material feeding trend monitoring and deep fluidization processing module includes a material feeding prediction subunit and a material feeding monitoring subunit; The feeding prediction subunit generates a theoretical feeding curve based on the first batch of ingredients after arch breaking; The feeding monitoring subunit generates an actual feeding curve in real time based on the first batch of ingredients after arch breaking, and compares the actual feeding curve with the theoretical feeding curve in real time. Based on the difference between the actual feeding curve and the theoretical feeding curve, it adaptively adjusts the feeding parameters.

[0013] Furthermore, the material feeding monitoring subunit reduces the material feeding speed based on the fact that the actual material feeding curve is higher than the theoretical material feeding curve; The material feeding monitoring subunit increases the material feeding speed based on the fact that the actual material feeding curve is lower than the theoretical material feeding curve. The material feeding monitoring subunit, based on the deviation between the actual material feeding curve and the theoretical material feeding curve exceeding a certain limit, initiates high-pressure gas injection to completely fluidize the material and allow it to settle and accumulate naturally again.

[0014] Furthermore, the digital twin security verification module is synchronized with the physical production line. All compensation parameters, disturbance strategies, and fluidized reconstruction instructions are first accelerated and simulated in the virtual model. Only after the multi-dimensional security score meets the standard can the module be issued for execution.

[0015] This invention proposes an automatic control method for feed production batching, applied to the aforementioned automatic control system for feed production batching, comprising the following steps: Before material feeding begins, a multi-dimensional sensor array collects various parameters and calculates the arching risk index. When the arching risk index is in the warning range, the disturbance initiation subunit activates the fluidization module to apply energy disturbance to the material. The relaxation judgment subunit determines whether the internal force chain of the material has been successfully relaxed based on the acoustic emission signal. If relaxation fails, the disturbance energy is gradually increased and the disturbance is applied again until relaxation is effective or the disturbance energy reaches the safety limit. At the same time, the first stage feeding speed is adjusted according to the disturbance level. After the batching is completed, the disturbance propagation feature calculation module extracts three types of features from the batching process: weight deviation, number of jog feedings, and stability of feeding time. These features are then used to generate quantitative results of the impact of disturbances on the first feeding speed, shut-off advance, and second jog feeding, thus forming a disturbance propagation feature set. When the disturbance exceeds the limit threshold, the disturbance compensation subunit will synchronously correct the first stage feeding speed, shut-off advance time and second stage inching feeding parameters of the next batch based on the historical disturbance response pattern of the silo. The hierarchical data processing subunit categorizes batch data into three types based on the magnitude of the disturbance: normal use, reduced weight, and direct removal. Feature records are established for high-disturbance events, and similarity comparisons are performed with historical events in the cloud. When the similarity reaches the standard, historical compensation schemes are invoked. In the first batch of materials after arch breaking, the feeding prediction subunit generates a theoretical feeding curve, and the feeding monitoring subunit generates an actual feeding curve in real time and compares it with the theoretical feeding curve in real time. When the actual feeding curve is higher than the theoretical feeding curve, the feeding speed is reduced; when the actual feeding curve is lower than the theoretical feeding curve, the feeding speed is increased. When the deviation between the two curves exceeds the limit, high-pressure gas injection is activated to completely fluidize the material and allow it to settle and accumulate naturally again. All compensation parameters, disturbance strategies, and fluidized reconstruction commands are first accelerated and simulated in the virtual model through the digital twin security verification module. Only after the multi-dimensional security score meets the standard can they be issued for execution. At the same time, the control strategy is iteratively trained through the reinforcement learning training module using the digital twin model as the training environment, and the iterated strategy is then transferred to the physical controller for execution.

[0016] Compared with the prior art, the beneficial effects of the present invention are that the present invention calculates the arching risk index in real time through multi-dimensional sensors, and actively intervenes with a small disturbance of no more than one-tenth of the normal arch-breaking energy when the arching is still in the stage of micro-stress chain accumulation. The disturbance intensity is adjusted according to the force chain relaxation feedback closed loop, and the arching is eliminated in the bud at a much lower energy cost, which significantly reduces the negative impact on material density and silo structure, while greatly reducing the energy consumption for arch breaking.

[0017] Furthermore, this invention extracts core data from three dimensions: weight deviation after batching, number of jog feeding times, and stability of feeding time. By using different weight combinations, it determines the degree of differential impact of disturbances on the first stage feeding speed, shut-off advance, and second stage jog parameters. When the disturbance exceeds the limit, it synchronously corrects multiple batching parameters of the next batch based on historical disturbance response patterns. This achieves an upgrade from "general compensation for single parameters" to "precise multi-point correction," significantly improving the recovery capability of batching accuracy after arch breaking.

[0018] Furthermore, this invention categorizes batch data into three levels of processing based on the magnitude of the disturbance: normal data entry, data entry with reduced weight, and data removal with triggering reverse correction. For the first time in the field of batching control, it achieves systematic identification and isolation of abnormal operating condition data, effectively solving the problem in existing technologies where iterative learning databases are contaminated by broken outlier data, leading to inaccurate batch matching for normal batches.

[0019] Furthermore, this invention establishes a disturbance event feature fingerprint database in the cloud, records the features of high disturbance events, and compares them with historical events for similarity. When the similarity meets the standard, the optimal historical compensation scheme is directly invoked without online trial and error. At the same time, preventive parameter adjustment is achieved through cluster analysis, enabling the system to have the ability to accumulate and reuse knowledge across batches and production lines.

[0020] Furthermore, this invention compares the actual feeding curve with the theoretical feeding curve in real time within the first batch of ingredients after arch breaking. When the deviation exceeds the limit, it initiates the overall fluidization reconstruction of the material—using high-pressure gas to completely fluidize the material and then allow it to settle and accumulate naturally again. This method serves as the last physical defense after all soft compensations fail, filling the gap in existing technologies where there is a lack of effective countermeasures when the material accumulation distortion exceeds the controllable range of the algorithm. At the same time, it ensures engineering feasibility through the constraint of the shortest cooling time.

[0021] Furthermore, this invention deploys a digital twin model synchronized with the physical production line in the cloud. All compensation strategies, parameter updates, and fluidized reconstruction instructions are first accelerated and simulated in the virtual model. Only after the multi-dimensional safety score meets the standard can the instructions be issued and executed, fundamentally eliminating the safety risks of online trial and error. At the same time, using this model as an environment, the optimal control strategy is automatically explored through reinforcement learning and migrated to the physical controller, realizing the continuous self-evolution of the control strategy and forming a closed-loop optimization system of "offline exploration - safety verification - online migration - online fine-tuning". Attached Figure Description

[0022] Figure 1 This is a structural block diagram of the feed dispensing control system according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the predictive perturbation arch breaking process according to an embodiment of the present invention. Figure 3This is a flowchart illustrating the perturbation feature calculation and multidimensional compensation in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the hierarchical data immunization and memory transfer process in an embodiment of the present invention. Figure 5 This is a flowchart illustrating the process of monitoring the material feeding trend and resetting the deep fluidization after arch breaking in an embodiment of the present invention. Figure 6 This is a block diagram of the closed-loop process for digital twin security verification and reinforcement learning in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0024] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] Please see Figures 1-6 As shown, this invention proposes an automatic control system for feed production batching, including a predictive micro-perturbation arch-breaking module, a perturbation propagation characteristic calculation module, a multi-dimensional compensation module, a graded data immunity module, a feed trend monitoring and deep fluidization processing module, a digital twin safety verification module, and a reinforcement learning training module. The predictive micro-perturbation arch-breaking module identifies arching risks and applies graded micro-perturbations before feed begins, adjusting the perturbation intensity based on the material force chain relaxation feedback loop, and simultaneously adjusting the first-stage feed speed according to the perturbation level. The perturbation propagation characteristic calculation module determines the impact of the perturbation on the first-stage feed speed, shut-off advance, and second-stage jogging feed after batching is completed. The multi-dimensional compensation module adjusts the feed speed for the next batch when the perturbation impact exceeds a defined range. The system synchronously corrects the batching parameters; the graded data immunity module is used to grade and classify disturbed batches and reuse corresponding compensation strategies based on historical disturbance events; the feeding trend monitoring and deep fluidization processing module is used to monitor the stable feeding state after arch breaking and to reconstruct the overall fluidization of materials inside the silo in case of abnormal over-limit conditions; the digital twin safety verification module is used to perform offline simulation and safety verification of all control parameters and execution strategies; the reinforcement learning training module is used to complete the iterative training of the optimal control strategy in a virtual environment and transfer the optimal strategy to the physical controller for execution; the system achieves full-process suppression of arching and long-term stability of batching accuracy through pre-event prediction and prevention, in-event precise compensation, post-event data immunity and physical reconstruction.

[0026] The predictive micro-perturbation arch-breaking module includes a perturbation initiation subunit, a relaxation determination subunit, and a fluidization module. The perturbation initiation subunit activates the fluidization module to apply energy perturbation to the material based on the arching risk index being in the warning range, thereby physically disturbing the material. The relaxation determination subunit determines whether the internal force chain of the material has been successfully relaxed based on acoustic emission signals. The relaxation determination subunit is also used to progressively increase the perturbation energy when relaxation fails, so as to apply perturbation to the material again until relaxation is determined to be effective or the perturbation energy reaches the safety limit. The predictive micro-perturbation arch-breaking module is equipped with a multi-dimensional sensor array, which can collect material characteristics, the internal state of the silo, the operating status of the conveying equipment, the correspondence between material level and mass, and environmental conditions, and comprehensively calculate the arching risk index.

[0027] The predictive micro-perturbation arch-breaking module is set with different levels of early warning intervals. The fluidization module includes, but is not limited to, arch-breaking methods such as acoustic waves, mechanical vibration, and pneumatic pulses. The disturbance initiation subunit uses only acoustic fluidization micro-perturbation when the arch risk index is in the low-risk interval. When the arch risk index is in the high-risk interval, it activates mechanical and pneumatic pulses for combined arch breaking. The disturbance energy increases step by step without exceeding the material safety tolerance range.

[0028] Understandably, the multi-dimensional sensor array includes an online moisture meter, a near-infrared particle size analyzer, an acoustic emission sensor, a screw conveyor current transformer, a temperature and humidity sensor, and a high-precision weighing sensor. The system collects data from these sensors once per second, first converting each raw data point into a dimensionless risk value between 0 and 1. The specific conversion method is as follows: Moisture and particle size risk values ​​are collected in real time by an online moisture meter and a near-infrared particle size analyzer. When the moisture content exceeds the critical moisture content of the corresponding material type, or the average particle size is smaller than the critical particle size, the material adhesion and internal friction angle increase significantly, and the probability of arching rises sharply. At this time, the risk value approaches 1. When the moisture content is lower than the safety threshold, the risk value is 0, and when it is higher than the high-risk threshold, the risk value is 1. Linear interpolation is used to complete the mapping within the interval.

[0029] The acoustic emission energy migration index in the radio frequency domain is calculated by performing a fast Fourier transform on the time-domain signal acquired by the acoustic emission sensor, determining the ratio of energy in the mid-to-high frequency band to that in the low-frequency band, and then normalizing it to the 0-1 range. During normal material flow, the sliding friction between particles is the main factor, and the acoustic emission energy is concentrated in the low-frequency band, so the index is close to 0. When the internal force chain of the material begins to accumulate, the acoustic emission energy migrates to the mid-to-high frequency band, and the corresponding index increases accordingly.

[0030] The auger current fluctuation rate is calculated by performing a sliding window analysis on the real-time signal collected by the current transformer, calculating the ratio of the standard deviation to the average value of the current value within a preset time window, and then dividing it by the empirical maximum value under the corresponding extreme blockage conditions to obtain a risk value between 0 and 1. The worse the material flowability, the greater the fluctuation range of the auger operating current, and the higher the corresponding risk value.

[0031] Material level-mass deviation is calculated by obtaining the volume of material in the silo through a level gauge and combining it with the standard bulk density to obtain the theoretical weight; at the same time, the actual weight is collected by a high-precision weighing sensor. The difference between the two is divided by the rated capacity of the silo to obtain a risk value between 0 and 1. If the difference shows a monotonically increasing trend over time, it indicates that the material in the silo has undergone local compaction, and the risk of arching has increased.

[0032] The temperature and humidity comprehensive index, based on a preset empirical mapping formula, converts ambient temperature and relative humidity into corresponding material adhesion risk levels, and then normalizes them to obtain the corresponding risk values.

[0033] The recent arch breakage frequency trend value is calculated by counting the number of times the arch breakage action is triggered within a preset historical batch window, and the result is normalized to the range of 0 to 1. When the arch breakage frequency increases, the risk value increases accordingly, reflecting the continuous change in the arching trend.

[0034] The spatial distribution heat map encoding of force chain fracture events is obtained by using a multi-channel acoustic emission sensor array distributed circumferentially and axially on the outer wall of the cone to obtain the spatial distribution of force chain fracture events using time difference positioning. When fracture events are concentrated on a certain side or at a certain height, the risk value increases.

[0035] Each of the seven risk values ​​is multiplied by a weighting coefficient, and then all the products are added together to obtain the structural risk index. The sum of all weighting coefficients is 1. The initial weighting coefficients are set based on experience, and the weights are dynamically optimized by the cloud self-learning module after the system is running.

[0036] In this embodiment, the predictive micro-perturbation arch breaking module sets two warning thresholds, namely a first warning threshold and a second warning threshold, and the first warning threshold is less than the second warning threshold. When the arching risk index is lower than the first warning threshold, it is determined to be a normal working condition, and no intervention is triggered, only continuous monitoring is performed. When the arching risk index is between the first warning threshold and the second warning threshold, it is determined that there is a slight arching risk, and the predictive micro-perturbation arch breaking is activated. At this time, the acoustic fluidization driver outputs low-energy pulses at the inherent frequency of the silo, which loosens the micro-force chains between material particles through resonance and prevents them from developing into macro-arching.

[0037] After the perturbation is applied, the system collects the vibration signal inside the material through an acoustic emission sensor. The decrease in the proportion of energy in the mid-to-high frequency band is used as the criterion for determining whether the stress chain is relaxed. This threshold is determined based on the statistical results of a large number of historical working conditions and can accurately distinguish between the successful and unsuccessful states of stress chain failure. When the decrease in the proportion of energy exceeds the set threshold, the stress chain is determined to be relaxed successfully and the perturbation action ends. If the decrease does not reach the threshold, it indicates that the current perturbation energy is insufficient to break the stress chain. The perturbation energy is then increased step by step according to the preset power step size and the perturbation is applied again until the stress chain is relaxed successfully or the perturbation energy reaches the preset safety limit.

[0038] By adopting this gradually increasing closed-loop control method, the arch-breaking effect can be achieved with the lowest effective energy. This avoids energy waste and excessive material vibration and compaction, and also prevents damage to the silo structure caused by excessive energy. It achieves low-energy and stable operation while ensuring the reliability of arch breaking.

[0039] When the arching risk index is greater than or equal to the second warning threshold, it is determined that there is a high risk. A single acoustic perturbation is not enough to effectively break the arch. At this time, the arch-breaking method is directly upgraded in stages. First, acoustic perturbation and low-amplitude mechanical vibration are started simultaneously, and then aerodynamic pulses are added to carry out joint intervention.

[0040] In this embodiment, regardless of the warning range, as long as a slight disturbance is output, the system immediately adjusts the first stage feeding parameters. Specifically, if the feeding efficiency is too low, it affects the production cycle; if the feeding efficiency is too high, the buffering effect is not obvious. Therefore, the rated operating frequency of the auger is reduced. The purpose of reducing the feeding speed is to provide a buffer period for material flow and avoid instantaneous material surge due to sudden relaxation of the stress chain. If a narrow allowable deviation band is maintained, the system will frequently trigger out-of-tolerance alarms and interrupt the batching process, resulting in a decrease in production efficiency. After appropriately widening the deviation band, most disturbance batches can successfully complete the first stage feeding without interruption. At the same time, the second stage jogging feeding still has sufficient adjustment margin and will not adversely affect the final batching accuracy. Therefore, widening the allowable deviation band of the first stage target weight can reserve more operating space for the second stage jogging feeding and avoid frequent interruptions of the batching process due to short-term out-of-tolerance.

[0041] Furthermore, the disturbance propagation feature calculation module can extract three core features from the batching process: weight deviation, number of jog feedings, and stability of feeding time, and generate quantitative results that reflect the degree of influence of disturbances on the first stage feeding speed, shut-off advance, and second stage jog feeding.

[0042] In this embodiment, after a batch of ingredients is completely finished, the system automatically acquires the following key data: the absolute value of the first segment actual deviation (in kilograms), the absolute value of the second segment actual deviation (in kilograms), the total target weight (in kilograms), the number of times the second segment is jog-feeded (in times), the reference number of jogs under normal working conditions -- statistics of the last 50 batches without arch-breaking disturbances, the standard deviation of the fluctuation of the material dropping time after arch breaking -- statistics of the last 10 batches with arch breaking time (in seconds), and the reference value of the fluctuation of the material dropping time under normal working conditions -- statistics of the last 50 normal batches.

[0043] The system first calculates three core ratios: the sum of the absolute values ​​of the first and second deviations, then divided by the target weight, to obtain the total deviation scale; the ratio of the actual number of jog feedings to the reference number of jog feedings under normal working conditions, to obtain the feeding difficulty; and the ratio of the standard deviation of the material drop time fluctuation after arch breaking to the reference value of the material drop time fluctuation under normal working conditions, to obtain the degree of deterioration of flow stability.

[0044] The system then uses three different weight configurations to calculate three disturbance influence coefficients: the first-stage material feeding speed influence coefficient, the shut-off advance influence coefficient, and the second-stage jogging parameter influence coefficient. The first-stage material feeding speed influence coefficient is obtained by weighted summation of three indicators: total deviation scale, material replenishment difficulty, and degree of deterioration of flow stability, according to their respective weights. The shut-off advance influence coefficient and the second-stage jogging parameter influence coefficient are also calculated based on the above three indicators, using weight combinations adapted to their respective influence characteristics.

[0045] The three coefficients mentioned above together constitute the disturbance propagation feature set. The first set of weights indicates that the total deviation scale has the greatest impact on the feeding speed; the second set of weights indicates that the difficulty of feeding has the most significant impact on the turn-off advance; and the third set of weights indicates that the difficulty of feeding has the greatest impact on the jogging parameters. These weight values ​​are obtained by artificially introducing arch-breaking disturbances of different intensities and performing multivariate regression calibration. Once calibrated, they are used as fixed parameters of the system.

[0046] Furthermore, the multi-dimensional compensation module includes a disturbance compensation subunit; the disturbance compensation subunit, based on the disturbance impact exceeding a limited threshold, synchronously corrects the first stage feeding speed, shut-off advance time, and second stage inching feeding parameters of the next batch according to the historical disturbance response pattern of the silo.

[0047] In this embodiment, the system sets a threshold for each control link. When the influence coefficient is lower than the corresponding threshold, the batching error of the next batch can still be kept within the standard allowable range even without compensation. When the influence coefficient exceeds the corresponding threshold, the batching error will increase with the coefficient and exceed the qualified standard. At this time, the system starts a multi-dimensional compensation mechanism. Each silo has a pre-stored disturbance response sensitivity matrix. This matrix is ​​used to characterize the compensation ratio required for the three control parameters when the three disturbance influence coefficients change. The elements in the matrix are obtained through multiple linear regression analysis of historical operating data and are updated and optimized by the system periodically.

[0048] The calculation process of the multidimensional compensation value is as follows: First, the three disturbance influence coefficients are arranged in order to form an eigenvector. Then, each row of the sensitivity matrix is ​​operated with the eigenvector to obtain the intermediate value of the feeding speed compensation, the intermediate value of the shut-off advance compensation, and the intermediate value of the jogging parameter compensation.

[0049] Finally, the three intermediate compensation values ​​are converted into absolute compensation amounts by combining them with the target weight of a single batch, resulting in the first stage of feeding speed compensation, the shut-off advance compensation, and the second stage of jogging parameter compensation. The batching parameters for the next batch are based on the benchmark parameters and the corresponding compensation amounts are added. Through this multi-dimensional linkage correction method, it is ensured that each control link can obtain targeted compensation that matches the disturbance propagation characteristics.

[0050] The hierarchical data immunity module includes a hierarchical data processing subunit; the hierarchical data processing subunit classifies batch data into normal use, reduced weight, and direct removal based on the magnitude of the disturbance impact; the hierarchical data processing subunit establishes feature records for high disturbance events, performs similarity comparison with historical events in the cloud, and calls the historical best compensation scheme based on the similarity level.

[0051] In this embodiment, the maximum value among the three influence coefficients is taken as the maximum disturbance coefficient. The system sets a first-level threshold and a second-level threshold. These two thresholds are determined based on the statistical distribution of a large number of historical disturbance intensities and are dynamically optimized by cloud learning.

[0052] The first level is low perturbation. When the maximum perturbation coefficient is less than the first level threshold, it is identified as a perturbation batch. All data in this batch are normally entered into the iterative learning database for cross-batch accuracy optimization. The second level is medium perturbation. When the maximum perturbation coefficient is greater than or equal to the first level threshold and less than the second level threshold, it is identified as a medium perturbation batch. Data in this batch is normally entered into the database, but in subsequent iterative learning updates, it is assigned a weight lower than that of normal batches. The weight decreases linearly as the maximum perturbation coefficient increases. The third level is high perturbation. When the maximum perturbation coefficient is greater than or equal to the second level threshold, it is identified as a high perturbation batch. Data in this batch is not entered into the iterative learning database, and the multi-dimensional compensation module is automatically triggered for reverse correction.

[0053] The system synchronously establishes a disturbance event feature fingerprint database in the cloud. The database records multi-dimensional feature information including material properties, environmental parameters, arch-breaking conditions, disturbance impact coefficient, compensation scheme, and final batching accuracy. Each time a high disturbance event occurs, the system extracts the feature fingerprint of the event and stores it in the database for archiving.

[0054] When a new high-disturbance event occurs, the system performs a similarity match between the feature fingerprint of the current event and the historical fingerprints in the database. The similarity calculation can use cosine similarity or normalized Euclidean distance. When the similarity reaches the set standard, indicating that the disturbance types are highly similar, the optimal compensation vector of the corresponding historical event is directly called as the initial value for this compensation. If the similarity is extremely high and the features of the two types of events are almost completely matched, the historical scheme is directly used without additional adjustments.

[0055] In addition, the system regularly performs cluster analysis on the fingerprint database. When it is found that the average disturbance level of a certain type of material in a high-disturbance batch in a specific environmental range is significantly higher than the overall average level and the qualified rate of the batching is significantly reduced, the system automatically lowers the preset feeding speed of the first stage of the material and expands the width of the protective band to provide more buffer space for material flow, reduce the impact of fluctuations caused by disturbances, and realize preventive parameter optimization based on immune memory.

[0056] Furthermore, the feeding trend monitoring and deep fluidization processing module includes a feeding prediction subunit and a feeding monitoring subunit. The feeding prediction subunit generates a theoretical feeding curve based on the first batch of materials after arch breaking. The feeding monitoring subunit generates an actual feeding curve in real time based on the first batch of materials after arch breaking, and compares the actual feeding curve with the theoretical feeding curve in real time. Based on the difference between the actual feeding curve and the theoretical feeding curve, the feeding parameters are adaptively adjusted. The feeding monitoring subunit reduces the feeding speed when the actual feeding curve is higher than the theoretical feeding curve, increases the feeding speed when the actual feeding curve is lower than the theoretical feeding curve, and initiates high-pressure gas injection when the deviation between the actual feeding curve and the theoretical feeding curve exceeds a limit value, so that the material is completely fluidized and then naturally settles and accumulates again.

[0057] Specifically, the material feeding prediction subunit calls a digital twin model, combines the current material characteristics, silo geometry, and applied feeding parameters to simulate and generate an ideal curve of cumulative weight change over time during the feeding process, i.e., the theoretical feeding curve, which serves as a reference benchmark for the feeding process. At the same time, real-time weight data is collected by a high-precision weighing sensor at a fixed sampling frequency to plot the actual feeding curve, and the actual curve is compared with the theoretical curve at each moment.

[0058] When the actual discharge curve is higher than the theoretical discharge curve, it indicates that the material falls faster than expected. The system will reduce the first-stage discharge speed to slow down the material's descent rate and avoid exceeding the tolerance. When the actual discharge curve is lower than the theoretical discharge curve, it indicates that the material's fluidity is weaker than expected and the material falls slower. The system will increase the first-stage discharge speed to ensure discharge efficiency. When the deviation between the actual discharge curve and the theoretical discharge curve exceeds the limit, it indicates that the material's accumulation state has been distorted and soft compensation can no longer effectively adjust it. The system will then activate the high-pressure gas injection process to allow the material to fully fluidize and then naturally settle and accumulate again.

[0059] Furthermore, the deviation is calculated using the maximum instantaneous deviation method: at each sampling moment, the absolute value of the difference between the actual weight and the theoretical weight is calculated, and the maximum value in the entire feeding process is taken as the final deviation. The limit value is dynamically set according to the material type and batching accuracy level, taking into account both batching accuracy and production efficiency. This avoids the threshold being too low, which would cause normal fluctuations to trigger frequent resets, and also prevents the threshold from being too high, which would prevent the identification of abnormal working conditions.

[0060] Before initiating deep fluidization reset, the system first verifies the minimum cooling time constraint. The minimum cooling time is determined by the digital twin model based on the material adhesion and silo structural stress simulation. If the interval since the last deep fluidization treatment is less than the minimum cooling time, an alternative adjustment method is adopted: the first stage feeding speed is further reduced, and an alarm is issued to prompt the operator to intervene manually. If the interval has met the minimum cooling time requirement, the deep fluidization treatment is officially executed.

[0061] The specific execution process of deep fluidization treatment is as follows: First, the feed valve is closed, pausing the current batching process; then, multiple high-pressure gas nozzles evenly arranged circumferentially in the cone section of the silo are used to simultaneously spray high-pressure gas at a pressure higher than that of conventional arch breaking, with the spraying duration being less than half the duration of conventional arch breaking; under the impact of the high-pressure airflow, the material in the silo is instantly and completely fluidized, and the original contact structure between particles is broken. After the spraying stops, the material naturally settles and accumulates again under the action of gravity, forming a new random accumulation state, completely eliminating the original distorted force chain structure.

[0062] After the reset, the system re-collects signals from the multi-dimensional sensor array, recalculates the arching risk index, and initializes the feeding parameters to the default safety values ​​corresponding to the material type, including lowering the auger operating frequency and expanding the allowable deviation band. In subsequent batches, the system is gradually iterated and optimized until it is finally restored to the normal operating parameters.

[0063] Furthermore, the digital twin safety verification module is synchronized with the physical production line. All compensation parameters, disturbance strategies, and fluidized reconstruction instructions are first accelerated and simulated in the virtual model. Only after the multi-dimensional safety score meets the standard can the instructions be issued for execution. The reinforcement learning training module uses the digital twin as the training environment and aims to achieve the optimal combination of long-term batching accuracy and arch-breaking energy consumption. It automatically explores combinations of different disturbance methods, compensation parameters, and reconstruction timings to generate the optimal control strategy and transfer it to the physical controller.

[0064] In this embodiment, a corresponding digital twin model is established for each physical production line in the cloud. The model is based on the discrete element method to simulate the movement, collision, and force chain formation and breakage of a large number of material particles in the silo. The model is synchronized with the physical production line in real time. The sensor data, control parameter settings, and actuator actions of the physical production line are all mapped to the virtual model in real time. The calculation results of the model can be used to predict the future state of material flow and arching.

[0065] All control strategies to be deployed to the physical controller, including compensation vectors generated by the multi-dimensional compensation module, weight parameters updated by cloud self-learning, and startup judgment results of the deep fluidization processing module, must undergo multiple batches of accelerated simulation and verification in the digital twin model before formal deployment. The model simulates the complete process of multiple consecutive batches of ingredients at a speed higher than real-time and calculates safety scores from multiple dimensions, covering three core dimensions: ingredient accuracy, equipment structural safety, and operating energy consumption. Only when the safety scores of all dimensions meet the passing standard will the system authorize the deployment of the control strategy to the physical controller for execution. If any dimension score fails to meet the standard, the strategy is rejected and corresponding optimization and adjustment suggestions are given, thus fundamentally avoiding the safety risks and production fluctuations caused by online trial and error.

[0066] The aforementioned digital twin model also serves as an offline training environment for reinforcement learning agents. The state space of the training environment covers multiple dimensions of operating parameters, including the current arching risk level, the state of each risk factor, and the perturbation propagation feature set of recent batches. The action space covers control actions such as perturbation energy level selection, arch breaking method combination, compensation parameter adjustment amount, and deep fluidization processing trigger judgment.

[0067] The reward function is designed with the core objective of achieving the optimal combination of improved batching accuracy and reduced arch-breaking energy consumption in the long term. This guides the agent to minimize arch-breaking energy consumption while ensuring batching accuracy, thus achieving a balance between accuracy and energy consumption. The agent is trained in a digital twin environment using a corresponding reinforcement learning algorithm, and can continuously run a large number of simulated batches to complete policy iterations. During training, the reward value gradually increases, and when the reward value no longer increases significantly after multiple consecutive runs, the policy training is considered to have converged.

[0068] The optimal control strategy parameters obtained through training convergence must first be verified by the digital twin security verification module. After verification, they are migrated to the edge controller of the physical production line in the form of parameter updates. After the strategy migration, the physical controller can also be fine-tuned online according to the on-site working conditions during actual operation to further adapt to the actual production environment. This forms a closed-loop optimization mechanism of "offline exploration - security verification - online migration - online fine-tuning" to achieve continuous self-evolution of the control strategy.

[0069] Please see Figures 1-5 As shown, this invention provides an automatic control method for feed production batching, applied to the aforementioned automatic control system for feed production batching, comprising: Step 1: Before material feeding begins, signals are collected using multi-dimensional sensors to calculate the arching risk index. When the index is in the warning range, graded micro-disturbances are applied to the silo, and the disturbance intensity is adjusted in a closed loop based on the force chain relaxation effect fed back by the acoustic emission sensor. At the same time, the first stage of material feeding speed is reduced according to the disturbance level, and the allowable deviation band is expanded. Step 2: After the batching is completed, extract the core data such as weight deviation, number of jog feedings, and stability of feeding time for each segment. Through different weight combinations, calculate three influence coefficients to reflect the degree of influence of the disturbance on the feeding speed of the first segment, the turn-off advance, and the jog feeding of the second segment, thus forming a disturbance propagation feature set. Step 3: When any coefficient in the disturbance propagation feature set exceeds its corresponding threshold, calculate the multidimensional compensation value based on the historical disturbance response pattern of the silo (i.e., the sensitivity matrix), and synchronously correct the first stage feeding speed, shut-off advance amount, and second stage jogging parameters of the next batch. Step 4: Classify the batch data according to the maximum value of the three impact coefficients: low disturbance is entered into the database normally, medium disturbance is entered into the database with reduced weight, and high disturbance is removed and reverse correction is triggered. The characteristics of high disturbance events are stored in the cloud database and matched with new events for similarity. When the match is successful, the historical best compensation scheme is directly reused. Step 5: In the first batch of materials after arch breaking, compare the actual feeding curve with the theoretical feeding curve in real time, calculate the deviation, and when the deviation exceeds the preset limit value, start the overall fluidization reconstruction of the material - after the material is completely fluidized by high-pressure gas blowing, it will settle and accumulate again. This operation is constrained by the shortest cooling time. Step six: All control parameters and execution strategies are first accelerated and verified by the cloud-based digital twin model. Only after the multi-dimensional safety scores meet the standards can they be issued for execution. At the same time, using the digital twin model as the environment, the optimal control strategy is generated through offline training via reinforcement learning, and the converged strategy is then transferred to the physical controller for execution.

[0070] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An automatic control system for feed production and batching, characterized in that, include, The predictive micro-perturbation arch breaking module is used to identify arching risks and apply graded micro-perturbations before material feeding begins. The perturbation intensity is adjusted according to the material force chain relaxation feedback closed loop, and the first stage feeding speed is adjusted according to the perturbation level. The disturbance propagation characteristic calculation module is used to determine the degree of impact of disturbance on the first stage feeding speed, shut-off advance amount, and second stage inching feeding after the batching is completed. The multi-dimensional compensation module is used to synchronously and in conjunction with the correction of the batching parameters for the next batch when the disturbance exceeds the limit. The graded data immunity module is used to grade and classify the disturbance batches and reuse the corresponding compensation strategies based on historical disturbance events. The feeding trend monitoring and deep fluidization processing module is used to monitor the stable state of feeding after arch breaking, and to perform overall fluidization reconstruction of the material inside the silo in case of abnormal over-limit conditions; The digital twin security verification module is used to perform offline simulation and security verification of all control parameters and execution strategies; The reinforcement learning training module is used to perform iterative training of the control strategy in a virtual environment and then transfer the iterated strategy to the physical controller for execution. The predictive perturbation arch-breaking module includes a perturbation initiation subunit, a relaxation determination subunit, and a fluidization module; The disturbance initiation subunit activates the fluidization module to apply energy disturbance to the material based on the arching risk index being in the warning range, thereby physically disturbing the material. The relaxation determination subunit determines whether the internal force chain of the material has been successfully relaxed based on the acoustic emission signal. The relaxation determination subunit is also used to gradually increase the disturbance energy when relaxation fails, so as to apply disturbance to the material again until relaxation is determined to be effective or the disturbance energy reaches the safety limit. The predictive micro-perturbation arch-breaking module is equipped with a multi-dimensional sensor array to collect material characteristics, silo internal status, conveying equipment operating status, material level and quality correspondence, and environmental conditions. The arching risk index is calculated by comprehensively considering the above parameters. The disturbance propagation feature calculation module can extract three types of features from the batching process: weight deviation, number of jog feedings, and stability of feeding time. These features are then used to generate quantitative results that reflect the degree of influence of the disturbance on the first stage feeding speed, the turn-off advance, and the second stage jog feeding. The material feeding trend monitoring and deep fluidization processing module includes a material feeding prediction subunit and a material feeding monitoring subunit; The feeding prediction subunit generates a theoretical feeding curve based on the first batch of ingredients after arch breaking; The feeding monitoring subunit generates an actual feeding curve in real time based on the first batch of ingredients after arch breaking, and compares the actual feeding curve with the theoretical feeding curve in real time. Based on the difference between the actual feeding curve and the theoretical feeding curve, it adaptively adjusts the feeding parameters.

2. The automatic feed production and batching control system according to claim 1, characterized in that, The predictive perturbation arch-breaking module sets different levels of early warning intervals, and the fluidization module includes at least two different arch-breaking methods; The disturbance initiation subunit uses only one arch-breaking method based on the fact that the arch risk index is in the low-risk range. Based on the fact that the arching risk index is in the high-risk range, two arch-breaking methods are used in combination to break the arch, with the disturbance energy increasing step by step without exceeding the material safety allowable range.

3. The automatic feed production and batching control system according to claim 2, characterized in that, The multidimensional compensation module includes a disturbance compensation subunit; The disturbance compensation subunit, based on the disturbance impact exceeding the limited threshold, synchronously corrects the first stage feeding speed, shut-off advance time, and second stage inching feeding parameters of the next batch according to the historical disturbance response pattern of the silo.

4. The automatic feed production and batching control system according to claim 3, characterized in that, The hierarchical data immunity module includes a hierarchical data processing subunit; The hierarchical data processing subunit categorizes batch data into normal use, reduced weight, and direct removal based on the magnitude of the disturbance impact. The hierarchical data processing subunit establishes feature records for high-disturbance events, compares them with historical events in the cloud, and invokes historical compensation schemes based on the similarity level.

5. The automatic feed production and batching control system according to claim 4, characterized in that, The material feeding monitoring subunit reduces the material feeding speed based on the fact that the actual material feeding curve is higher than the theoretical material feeding curve. The material feeding monitoring subunit increases the material feeding speed based on the fact that the actual material feeding curve is lower than the theoretical material feeding curve. The material feeding monitoring subunit, based on the deviation between the actual material feeding curve and the theoretical material feeding curve exceeding a certain limit, initiates high-pressure gas injection to completely fluidize the material and allow it to settle and accumulate naturally again.

6. The automatic feed production and batching control system according to claim 5, characterized in that, The digital twin security verification module is synchronized with the physical production line. All compensation parameters, disturbance strategies, and fluidized reconstruction instructions are first accelerated in the virtual model. Only after the multi-dimensional security score meets the standard can they be issued for execution.

7. An automatic control method for feed production batching, employing the automatic control system for feed production batching as described in any one of claims 1-6, characterized in that, Includes the following steps: Before material feeding begins, multiple parameters are collected using a multi-dimensional sensor array, and the arching risk index is calculated comprehensively. When the arching risk index is in the warning range, the disturbance initiation subunit starts the fluidization module to apply energy disturbance to the material. The relaxation judgment subunit judges whether the internal force chain of the material has been successfully relaxed based on the acoustic emission signal. If the relaxation fails, the disturbance energy is gradually increased and the disturbance is applied again until the relaxation is effective or the disturbance energy reaches the safety limit. At the same time, the first stage feeding speed is adjusted according to the disturbance level. After the batching is completed, the disturbance propagation feature calculation module extracts three types of features from the batching process: weight deviation, number of jog feedings, and stability of feeding time. These features are then used to generate quantitative results of the impact of disturbances on the first feeding speed, shut-off advance, and second jog feeding, thus forming a disturbance propagation feature set. When the disturbance exceeds the limit threshold, the disturbance compensation subunit will synchronously correct the first stage feeding speed, shut-off advance time and second stage inching feeding parameters of the next batch based on the historical disturbance response pattern of the silo. The hierarchical data processing subunit categorizes batch data into three types based on the magnitude of the disturbance: normal use, reduced weight, and direct removal. Feature records are established for high-disturbance events, and similarity comparisons are performed with historical events in the cloud. When the similarity reaches the standard, historical compensation schemes are invoked. In the first batch of materials after arch breaking, the feeding prediction subunit generates a theoretical feeding curve, and the feeding monitoring subunit generates an actual feeding curve in real time and compares it with the theoretical feeding curve in real time. When the actual feeding curve is higher than the theoretical feeding curve, the feeding speed is reduced; when the actual feeding curve is lower than the theoretical feeding curve, the feeding speed is increased. When the deviation between the two curves exceeds the limit, high-pressure gas injection is activated to completely fluidize the material and allow it to settle and accumulate naturally again. All compensation parameters, disturbance strategies, and fluidized reconstruction commands are first accelerated and simulated in the virtual model through the digital twin security verification module. Only after the multi-dimensional security score meets the standard can they be issued for execution. At the same time, the control strategy is iteratively trained through the reinforcement learning training module using the digital twin model as the training environment, and the iterated strategy is then transferred to the physical controller for execution.

Images