Production line ingredient automatic control system and method for compounding spices

Abstract

The invention discloses a production line ingredient automatic control system and method for spice compounding, and belongs to the technical field of production line control. The method comprises the following steps: acquiring sensor data, defining a state vector and fusing the sensor data; establishing an anti-blocking control model, performing model adaptive correction, and outputting an anti-blocking control strategy; agglomerated particles are identified, particle size distribution is output, and particle size grades are divided; executing a disassembly strategy based on the particle size grade, and verifying the disassembly effect to obtain a caking verification result; when it is confirmed that no caking exists in the feeder, the lag time caused by the viscous ingredients is updated; adding fuzzy adaptive compensation to obtain compensation time; establishing a lag compensation effect evaluation index, carrying out no-load and full-load lag test on each spice based on the lag time, and verifying the compensated coupling error; and through relative gain matrix analysis, establishing a cross compensation matrix for the strong coupling pair, and generating a compensated rotating speed instruction of the feeder.

Description

Technical Field

[0001] The present invention relates to the technical field of production line control, and specifically to an automatic control system and method for ingredient batching in a production line for spice compounding. Background Art

[0002] Spice compounding is a key link in determining the flavor consistency of products, and its automatic control of ingredient batching directly affects production efficiency and product quality. Most spices are powdery or granular materials, which have strong hygroscopicity and poor fluidity. During storage, transportation, and ingredient batching, they are extremely prone to caking due to factors such as electrostatic adsorption and deliquescence, resulting in uneven feeding of the feeder and large deviation in ingredient batching accuracy. At the same time, the viscosity differences of different spices are significant, and the lag effect of the ingredient batching system will further exacerbate the coupling error of multi-component batching, causing the compounding ratio to deviate from the preset formula and affecting the flavor stability of the product.

[0003] Traditional anti-caking methods rely on mechanical stirring or vibration loosening at a fixed frequency, without combining real-time material state dynamic adjustment strategies. For viscous spices such as chili powder and ginger powder, existing lag compensation models are mostly based on fixed parameters and do not track the changes in material characteristics in real time. Traditional control systems adopt a single-loop independent control mode and do not consider the dynamic coupling between multiple feeders. Summary of the Invention

[0004] The purpose of the present invention is to provide an automatic control system and method for ingredient batching in a production line for spice compounding to solve the problems raised in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solutions: In the first aspect, the present application provides an automatic control method for ingredient batching in a production line for spice compounding, including the following steps: Obtain sensor data, including spice state and the thickness of the wall adhesion layer, define a state vector, and fuse the sensor data using the extended Kalman filter; establish an anti-caking control model based on the fused sensor data and perform model adaptive correction; the anti-caking control model outputs a caking prevention control strategy and performs hierarchical anti-caking control; When there is caking in the feeder, identify the caking particles through the YOLOv8 model, output the particle size distribution, and divide the particle size grades; execute the disassembly strategy based on the particle size grades and verify the disassembly effect to obtain a caking verification result; Based on the caking verification result, when it is confirmed that there is no caking in the feeder, dynamically update the lag time caused by viscous ingredient batching using the double exponential smoothing method; increase fuzzy adaptive compensation based on the lag time to obtain the compensation time; Establish a lag compensation effect evaluation index, conduct no-load and full-load lag tests on each spice based on the lag time, and verify the coupled error after compensation; based on the coupled error and compensation time, through relative gain matrix analysis, establish a cross-compensation matrix for strong coupling pairs to generate the feed machine speed command after compensation.

[0006] Combined with the first aspect, in the first implementation manner of the first aspect of this application, the obtaining of sensor data includes the spice state and the thickness of the wall adhesion layer. Define a state vector and use the extended Kalman filter to fuse the sensor data, including: The spice state includes the humidity, particle size, viscosity, and density of the spice. Define the sensor data as a state vector; set a random noise term for each state vector to reflect the unmodeled dynamics, and establish a state equation based on the physical laws of the state vector change; establish a mapping relationship between the sensor data and the true state, and set a covariance matrix for each sensor data to reflect the measurement uncertainty. Based on the state estimate value at the previous moment, predict the state at the current moment based on the state equation, and calculate the covariance matrix of the predicted state at the current moment; compare the difference between the measured value and the predicted value of the sensor, calculate the Kalman gain, and correct the predicted state based on the Kalman gain to obtain the optimal estimate value; update the covariance matrix to complete a filtering cycle; iterate the extended Kalman filter to fuse the sensor data.

[0007] Combined with the first aspect, in the second implementation manner of the first aspect of this application, the establishment of an anti-caking control model based on the fused sensor data and the model adaptive correction include: The anti-caking control model includes a data layer, a model layer, and a strategy layer; in the data layer, receive the fused sensor data and perform normalization processing; in the model layer, use a fuzzy logic controller to construct a nonlinear mapping relationship, define linguistic variables, and establish an IF-THEN rule base (production rule base); in the strategy layer, convert the model output into specific device control instructions; the data source for training the model is historical production data, specifically the sensor data and manual processing records corresponding to the caking events of different spices, and the training objective is to make the strategy level output by the model negatively correlated with the actual caking incidence rate. Define the correction trigger condition, and according to the real-time caking data, increase the rule priority of the high-frequency occurrence scenario and dynamically adjust the membership function.

[0008] Combined with the first aspect, in the third implementation manner of the first aspect of this application, when there is caking in the feed machine, identify the caking particles through the YOLOv8 model, output the particle size distribution, and divide the particle size grades, including: Collect typical caked samples, take images in the laboratory under different lighting conditions, spice states, and background interferences, and use the LabelImg tool to annotate the caked particles as the training set; rotate, scale, and adjust the brightness of the training set to expand it; for the YOLOv8 model, load the COCO pre-trained weights, set the training parameters, recalculate the prior box sizes using K-means clustering, and train the YOLOv8 model; input the preprocessed images into the trained YOLOv8 model to output the bounding box coordinates and class confidence of the caked particles; establish the mapping relationship between pixels and actual sizes through a calibration board, and for each detection box, take the geometric mean of its width and height as the equivalent diameter; perform perspective correction on the detection boxes in the edge area based on the camera installation angle; for the adhered particles within the detection box, use the watershed algorithm for secondary segmentation to improve the particle size calculation accuracy; Statistically analyze the particle sizes of all caked particles in the current frame, generate a particle size histogram at intervals, and output the particle size distribution, including the average particle size, D10, D50, and D90; according to the predefined particle size grades, statistically analyze the proportion of the number and mass of particles in each grade to divide the particle size grades.

[0009] Combined with the first aspect, in the fourth implementation manner of the first aspect of this application, the implementation of the disassembly strategy based on the particle size grade is carried out to verify the disassembly effect, and the caking verification result is obtained, including: The particle size grades include primary caking, secondary caking, tertiary caking, quaternary caking, and quinary caking; for primary caking, start the low-frequency vibrating screen at the outlet of the feeder, loosen and separate the particle agglomerates through vibration, and turn on the ultrasonic vibration device on the inner wall of the pipeline to emit pulses to break the weak adhesion between particles; for secondary caking, start the double-roll crusher at the inlet of the feeder, shear the caked particles through differential rotation, and synchronously turn on the pneumatic vibrator at the bottom of the bin to prevent subsequent batching accumulation from affecting the crushing effect; for tertiary caking, based on the treatment measures for secondary caking, turn on the high-pressure air knife to blow the crushed particles away from the blades to avoid secondary adhesion; for quaternary caking, the double-roll crusher switches to the reverse rotation mode, breaks the hard caking through rubbing, and starts the centrifugal disperser at the end of the pipeline to use centrifugal force to throw the large particles towards the wear-resistant lining plate for impact crushing; for quinary caking, stop the feeder, close the valve at the outlet of the bin, trigger an audible and visual alarm, start the 360° rotating cleaning arm inside the bin, and try to physically remove the super-large caking. When it is not removed within the set time, trigger the manual maintenance process; Obtain the current particle size distribution. When it meets the qualified standard, it is determined that there is no caking in the feeder, and the caking verification result is output.

[0010] Combined with the first aspect, in the fifth implementation manner of the first aspect of this application, based on the caking verification result, when it is confirmed that there is no caking in the feeder, the double exponential smoothing method is used to dynamically update the lag time caused by the viscous batching, including: The lag time of the viscous ingredient is specifically the time required for the actual flow rate of the ingredient to reach the target value after the feeder speed adjustment instruction is issued, which reflects the transmission delay of the viscous ingredient in the pipeline; in the no-load state of the feeder, the speed is increased from 0 to A% of the rated value, and the pressure rise curve is recorded by the pressure sensor at the end of the pipeline, and the time when the pressure first reaches B% of the steady-state value is calculated as the initial lag time, where A and B are positive integers set by the user and A > B; repeated tests are carried out for different spices, and the average value is taken as the reference lag time of the ingredient and stored in the ingredient characteristic database. The lag time is regarded as a dynamic sequence including the current lag value and the rate of change of the lag time; the state is defined, including a horizontal component and a trend component. The horizontal component is the smoothed value of the current lag time, reflecting the immediate lag level, and the trend component is the change trend of the lag time, reflecting the increase and decrease rate of the lag time caused by the change in viscosity; the weights of the horizontal component and the trend component are set and dynamically adjusted according to the viscosity coefficient of the ingredient where, , T represents the thickness of the wall adhesion layer, H represents the humidity of the spice, and D represents D50 in the particle size distribution; the measured value of the current lag time, the viscosity coefficient, the horizontal component at the previous moment, and the trend component at the previous moment are collected; when the trend remains unchanged, the horizontal predicted value is the horizontal component at the previous moment plus the trend component at the previous moment, and the current trend component is set as the trend component at the previous moment; according to the difference between the measured value of the current lag time and the horizontal predicted value, the current horizontal component is obtained; combining the difference between the current horizontal component and the horizontal value at the previous moment, the current trend component is adjusted; the lag time is the sum of the current horizontal component and the current trend component.

[0011] Combined with the first aspect, in the sixth implementation manner of the first aspect of the present application, the fuzzy adaptive compensation based on the lag time is increased to obtain the compensation time, including: Define fuzzy linguistic values for the input variable and the output variable, including: negative large, negative small, zero, positive small, and positive large. The input variable adopts a trapezoidal membership function, and the output variable adopts a singleton fuzzy set; based on the viscous ingredient control experience and the fuzzy linguistic values, establish fuzzy rules for adjusting the compensation gain and the compensation time; convert the input variable into the membership degree of each fuzzy set, activate the corresponding fuzzy rules according to the membership degree, and perform weighted averaging on the outputs of the activated rules to obtain a fuzzy output; calculate the compensation gain based on the fuzzy output using the centroid method; Calculate the basic compensation time based on the lag time and the compensation gain, add a deviation change rate correction amount to correct the basic compensation time, and perform boundary constraint to obtain the compensation time.

[0012] In combination with the first aspect, in a seventh implementation of the first aspect of the present application, the establishment of a hysteresis compensation effect evaluation index, performing no-load and full-load hysteresis tests on each spice based on the hysteresis time, and verifying the compensated coupling error, includes: The evaluation indicators of hysteresis compensation effect include hysteresis time estimation error, batching accuracy error, coupling error and system response time. The hysteresis time estimation error is the deviation percentage between the hysteresis time obtained by double exponential smoothing method and the actual hysteresis time. The batching accuracy error is the deviation percentage between the actual discharge amount and the target discharge amount. The coupling error is the flow deviation percentage of another ingredient caused by the adjustment of a certain ingredient in the strongly coupled batching pair. The system response time is the time from the viscosity mutation to the effectiveness of the compensation strategy. Empty the silo, close all ingredient inlets, and ensure that there is no residual ingredient in the pipeline; step the feeder speed from 0 to A% of the rated speed, synchronously record the timestamp of the feeder control signal and the response timestamp of the pressure sensor at the end of the pipeline, and calculate the inherent lag time; record the inherent lag time at different speeds to form a comparison table of speed and inherent lag time; group by viscosity coefficient, perform benchmark test and compensation test, compare data, and verify the coupling error after compensation.

[0013] In combination with the first aspect, in an eighth implementation manner of the first aspect of the present application, based on the coupling error and the compensation time, a cross compensation matrix is established for the strong coupling pair through relative gain matrix analysis to generate a compensated feeder speed instruction, including: The feeder system is abstracted as a MIMO system, a unit step input is applied to the MIMO system, the steady-state response data is recorded, and the strong coupling pairs are determined; an n×n cross-compensation matrix C is constructed, and the elements in the cross-compensation matrix C represents the compensation coefficient of the j-th feeder to the i-th feeder, and n represents the number of mutually coupled feeders; the compensation intensity factor of the non-diagonal elements in the cross compensation matrix C is adjusted in real time according to the coupling error; for the strongly coupled pair (i, j), the compensation time ratio is introduced as the weight to dynamically update the compensation coefficient; Based on the original speed command vector output by the PID controller, the cross compensation is calculated and physical constraints are imposed; the compensation effect is verified in real time and the matrix is optimized to generate the compensated feeder speed command.

[0014] In a second aspect, the present application provides an automated control system for mixing ingredients in a production line for compounding spices, comprising: Hierarchical Anti-Caking Control Module: It includes: a data fusion unit, an anti-caking control model construction unit, and a hierarchical anti-caking control unit; among them, the data fusion unit acquires sensor data, including the state of the spice and the thickness of the wall adhesion layer, defines a state vector, and fuses the sensor data using the extended Kalman filter; the anti-caking control model construction unit establishes an anti-caking control model based on the fused sensor data and performs model adaptive correction; in the hierarchical anti-caking control unit, the anti-caking control model outputs a caking prevention control strategy to perform hierarchical anti-caking control; Caking Verification Module: It includes: a particle size grading unit and a caking verification unit; among them, when there is caking in the feeder, the particle size grading unit identifies the caking particles through the YOLOv8 model, outputs the particle size distribution, and divides the particle size grades; the caking verification unit executes a disassembly strategy based on the particle size grades to verify the disassembly effect and obtains the caking verification result; Lag and Compensation Calculation Module: It includes: a lag time calculation unit and a compensation time calculation unit; among them, the lag time calculation unit, based on the caking verification result, when it is confirmed that there is no caking in the feeder, dynamically updates the lag time caused by the viscous batching using the double exponential smoothing method; the compensation time calculation unit adds fuzzy adaptive compensation based on the lag time to obtain the compensation time; Feeder Speed Command Generation Module: It includes: a coupling error verification unit and a command generation unit; among them, the coupling error verification unit establishes an evaluation index for the lag compensation effect, conducts lag tests on each spice under no-load and full-load conditions based on the lag time, and verifies the compensated coupling error; the command generation unit, based on the coupling error and the compensation time, analyzes through the relative gain matrix, establishes a cross-compensation matrix for the strongly coupled feeder pair, and generates a compensated feeder speed command.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In this application, multi-source sensor data is fused through the extended Kalman filter, and an adaptive anti-caking control model is established, which can dynamically adjust parameters such as the stirring frequency and vibration amplitude in real time to reduce the caking incidence rate; the identification and particle size grading of caking particles based on the YOLOv8 model can quickly complete caking detection and output the particle size distribution, and execute different disassembly strategies for cakings of different grades.

[0016] 2. In this application, the double exponential smoothing method is used to update the lag time of the viscous batching in real time, combined with the fuzzy adaptive compensation algorithm, to control the lag time prediction error; through the relative gain matrix analysis of the multi-variable coupling strength, a cross-compensation matrix is established for the strongly coupled feeder pair, reducing the coupling error of the multi-component batching and improving the consistency of the compounding ratio.

[0017] 3. This application integrates sensor fusion, visual recognition, adaptive control, and multivariable decoupling technologies to achieve full-process automation from caking prevention, real-time detection, dynamic compensation to multi-machine collaborative control, reducing manual intervention and improving batching efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of the steps of an automatic control method for batching in a production line for spice compounding according to the present invention; Figure 2 is a system structure diagram of an automatic control system for batching in a production line for spice compounding according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] Embodiment 1: As Figure 1 shown in the schematic diagram of the steps of an automatic control method for batching in a production line for spice compounding, this application provides an automatic control method for batching in a production line for spice compounding, including the following steps: Step S100: Obtain sensor data, including the state of the spice and the thickness of the wall adhesion layer, define a state vector, and fuse the sensor data using an extended Kalman filter; establish an anti-caking control model based on the fused sensor data and perform model adaptive correction; the anti-caking control model outputs an anti-caking control strategy for hierarchical anti-caking control; Specifically, the state of the spice includes the humidity, particle size, viscosity, and density of the spice. Define the sensor data as a state vector; set a random noise term for each state vector to reflect the unmodeled dynamics, and establish a state equation based on the physical laws of the state vector change; establish a mapping relationship between the sensor data and the true state, and set a covariance matrix for each sensor data to reflect the measurement uncertainty; Based on the state estimate value at the previous moment, predict the state at the current moment based on the state equation, and calculate the covariance matrix of the predicted state at the current moment; compare the difference between the measured value of the sensor and the predicted value, calculate the Kalman gain, and correct the predicted state based on the Kalman gain to obtain the optimal estimate value; update the covariance matrix to complete a filtering cycle; iterate the extended Kalman filter to fuse the sensor data.

[0021] Furthermore, the anti-caking control model includes a data layer, a model layer, and a policy layer. In the data layer, the fused sensor data is received and normalized. In the model layer, a fuzzy logic controller is used to construct a non-linear mapping relationship, define linguistic variables, and establish an IF-THEN rule base. In the policy layer, the model output is converted into specific device control instructions. The data source for training the model is historical production data, specifically the sensor data and manual processing records corresponding to the caking events of different spices. The training objective is to make the policy level output by the model negatively correlated with the actual caking incidence rate. Define the correction trigger condition. According to the real-time caking data, increase the rule priority of the high-frequency occurrence scenarios and dynamically adjust the membership function.

[0022] Example 2: The historical data covers 5 kinds of spices (chili powder, ginger powder, black pepper powder, Chinese prickly ash powder, cumin powder), including more than 2000 caking event samples. Among them, the slightly caked (particle size ≤ 2mm) accounts for 45%, and the corresponding sensor characteristics are humidity ≤ 15%RH and adhesion layer thickness ≤ 0.8mm; the moderately caked (2 - 5mm) accounts for 35%, corresponding to humidity 15% - 20%RH and adhesion layer thickness 0.8 - 1.5mm; the severely caked (> 5mm) accounts for 20%, corresponding to humidity > 20%RH and adhesion layer thickness > 1.5mm.

[0023] In the fuzzy logic controller, the input variables include humidity (linguistic values: low, medium, high), universe of discourse [0, 30%RH]; wall adhesion layer thickness (linguistic values: thin, medium, thick), universe of discourse [0, 3mm]; viscosity (linguistic values: low, medium, high), universe of discourse [0, 800mPa・s]. The output variable is the control policy level (linguistic values: level 1 - gentle stirring, level 2 - moderate vibration, level 3 - strong vibration + stirring), universe of discourse [1, 3].

[0024] An example of the rule base is as follows: IF humidity = high AND adhesion layer thickness = thick THEN policy level = 3; IF humidity = medium AND viscosity = high THEN policy level = 2; IF adhesion layer thickness = thin AND viscosity = low THEN policy level = 1.

[0025] The prediction accuracy of the caking incidence rate of the trained model reaches 88%, and the negative correlation coefficient between the policy level and the actual caking incidence rate increases from -0.52 to -0.79. When it is detected that caking still occurs continuously 3 times with humidity > 22%RH and policy level = 3, the system automatically shifts the membership function of "humidity = high" 2%RH to the left (i.e., triggers the high-level policy in advance). After correction, the caking incidence rate in this scenario drops from 15% to 4%.

[0026] Step S200: When there are lumps in the feeder, identify the lump particles through the YOLOv8 model, output the particle size distribution, and divide the particle size grades; execute the disassembly strategy based on the particle size grades, verify the disassembly effect, and obtain the lump verification result. Specifically, collect typical lump samples, take images in the laboratory under different lighting conditions, spice states, and background interferences, and use the LabelImg tool to label the lump particles as the training set; rotate, scale, and adjust the brightness of the training set to expand the training set; for the YOLOv8 model, load the COCO pre-trained weights, set the training parameters, use K-means clustering to recalculate the prior box size, and train the YOLOv8 model; input the pre-processed images into the trained YOLOv8 model, and output the bounding box coordinates and class confidence of the lump particles; establish the mapping relationship between pixels and actual sizes through a calibration board, and for each detection box, take the geometric mean of its width and height as the equivalent diameter; perform perspective correction on the detection boxes in the edge area based on the camera installation angle; for the adhering particles within the detection box, use the watershed algorithm for secondary segmentation to improve the particle size calculation accuracy. Statistically analyze the particle sizes of all lump particles in the current frame, generate a particle size histogram at intervals, and output the particle size distribution, including the average particle size, D10, D50, and D90; according to the predefined particle size grades, statistically analyze the quantity ratio and mass ratio of particles in each grade, and divide the particle size grades.

[0027] Furthermore, the particle size grades include first-level lumps, second-level lumps, third-level lumps, fourth-level lumps, and fifth-level lumps; for first-level lumps, start the low-frequency vibrating screen at the feeder outlet, loosen and separate the particle clusters through vibration, and turn on the ultrasonic vibration device on the inner wall of the pipeline to emit pulses to break the weak adhesion force between particles; for second-level lumps, start the double-roll crusher at the feeder inlet, shear the lump particles through differential rotation, and synchronously turn on the pneumatic vibrator at the bottom of the silo to prevent subsequent batching accumulation from affecting the crushing effect; for third-level lumps, based on the treatment measures for second-level lumps, turn on the high-pressure air knife to blow the crushed particles away from the blades to avoid secondary adhesion; for fourth-level lumps, switch the double-roll crusher to the reverse rotation mode, crush the hard lumps through rubbing, start the centrifugal disperser at the end of the pipeline, and use the centrifugal force to throw the large particles towards the wear-resistant lining plate for impact crushing; for fifth-level lumps, stop the feeder, close the valve at the silo outlet, trigger an audible and visual alarm, start the 360° rotating cleaning arm inside the silo, and try to physically remove the super-large lumps. When the super-large lumps are not removed within the set time, trigger the manual maintenance process. Obtain the current particle size distribution. When the qualified standard is met, determine that there are no lumps in the feeder and output the lump verification result.

[0028] Example 3: The caking grades are classified as follows: Grade 1 caking: equivalent diameter ≤ 1mm; Grade 2 caking: 1mm < equivalent diameter ≤ 3mm; Grade 3 caking: 3mm < equivalent diameter ≤ 5mm; Grade 4 caking: 5mm < equivalent diameter ≤ 10mm; Grade 5 caking: equivalent diameter > 10mm.

[0029] Collect 3000 caking images of black pepper powder, covering scenarios with humidity of 8% - 22%RH and particle size D50 = 50 - 150μm, and label more than 12,000 caking particles. After data augmentation, the training set is expanded to 12,000 images, including variant samples with rotation of ±15°, scaling of 0.8 - 1.2 times, and brightness of ±20%. The input image size is 640×640, the batch size is 16, the number of training epochs is 100, the initial learning rate is 0.001, and cosine annealing decay is adopted. K-means clustering obtains 5 prior box sizes: (12,16), (24,32), (40,56), (72,96), (128,160). The mAP@0.5 of the test set reaches 95.2%, and the average detection time is 28ms / frame.

[0030] There is an agglomerated caking mass in the input image. Before perspective correction, it is detected as a Grade 4 caking with an equivalent diameter of 8mm. After being segmented by the watershed algorithm, it is identified as 3 particles (with diameters of 3mm, 4mm, and 5mm respectively), and is corrected to 2 Grade 3 cakings and 1 Grade 4 caking.

[0031] The initial particle size distribution is as follows: average particle size 6.8mm, D10 = 2.1mm, D50 = 5.5mm, D90 = 11.2mm; grade proportion: Grade 3 caking (3 - 5mm) accounts for 35%, Grade 4 caking (5 - 10mm) accounts for 50%, and Grade 5 caking (>10mm) accounts for 15%.

[0032] A Grade 5 caking with an equivalent diameter of 12mm is detected, triggering the rotation of the cleaning arm (rotation speed 10rpm). The clearance rate within 30 seconds is 90%, and the remaining 1 caking with a diameter of 8mm is downgraded to Grade 4. The double-roll crusher switches to reverse rotation (rotation speed difference 300rpm), and the centrifugal disperser rotates at 2000rpm. After processing, 90% of the particles are < 5mm, and the remaining 10% are Grade 3 cakings of 3 - 5mm. Start the forward shear of the double-roll crusher (rotation speed 1500rpm) + high-pressure air knife (air pressure 0.6MPa). After processing, 85% of the particles are < 3mm, and 15% are Grade 2 cakings of 1 - 3mm. The double-roll crusher maintains forward rotation, and start the pneumatic vibrator (frequency 50Hz). After processing, 98% of the particles are < 1mm, meeting the standard within the range of Grade 1 caking.

[0033] Particle size distribution after processing: average particle size 0.8 mm, D10 = 0.3 mm, D50 = 0.7 mm, D90 = 1.1 mm; grade ratio: primary agglomerates (≤1 mm) account for 99%, secondary agglomerates account for 1%, and there are no agglomerates of grades three to five; disassembly efficiency: it takes 120 seconds from detection to complete processing, improving the efficiency by 50% compared with the traditional single crushing method.

[0034] Step S300: Based on the caking verification result, when it is confirmed that there is no caking in the feeder, the double exponential smoothing method is used to dynamically update the lag time caused by viscous batching; based on the lag time, a fuzzy adaptive compensation is added to obtain the compensation time; Specifically, the lag time of viscous batching is specifically the time required for the actual flow rate of the batching to reach the target value after the feeder speed adjustment instruction is issued, reflecting the transmission delay of the viscous batching in the pipeline; in the no-load state of the feeder, the speed is increased from 0 to A% of the rated value, and the pressure rise curve is recorded by the pressure sensor at the end of the pipeline, and the time when the pressure first reaches B% of the steady-state value is calculated as the initial lag time, where A and B are positive integers set by the user and A > B; repeated tests are carried out for different spices, and the average value is taken as the reference lag time of the batching and stored in the batching characteristic database; The lag time is regarded as a dynamic sequence including the current lag value and the change rate of the lag time; the state is defined, including the horizontal component and the trend component. The horizontal component is the smoothed value of the current lag time, reflecting the immediate lag level, and the trend component is the change trend of the lag time, reflecting the increase and decrease rate of the lag time caused by the change of viscosity; the weights of the horizontal component and the trend component are set and dynamically adjusted according to the viscosity coefficient of the batching where, , T represents the thickness of the wall adhesion layer, H represents the humidity of the spice, and D represents D50 in the particle size distribution; the measured value of the current lag time, the viscosity coefficient, the horizontal component at the previous moment, and the trend component at the previous moment are collected; when the trend remains unchanged, the horizontal predicted value is the horizontal component at the previous moment plus the trend component at the previous moment, and the current trend component is set to the trend component at the previous moment; according to the difference between the measured value of the current lag time and the horizontal predicted value, the current horizontal component is obtained; combining the difference between the current horizontal component and the horizontal value at the previous moment, the current trend component is adjusted; the lag time is the sum of the current horizontal component and the current trend component.

[0035] Further, define fuzzy linguistic values for the input variables and output variables, including: negative large, negative small, zero, positive small, and positive large. The input variables adopt trapezoidal membership functions, and the output variables adopt singleton fuzzy sets. Based on the viscous batching control experience and the fuzzy linguistic values, establish fuzzy rules for adjusting the compensation gain and compensation time. Convert the input variables into the membership degrees of each fuzzy set, activate the corresponding fuzzy rules according to the membership degrees, and perform weighted averaging on the outputs of the activated rules to obtain a fuzzy output. Calculate the compensation gain based on the fuzzy output using the centroid method. Calculate the basic compensation time based on the lag time and compensation gain, add a deviation change rate correction amount to correct the basic compensation time, and perform boundary constraints to obtain the compensation time.

[0036] Example 4: Feeder parameters: rated speed 200 rpm, pipe length 5 m, inner diameter 50 mm. Initial test parameters: A = 80 (speed increased from 0 to 80% of the rated value), B = 90 (pressure reached 90% of the steady-state value). The viscosity coefficient calculation formula is viscosity × humidity / D50 (viscosity unit mPa・s, humidity %RH, D50 unit μm).

[0037] No-load test data are as follows (repeated 5 times): Test 1: lag time 4.2 s; Test 2: 4.5 s; Test 3: 4.0 s; Test 4: 4.3 s; Test 5: 4.4 s; Average value: 4.28 s, which is stored in the database as the reference lag time.

[0038] Full-load test data are as follows (humidity 18%RH, D50 = 60 μm, viscosity 450 mPa・s): Actual lag time: 6.8 s (due to significantly higher material viscosity than no-load, the transmission delay increases).

[0039] Use the double exponential smoothing method to dynamically update the lag time. For input 1: lag time deviation (measured value - reference value), current deviation: 7.2 - 4.28 = 2.92 s, linguistic value "positive large". For input 2: deviation change rate (this deviation - last deviation), change rate: 2.92 - 2.72 (last deviation was 7.0 - 4.28 = 2.72) = 0.2, linguistic value "positive small". Output: compensation gain K, fuzzy rule: IF deviation = positive large AND change rate = positive small THEN K = 1.2 (enhanced compensation).

[0040] Basic compensation time: lag time estimate × compensation gain = 7.26 × 1.2 = 8.71 s; Deviation change rate correction amount: +0.5 s (due to positive change rate, appropriately extend the compensation); Boundary constraint (maximum 10 s): Final compensation time = min(8.71 + 0.5, 10) = 9.21 s.

[0041] Step S400: Establish an evaluation index for the lag compensation effect. Based on the lag time, conduct no-load and full-load lag tests on each spice to verify the coupled error after compensation. Based on the coupled error and compensation time, analyze through the relative gain matrix, establish a cross-compensation matrix for the strong coupling pair, and generate the feed machine speed command after compensation.

[0042] Specifically, the evaluation index for the lag compensation effect includes the lag time estimation error, batching accuracy error, coupled error, and system response time. The lag time estimation error is the deviation percentage between the lag time obtained by the double exponential smoothing method and the actual lag time. The batching accuracy error is the deviation percentage between the actual feeding amount and the target feeding amount. The coupled error is the flow deviation percentage of another ingredient caused by the adjustment of one ingredient in the strong coupling batching pair. The system response time is the time from the viscosity mutation to the effective compensation strategy. Empty the bin, close all batching inlets, and ensure that there is no residual batching in the pipeline. The speed of the feed machine steps from 0 to A% of the rated speed. Synchronously record the time stamp of the feed machine control signal and the response time stamp of the pressure sensor at the end of the pipeline, and calculate the inherent lag time. Record the inherent lag time at different speeds to form a comparison table of speed and inherent lag time. Group according to the viscosity coefficient, conduct benchmark tests and compensation tests, and compare the data to verify the coupled error after compensation.

[0043] Furthermore, abstract the feed machine system as a MIMO system, apply a unit step input to the MIMO system, record the steady-state response data, and determine the strong coupling pair. Construct an n×n cross-compensation matrix C. The element in the cross-compensation matrix C represents the compensation coefficient of the j-th feed machine for the i-th feed machine, and n represents the number of mutually coupled feed machines. Adjust the compensation intensity factor of the non-diagonal elements in the cross-compensation matrix C in real time according to the coupled error. For the strong coupling pair (i,j), introduce the compensation time ratio as the weight to update the dynamic compensation coefficient. Based on the original speed command vector output by the PID controller, calculate the cross-compensation and apply physical constraints. Conduct real-time verification of the compensation effect and matrix optimization to generate the feed machine speed command after compensation.

[0044] Example 5: Take the two-component batching system of chili powder (low viscosity) and ginger powder (high viscosity) as an example, and set the following experimental conditions: Number of feed machines: 2 (numbered M1, M2), respectively conveying chili powder and ginger powder, and the rated speed of both is 300 rpm. Viscosity coefficient grouping: Chili powder (viscosity coefficient < 200), ginger powder (viscosity coefficient ≥ 200). Strong coupling determination threshold: A coupling error > 10% is regarded as a strong coupling pair. Conduct lag tests and compensation effect evaluations.

[0045] When a step signal is input to M1 (rotation speed from 150 rpm → 200 rpm), at steady state: The actual flow rate change of M1: +20% (target value). The actual flow rate change of M2: -12% (coupling interference). When a step signal is input to M2 (rotation speed from 150 rpm → 200 rpm), at steady state: The actual flow rate change of M2: +18% (target value). The actual flow rate change of M1: +8% (coupling interference).

[0046] Calculate the RGA matrix. If the absolute value of the off-diagonal element > 0.1, it is determined that M1 and M2 are a strongly coupled pair. Initialize the cross-compensation matrix. The element C[1,2] represents the compensation coefficient of M2 to M1, and C[2,1] represents the compensation coefficient of M1 to M2.

[0047] When it is detected that the adjustment of M1 causes a 4.2% flow rate deviation of M2 (target ≤ 5%), maintain the current compensation coefficient; if the deviation > 5%, update the compensation coefficient.

[0048] Generate the compensated rotation speed command: Original PID command: M1 = 200 rpm, M2 = 200 rpm.

[0049] Cross-compensation calculation: Command of M1 after compensation: 200 + (-0.3)×200 = 140 rpm (canceling the coupling interference of M2).

[0050] Command of M2 after compensation: 200 + (-0.5)×200 = 100 rpm (canceling the coupling interference of M1).

[0051] Physical constraint correction: The rotation speed needs to be ≥ 50 rpm. Final command: M1 = 140 rpm, M2 = 100 rpm.

[0052] Example 6: As Figure 2 Shown in the system structure diagram of an automatic control system for ingredient batching in a spice compounding production line, the present application provides an automatic control system for ingredient batching in a spice compounding production line, including: Hierarchical anti-caking control module: including: a data fusion unit, an anti-caking control model construction unit, and a hierarchical anti-caking control unit; wherein, the data fusion unit acquires sensor data, including spice state and wall adhesion layer thickness, defines a state vector, and fuses the sensor data using an extended Kalman filter; the anti-caking control model construction unit establishes an anti-caking control model based on the fused sensor data and performs model adaptive correction; the anti-caking control model in the hierarchical anti-caking control unit outputs an anti-caking prevention control strategy for hierarchical anti-caking control; Caking verification module: including: particle size grade division unit and caking verification unit; among them, when there is caking in the feeder, the particle size grade division unit identifies caking particles through the YOLOv8 model, outputs the particle size distribution, and divides the particle size grades; the caking verification unit executes the disassembly strategy based on the particle size grades, verifies the disassembly effect, and obtains the caking verification result; Lag and compensation calculation module: including: lag time calculation unit and compensation time calculation unit; among them, the lag time calculation unit, based on the caking verification result, when it is confirmed that there is no caking in the feeder, uses the double exponential smoothing method to dynamically update the lag time caused by viscous batching; the compensation time calculation unit adds fuzzy adaptive compensation based on the lag time to obtain the compensation time; Feeder speed command generation module: including: coupling error verification unit and command generation unit; among them, the coupling error verification unit establishes an evaluation index for the lag compensation effect, conducts lag tests on each spice under no-load and full-load conditions based on the lag time, and verifies the coupling error after compensation; the command generation unit, based on the coupling error and the compensation time, analyzes through the relative gain matrix, establishes a cross-compensation matrix for the strong coupling pair, and generates the compensated feeder speed command.

[0053] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed claims.

Claims

1. An automatic control method for production line batching of spice compounding, characterized in that, It includes the following steps: Obtain sensor data, including the state of spices and the thickness of the wall adhesion layer, define a state vector, and fuse the sensor data using the extended Kalman filter; establish an anti-caking control model based on the fused sensor data and perform model adaptive correction; the anti-caking control model outputs an anti-caking control strategy and performs hierarchical anti-caking control; When there are caked masses in the feeder, identify the caked particles through the YOLOv8 model, output the particle size distribution, and divide the particle size grades; execute a disassembly strategy based on the particle size grades, verify the disassembly effect, and obtain a caking verification result; Based on the caking verification result, when it is confirmed that there are no caked masses in the feeder, use the double exponential smoothing method to dynamically update the lag time caused by viscous batching; Increase the fuzzy adaptive compensation based on the lag time to obtain the compensation time; Establish an evaluation index for the lag compensation effect, conduct lag tests for each spice under no-load and full-load conditions based on the lag time, and verify the coupled error after compensation; Based on the coupled error and the compensation time, through relative gain matrix analysis, establish a cross-compensation matrix for strongly coupled pairs and generate a compensated feeder speed command.

2. The automatic control method for production line batching of spice compounding according to claim 1, wherein, The obtaining of the sensor data, including the state of spices and the thickness of the wall adhesion layer, defining a state vector, and fusing the sensor data using the extended Kalman filter, includes: The state of spices includes the humidity, particle size, viscosity, and density of the spices. Define the sensor data as a state vector; set a random noise term for each state vector to reflect the unmodeled dynamics, and establish a state equation based on the physical laws of the state vector change; establish a mapping relationship between the sensor data and the true state, and set a covariance matrix for each sensor data to reflect the measurement uncertainty; Based on the state estimate value at the previous moment, predict the state at the current moment according to the state equation, and calculate the covariance matrix of the predicted state at the current moment; compare the difference between the measured value and the predicted value of the sensor, calculate the Kalman gain, and correct the predicted state based on the Kalman gain to obtain the optimal estimate value; update the covariance matrix to complete a filtering cycle; iterate the extended Kalman filter to fuse the sensor data.

3. A method for automatic control of production line batching for spice compounding according to claim 1, characterized in that, The establishing of the anti-caking control model based on the fused sensor data and performing model adaptive correction includes: The anti-caking control model includes a data layer, a model layer, and a strategy layer; in the data layer, receive the fused sensor data and perform normalization processing; in the model layer, use a fuzzy logic controller to construct a nonlinear mapping relationship, define linguistic variables, and establish an IF-THEN rule base; in the strategy layer, convert the model output into specific device control instructions; the data source for training the model is historical production data, specifically the sensor data and manual processing records corresponding to the caking events of different spices, and the training objective is to make the strategy level output by the model negatively correlated with the actual caking incidence rate; Define the correction trigger condition, and according to the real-time caking data, increase the rule priority of high-frequency occurrence scenarios and dynamically adjust the membership function.

4. The automatic control method for production line batching of spice compounding according to claim 1, wherein, When there are caked masses in the feeder, identifying the caked particles through the YOLOv8 model, outputting the particle size distribution, and dividing the particle size grades, includes: Collect typical caked samples, take images in the laboratory under simulated different lighting conditions, spice states, and background interferences, and use the LabelImg tool to annotate the caked particles as the training set; rotate, scale, and adjust the brightness of the training set to expand the training set; for the YOLOv8 model, load the COCO pre-trained weights, set the training parameters, recalculate the prior box sizes using K-means clustering, and perform YOLOv8 model training; input the preprocessed images into the trained YOLOv8 model to output the bounding box coordinates and class confidence of the caked particles; establish the mapping relationship between pixels and actual sizes through a calibration board, and for each detection box, take the geometric mean of its width and height as the equivalent diameter; based on the camera installation angle, perform perspective correction on the detection boxes in the edge area; for the adhered particles within the detection box, use the watershed algorithm for secondary segmentation to improve the particle size calculation accuracy; Statistically calculate the particle sizes of all caked particles in the current frame, generate a particle size histogram at intervals, and output the particle size distribution, including the average particle size, D10, D50, and D90; according to the predefined particle size grades, statistically calculate the proportion of the number and mass of particles in each grade and divide the particle size grades.

5. The automated control method for ingredient preparation in a production line for spice compounding according to claim 1, characterized in that, Execute the disassembly strategy based on the particle size grades, conduct disassembly effect verification, and obtain the caking verification results, including: The particle size grades include first-level caking, second-level caking, third-level caking, fourth-level caking, and fifth-level caking; for first-level caking, start the low-frequency vibrating screen at the outlet of the feeder, loosen and separate the particle clusters through vibration, and turn on the ultrasonic vibration device on the inner wall of the pipeline to emit pulses to break the weak adhesion force between particles; for second-level caking, start the double-roll crusher at the inlet of the feeder, shear the caked particles through differential rotation, and simultaneously turn on the pneumatic vibrator at the bottom of the bin to prevent subsequent batching accumulation from affecting the crushing effect; for third-level caking, based on the treatment measures for second-level caking, turn on the high-pressure air knife to blow the crushed particles away from the blades to avoid secondary adhesion; for fourth-level caking, switch the double-roll crusher to the reverse rotation mode, crush the hard caking through rubbing, and start the centrifugal disperser at the end of the pipeline to use centrifugal force to throw the large particles towards the wear-resistant lining plate for impact crushing; for fifth-level caking, stop the feeder, close the valve at the outlet of the bin, trigger an audible and visual alarm, start the 360° rotating cleaning arm inside the bin, and attempt to physically remove the super-large caking. When it is not removed within the set time, trigger the manual maintenance process; Obtain the current particle size distribution. When it meets the qualified standard, determine that there is no caking in the feeder and output the caking verification results.

6. The automatic control method for production line batching of spice compounding according to claim 1, characterized in that, Based on the caking verification results, when it is confirmed that there is no caking in the feeder, use the double exponential smoothing method to dynamically update the lag time caused by viscous batching, including: The lag time of the viscous ingredient is specifically the time required for the actual flow rate of the ingredient to reach the target value after the feeder speed adjustment command is issued, reflecting the transmission delay of the viscous ingredient in the pipeline. In the no-load state of the feeder, the speed is increased from 0 to A% of the rated value, and the pressure rise curve is recorded through the pressure sensor at the end of the pipeline. The time when the pressure first reaches B% of the steady-state value is calculated as the initial lag time, where A and B are positive integers set by the user and A > B. Repeat the test for different spices and take the average value as the reference lag time of the ingredient, which is stored in the ingredient characteristic database. The lag time is regarded as a dynamic sequence including the current lag value and the change rate of the lag time; states are defined, including a horizontal component and a trend component, where the horizontal component is the smoothed value of the current lag time, reflecting the instantaneous lag level, and the trend component is the change trend of the lag time, reflecting the increase or decrease rate of the lag time caused by the change in viscosity; the weights of the horizontal component and the trend component are set and dynamically adjusted according to the viscosity coefficient of the ingredients where , T represents the thickness of the wall adhesion layer, H represents the humidity of the spice, and D represents D50 in the particle size distribution; the measured value of the current lag time, the viscosity coefficient, the horizontal component at the previous moment, and the trend component at the previous moment are collected; when the trend remains unchanged, the horizontal predicted value is the horizontal component at the previous moment plus the trend component at the previous moment, and the current trend component is set as the trend component at the previous moment; the current horizontal component is obtained according to the difference between the measured value of the current lag time and the horizontal predicted value; the current trend component is adjusted by combining the difference between the current horizontal component and the horizontal value at the previous moment; the lag time is the sum of the current horizontal component and the current trend component.

7. The automatic control method for production line batching of spice compounding according to claim 1, characterized in that, Increasing fuzzy adaptive compensation based on the lag time to obtain the compensation time includes: Defining fuzzy linguistic values for the input variable and the output variable, including: negative large, negative small, zero, positive small, and positive large. The input variable uses a trapezoidal membership function, and the output variable uses a singleton fuzzy set. Based on the viscous ingredient control experience and the fuzzy linguistic values, establish fuzzy rules for adjusting the compensation gain and the compensation time. Convert the input variable into the membership degrees of each fuzzy set, activate the corresponding fuzzy rules according to the membership degrees, and perform weighted averaging on the outputs of the activated rules to obtain the fuzzy output. Calculate the compensation gain based on the fuzzy output using the centroid method. Calculate the basic compensation time based on the lag time and the compensation gain, add the deviation change rate correction amount to correct the basic compensation time, and perform boundary constraints to obtain the compensation time.

8. An automatic control method for production line batching of spice compounding according to claim 1, characterized in that, Establishing the lag compensation effect evaluation index, conducting lag tests on each spice under no-load and full-load conditions based on the lag time, and verifying the coupling error after compensation, including: The lag compensation effect evaluation index includes the lag time estimation error, the ingredient accuracy error, the coupling error, and the system response time. The lag time estimation error is the percentage deviation between the lag time obtained by the double exponential smoothing method and the actual lag time. The ingredient accuracy error is the percentage deviation between the actual feeding amount and the target feeding amount. The coupling error is the percentage flow deviation of one ingredient caused by the adjustment of another ingredient in the strong coupling ingredient pair. The system response time is the time from the viscous mutation to the effective compensation strategy. Empty the silo, close all ingredient inlets, and ensure that there is no residual ingredient in the pipeline. Step the feeder speed from 0 to A% of the rated speed, synchronously record the time stamp of the feeder control signal and the response time stamp of the pressure sensor at the end of the pipeline, and calculate the inherent lag time. Record the inherent lag time at different speeds to form a comparison table of speed and inherent lag time. Group by the viscosity coefficient, conduct the reference test and the compensation test, and compare the data to verify the coupling error after compensation.

9. An automatic control method for production line batching of spice compounding according to claim 1, characterized in that Based on the coupling error and the compensation time, through relative gain matrix analysis, establish a cross-compensation matrix for the strong coupling pair and generate the compensated feeder speed command, including: Abstract the feeder system as a MIMO system, apply a unit step input to the MIMO system, record the steady-state response data, and determine the strong coupling pairs; construct an n×n cross-compensation matrix C, and the elements in the cross-compensation matrix C represent the compensation coefficient of the jth feeder to the ith feeder, and n represents the number of feeders coupled with each other; adjust the compensation strength factor of the non-diagonal elements in the cross-compensation matrix C in real time according to the coupling error; for the strong coupling pair (i, j), introduce the compensation time ratio as the weight to update the dynamic compensation coefficient; Based on the original speed command vector output by the PID controller, calculate the cross-compensation and apply physical constraints. Conduct real-time verification of the compensation effect and matrix optimization to generate the compensated feeder speed command.

10. An automatic control system for production line batching of spice compounding, which uses an automatic control method for production line batching of spice compounding according to any one of claims 1-9, characterized in that Including: Hierarchical anti-caking control module: It includes: a data fusion unit, an anti-caking control model construction unit, and a hierarchical anti-caking control unit; among them, the data fusion unit acquires sensor data, including the state of spices and the thickness of the wall adhesion layer, defines a state vector, and fuses the sensor data using the extended Kalman filter; the anti-caking control model construction unit establishes an anti-caking control model based on the fused sensor data and performs model adaptive correction; in the hierarchical anti-caking control unit, the anti-caking control model outputs a caking prevention control strategy to perform hierarchical anti-caking control; Caking verification module: It includes: a particle size grade division unit and a caking verification unit; among them, when there is caking in the feeder, the particle size grade division unit identifies caking particles through the YOLOv8 model, outputs the particle size distribution, and divides the particle size grades; the caking verification unit executes a disassembly strategy based on the particle size grades to verify the disassembly effect and obtains a caking verification result; Lag and compensation calculation module: It includes: a lag time calculation unit and a compensation time calculation unit; among them, the lag time calculation unit, based on the caking verification result, when it is confirmed that there is no caking in the feeder, dynamically updates the lag time caused by viscous batching using the double exponential smoothing method; the compensation time calculation unit adds fuzzy adaptive compensation based on the lag time to obtain the compensation time; Feeder speed command generation module: It includes: a coupling error verification unit and a command generation unit; among them, the coupling error verification unit establishes an evaluation index for the lag compensation effect, conducts lag tests for each spice under no-load and full-load conditions based on the lag time, and verifies the coupling error after compensation; the command generation unit, based on the coupling error and the compensation time, analyzes through the relative gain matrix, establishes a cross compensation matrix for the strong coupling pair, and generates a compensated feeder speed command.

Images