A high-precision intelligent formula dispensing collaborative control method and system for fragrances and flavors

Abstract

The application discloses a kind of essence and fragrance high-precision intelligent formula batching collaborative control method and system, belong to essence intelligent batching technical field, by the feature extraction of encrypted material formula data is determined after target feeding quantity, real-time monitoring remaining material quantity in container in feeding process, when detecting that remaining amount is lower than the preset trace material weight threshold, automatically generate segmented adjustment instruction, utilize the material flow compensation mapping set established in advance dynamically set nozzle control length, obtain the actual discharge length after time compensation, and the discharge deviation of current feeding cycle is input gradient promotion integrated prediction model and carries out closed-loop correction, to obtain the corrected feeding quantity of next feeding cycle, realizes the accurate control and self-adapting optimization of trace material feeding, effectively improves batching precision and production stability.

Description

Technical Field

[0001] This invention relates to the field of intelligent flavor formulation technology, and in particular discloses a method and system for high-precision intelligent formulation and ingredient coordination control of flavors and fragrances. Background Technology

[0002] Existing flavor ingredient technology generally adopts a crude control mode of fixed nozzle duration and static weighing compensation. It relies heavily on manual cup changing, manual formula recording, and manual flow calibration. It has not achieved encrypted verification of formula data and intelligent modeling of physical properties to predict the amount of feed. It also lacks an integrated high-precision ingredient control mechanism that includes adaptive switching between coarse and fine feed, dynamic flow compensation, correction of volatilization loss time, suppression of alignment error, and multi-station collaborative closed-loop correction.

[0003] The core defects and corresponding technical consequences of the existing technology are as follows: First, the use of fixed discharge parameters cannot adapt to dynamic flow fluctuations caused by different material viscosities, temperatures, pipeline residues, and nozzle wear. Errors are significantly amplified under micro-feeding conditions, easily leading to imbalances in aroma proportions and batch-to-batch aroma inconsistencies. Second, relying solely on passive correction based on back-end weighing results cannot dynamically compensate for flow deviations in real time during the discharge process, and cannot eliminate systematic errors caused by dynamic disturbances. Third, the lack of time compensation for the volatile characteristics of fragrances results in uncontrollable volatilization losses, further reducing the accuracy of micro-feeding. Fourth, manual participation in the entire process of transfer, alignment, and recording not only leads to feeding deviations caused by alignment misalignment, material splashing, and container wall adhesion, but also easily introduces external contamination and damages product purity. Furthermore, the circulation of paper formulas poses a serious risk of technology leakage. Fifth, the lack of multi-station collaborative scheduling and cyclic closed-loop iterative correction mechanisms makes it impossible to continuously optimize the accuracy of the ingredients, resulting in poor mass production stability and difficulty in meeting the high-precision, high-confidentiality, clean, and automated production requirements of high-end fragrances and flavors.

[0004] Therefore, how to achieve secure analysis of encrypted formulas, accurate feeding prediction based on multi-dimensional material properties, dynamically adapt to working condition disturbances to complete flow and volatilization compensation, adaptively switch between coarse and fine feeding modes, and achieve multi-station collaborative closed-loop accurate feeding are key technical challenges that urgently need to be overcome in the field of intelligent ingredient mixing of high-end flavors and fragrances. Summary of the Invention

[0005] This invention provides a method and system for high-precision intelligent formulation and ingredient coordination control of flavorings and fragrances, aiming to solve the technical defect that it is difficult to guarantee the accuracy of micro-material dosing.

[0006] One aspect of this invention relates to a high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors, comprising the following steps:

[0007] S100. Obtain encrypted material formula data, perform decryption and integrity verification on the encrypted material formula data, extract the physical and chemical dimension features of each ingredient, and construct a multi-dimensional material property feature set; use a preset machine learning algorithm to fit the multi-dimensional material property feature set and accurately predict the target feed amount of each material.

[0008] S200: Match the material holding container of the corresponding specification based on the target feeding amount, drive the material transfer actuator to complete the transfer and positioning of the material holding container, and obtain the container positioning coordinates corresponding to the material holding container;

[0009] S300: Based on the container positioning coordinates, start the weight monitoring module of the corresponding workstation to collect the weight data of the material already put into the material holding container in real time, and obtain the amount of remaining material to be added under the current batching condition.

[0010] S400: Compare the remaining amount of material to be fed with the preset weight threshold of trace material, and adaptively switch between coarse feeding mode and trace fine feeding mode based on the comparison result, and generate segmented adjustment instructions for the conveying speed corresponding to the working condition.

[0011] S500: Based on the segmented adjustment command, retrieve the pre-built material flow compensation mapping set, and combine the current material property parameters with the remaining amount of material to be added to obtain the real-time flow compensation value.

[0012] S600: The basic opening and closing control time of the feeding nozzle is dynamically corrected and adjusted by using real-time flow compensation value, and the time compensation is combined with the material volatilization loss characteristics to obtain the actual discharge time that is adapted to the real-time working conditions.

[0013] S700 completes a single batching operation according to the actual discharge time and collects the discharge deviation data of the current batching cycle. The discharge deviation data, physical and chemical characteristics of the materials, and station operating parameters are input into the trained gradient boosting integrated prediction model to complete the closed-loop correction of the batching amount and obtain the corrected batching amount for the next batching cycle. Based on the corrected batching amount, multi-batching station collaborative scheduling and high-precision continuous batching control are realized.

[0014] Furthermore, step S100 specifically includes:

[0015] S110. The RSA asymmetric encryption and decryption algorithm is used to perform modular exponentiation decryption on the encrypted material formula ciphertext data and the local preset private key to obtain the decrypted plaintext formula dataset. The data integrity is verified by combining the SHA-256 hash digest operator with the unique timestamp and the original check code. Only the compliant formula data that passes the verification is retained, and the tampered abnormal data is removed.

[0016] S120: Extract four core physicochemical parameters—material density, dynamic viscosity, volatility coefficient, and solubility—from compliant formulation data to construct a structured multidimensional material property feature vector.

[0017] S130. Use a random forest regression model containing decision trees to integrate and fit the multidimensional physical property feature vectors of materials for prediction. Take the mean of all decision tree outputs and constrain the results to be non-negative to obtain the target feed amount for each ingredient.

[0018] Furthermore, the material holding container is a material cup, and the material transfer actuator is a transfer robot. Step S200 specifically includes:

[0019] S210. Calculate the theoretical filling volume of the material by combining the target feeding amount with the material density, set the safety margin coefficient, screen the material cups whose rated volume meets the constraints as matching containers, and establish a one-to-one mapping relationship between feeding amount, material cup number, and material batching station coordinates.

[0020] S220 drives the transfer robot to complete the grasping, shifting and placement of the cup. It collects the pixel coordinates of the cup image through an industrial camera, and uses the pre-calibrated camera intrinsic parameter matrix and pixel physical scaling factor to convert the pixel coordinates into the three-dimensional world coordinates of the production line to obtain the precise positioning coordinates of the container.

[0021] S230. Calculate the Euclidean distance between the current material cup position and the reference coordinates of each idle preset workstation as the workstation allocation cost, select the workstation number corresponding to the minimum cost value as the target allocation workstation, and complete the dynamic allocation of multiple workstations.

[0022] Further, step S300 includes:

[0023] S310. Activate the high-precision weighing sensor at the corresponding workstation based on the container positioning coordinates, and perform sliding mean filtering and noise reduction processing on the original weighing sampling sequence using a fixed sliding time window to eliminate weighing pulse noise caused by mechanical vibration and electromagnetic interference, and obtain the real-time net weight of the material after filtering.

[0024] S320. Subtract the filtered real-time net weight of the material from the target feed amount, and ensure that the value is non-negative by the maximum value constraint, to calculate the remaining amount of material to be fed under the current working condition.

[0025] Furthermore, step S400 specifically includes:

[0026] Preset the weight threshold for trace materials, the conveying speed for coarse feeding, and the conveying speed for fine feeding;

[0027] When the remaining amount of material to be fed is greater than the preset micro-material weight threshold, the coarse feeding mode is executed, and the high-speed conveyor gear is matched to complete the rapid feeding.

[0028] When the remaining amount of material to be added is less than or equal to the preset micro-material weight threshold, the system switches to micro-precision feeding mode, matches the low-speed conveyor gear to complete precise feeding, and realizes coarse and fine dual-stage adaptive segmented speed regulation control.

[0029] Furthermore, the method for constructing the material flow compensation mapping set in step S500 includes:

[0030] Actual discharge deviation data of different viscosities and volatile flavors under multiple conveying speeds and multiple sets of remaining feed amounts are collected in advance. Discrete lookup table mapping function is constructed using material dynamic viscosity, real-time conveying speed and remaining feed amount as three-dimensional indexes. During online operation, the three sets of parameters of current working condition material dynamic viscosity, real-time conveying speed and remaining feed amount are input into the discrete lookup table mapping function to match and obtain real-time flow compensation value.

[0031] Furthermore, step S600 specifically includes:

[0032] S610. Based on the preset basic opening and closing time of the nozzle, the flow rate deviation is linearly corrected by combining the flow rate-time linear compensation coefficient and the real-time flow rate compensation value to obtain the nozzle opening and closing time after the initial flow rate correction.

[0033] S620. Combining the material volatility coefficient and the volatilization time compensation coefficient, calculate the supplementary discharge time required for volatilization loss based on the initial correction time of the flow rate.

[0034] S630. Add the flow correction time and the evaporation compensation time, and constrain them within the safe time interval of the equipment by the value range truncation operator to obtain the actual discharge time of the nozzle that is adapted to the real-time working conditions.

[0035] Furthermore, step S700 specifically includes:

[0036] S710. Collect the final actual net weight of material fed in this cycle, calculate the absolute and relative deviations of single feeding respectively, and construct a standardized discharge deviation feature vector.

[0037] S720. Input the fused feature vector of the material discharge deviation into the gradient boosting integrated prediction model containing the weak regression tree, and impose a value range truncation constraint on the model output to obtain the material discharge deviation correction coefficient.

[0038] S730: The target feeding amount for the current cycle is scaled and corrected using the feeding deviation correction coefficient to obtain the corrected feeding amount for the next feeding cycle; at the same time, the current working condition parameters and discharge deviation data are updated to the material flow compensation mapping set to achieve iterative optimization of the compensation knowledge base;

[0039] S740. Taking the time balance of the batching end time of all online batching stations as the optimization objective, construct the multi-station collaborative time balance scheduling objective function, solve the optimal correction feeding quantity distribution scheme of each station, and minimize the total deviation between the batching end time of each station and the average end time.

[0040] Another aspect of the present invention relates to a high-precision intelligent formulation and ingredient coordination control system for fragrances and flavors, used to execute the above-mentioned high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors, comprising:

[0041] The target feed quantity prediction module is used to acquire encrypted material formula data, perform decryption and integrity verification on the encrypted material formula data, extract the physical and chemical features of each ingredient, and construct a multi-dimensional material property feature set; a preset machine learning algorithm is used to fit the multi-dimensional material property feature set to accurately predict the target feed quantity of each material.

[0042] The container positioning coordinate acquisition module is used to match the corresponding specifications of the material holding container based on the target feeding amount, drive the material transfer actuator to complete the transfer and positioning of the material holding container, and obtain the container positioning coordinates corresponding to the material holding container.

[0043] The remaining material to be added module is used to activate the weight monitoring module of the corresponding workstation based on the container positioning coordinates, collect the weight data of the material already added in the material holding container in real time, and obtain the remaining material to be added under the current batching condition.

[0044] The conveying speed segment adjustment instruction generation module is used to compare the remaining amount of material to be fed with the preset micro-material weight threshold, adaptively switch between coarse feeding mode and micro-fine feeding mode based on the comparison result, and generate conveying speed segment adjustment instructions for the corresponding working conditions.

[0045] The real-time flow compensation value acquisition module is used to retrieve a pre-built material flow compensation mapping set based on the segmented adjustment command, and match the current material property parameters and the remaining amount of material to be added to obtain the real-time flow compensation value.

[0046] The actual discharge time acquisition module is used to dynamically correct and adjust the basic opening and closing control time of the batching nozzle using real-time flow compensation values, and to compensate for the time by combining the material volatilization loss characteristics to obtain the actual discharge time adapted to the real-time working conditions.

[0047] The workstation collaborative scheduling and continuous batching control module is used to complete a single batching operation according to the actual discharge time and collect the discharge deviation data of the current batching cycle. The discharge deviation data, physical and chemical characteristics of the materials, and workstation operating parameters are input into the trained gradient boosting integrated prediction model to complete the closed-loop correction of the batching amount and obtain the corrected batching amount for the next batching cycle. Based on the corrected batching amount, multi-batching workstation collaborative scheduling and high-precision continuous batching control are realized.

[0048] The beneficial effects achieved by this invention are as follows:

[0049] The present invention provides a high-precision intelligent formulation and ingredient coordination control method and system for flavorings and fragrances. Addressing the challenge of ensuring accurate micro-material dispensing in business scenarios, this method extracts features from encrypted material formulation data to determine the target dispensing amount. During the dispensing process, it monitors the remaining material in the container in real time. When the remaining amount is detected to be below a preset micro-material weight threshold, it automatically generates segmented adjustment instructions. A pre-established material flow compensation mapping set is used to dynamically adjust the nozzle control duration, obtaining the actual dispensing duration after time compensation. The dispensing deviation of the current dispensing cycle is input into a gradient boosting integrated prediction model for closed-loop correction, thereby obtaining the corrected dispensing amount for the next dispensing cycle. This achieves precise control and adaptive optimization of micro-material dispensing, effectively improving ingredient accuracy and production stability. The specific beneficial effects are as follows:

[0050] 1. This invention first decrypts and verifies the integrity of encrypted formula data to ensure the security and integrity of formula data transmission and retrieval. At the same time, it comprehensively extracts the physical and chemical multidimensional physical properties of materials, constructs a complete material characteristic system, and relies on machine learning algorithms to fit the intrinsic relationship between physical properties and feed amount to achieve intelligent and accurate prediction of target feed amount. This avoids the basic feed deviation caused by the failure of traditional fixed formula feed methods to adapt to the differences in material properties, and improves the accuracy of formula feed benchmark from the source.

[0051] 2. This invention adaptively matches the corresponding specification of material container based on the predicted feeding amount, and obtains the precise coordinate position of the container by combining transfer positioning, so as to realize the precise matching and positioning docking of the batching station, container and material, and solve the problems of spillage and material deviation caused by the arbitrary adaptation of traditional batching containers and large positioning deviation. It provides a reliable workstation foundation for subsequent accurate weight monitoring and stable feeding operation.

[0052] 3. This invention relies on a high-precision weight monitoring module to collect the weight of the container in real time, dynamically calculate the amount of remaining material to be added, realize the quantitative monitoring of the entire feeding process, track the feeding progress status in real time, and accurately provide feedback on the dynamic feeding conditions, thus overcoming the shortcomings of traditional timed feeding which cannot detect the remaining material and has delayed feedback on operating conditions.

[0053] 4. This invention achieves adaptive switching between coarse feeding and fine feeding modes by comparing the remaining material quantity threshold, and generates segmented speed adjustment commands to meet the feeding requirements of different dosages of flavorings and fragrances. It balances the efficiency of batching operations with the accuracy of fine feeding, solves the problems of low efficiency for large dosages and serious deviations for small dosages in traditional single feeding speed, and is suitable for batching conditions across the entire range.

[0054] 5. This invention is based on a pre-constructed flow compensation mapping set, which combines real-time material properties and remaining feed quantity to dynamically match flow compensation values. At the same time, it combines the material volatilization loss characteristics to make dual corrections to the nozzle opening and closing time, effectively offsetting the feeding errors caused by the volatility of fragrances and flavors, unstable flow, and differences in physical properties, and greatly improving the fit and accuracy of discharge control under dynamic working conditions.

[0055] 6. This invention achieves precise feeding based on the corrected actual discharge time and collects single discharge deviation data in real time. By integrating multi-dimensional parameters through a gradient enhancement integrated prediction model, it completes deviation tracing and closed-loop correction, dynamically iteratively updates the feeding parameters for the next cycle, and realizes an iterative control mechanism for single deviation correction and continuous working condition optimization, effectively suppressing the cumulative error caused by long-term continuous feeding.

[0056] 7. This invention relies on continuously corrected, precise feeding quantities to achieve multi-station collaborative scheduling and continuous, high-precision batching control. It overcomes the shortcomings of traditional single-station independent control and poor consistency in batching ratios, realizing multi-batch station operation linkage, parameter coordination, and synchronous error correction. This invention is fully adaptable to the process characteristics of flavor and fragrance materials, such as high volatility, large differences in physical properties, and high requirements for micro-scale batching precision. It achieves intelligent management and control of the entire process, including formula safety verification, intelligent prediction, segmented precise feeding, dynamic compensation, and closed-loop collaborative control. This effectively improves the accuracy, consistency, and automation level of flavor and fragrance formulations, possessing strong process adaptability and practicality for production line engineering. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating an embodiment of the high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances of the present invention.

[0058] Figure 2 This is a functional block diagram of an embodiment of the high-precision intelligent formulation and ingredient coordination control system for fragrances and flavors of the present invention.

[0059] Explanation of icon numbers:

[0060] 10. Target feeding quantity prediction module; 20. Container positioning coordinate acquisition module; 30. Remaining material to be fed acquisition module; 40. Conveying speed segmented adjustment instruction generation module; 50. Real-time flow compensation value acquisition module; 60. Actual discharge time acquisition module; 70. Workstation collaborative scheduling and continuous batching control module. Detailed Implementation

[0061] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0062] like Figure 1 As shown, the first embodiment of this invention proposes a high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances. This method is applicable to automated and precise ingredient control for multi-component fine materials such as daily chemical flavorings, edible flavorings, and tobacco flavorings. It addresses industry pain points such as large differences in the physical properties of multiple materials, volatility, feeding lag, flow drift, feeding sequence conflicts at multiple workstations, and formula data leakage. This invention achieves safe, efficient, and high-precision automated and coordinated ingredient control for multi-component flavorings and fragrances through encrypted formula decryption verification, multi-dimensional material property modeling to predict target feeding amount, precise container positioning and weighing monitoring, adaptive switching between coarse and fine feeding modes, dynamic compensation for material flow, dual-dimensional tuning of nozzle discharge time, intelligent closed-loop correction of feeding deviation, and multi-workstation collaborative scheduling. The method includes the following steps:

[0063] Step S100: Obtain encrypted material formula data, perform decryption and integrity verification on the encrypted material formula data, extract the physical and chemical dimension features of each ingredient, and construct a multi-dimensional material property feature set; use a preset machine learning algorithm to fit the multi-dimensional material property feature set to accurately predict the target feed amount of each material.

[0064] The system reads encrypted material formula data from the formula encryption database or encrypted transmission terminal, and performs symmetric decryption using a local preset key. It then uses the MD5 hash verification algorithm to verify the integrity of the decrypted plaintext formula, comparing it with the original verification code to determine if the formula has been tampered with, is missing, or has been distorted during transmission. After successful verification, it extracts the physical and chemical parameters of each flavoring ingredient in batches from the compliant formula file, integrating and structurally constructing a multi-dimensional material property feature set covering all ingredients. Based on a pre-trained, convergent machine learning regression model, it performs nonlinear fitting calculations on the multi-dimensional property features, uncovering the correlation between material properties and actual feed loss, conveying adhesion, and volatilization loss. Based on the fitted mapping relationship, it predicts and compensates for losses under various operating conditions, ultimately outputting the target feed amount required for compliant production of each ingredient.

[0065] Encrypted material formula data consists of encrypted data such as the proportion of flavor components, raw material numbers, theoretical feeding benchmarks, and basic material properties, processed by a symmetric encryption algorithm. This data is used to prevent the core formula from being leaked or illegally tampered with.

[0066] Integrity verification is a security mechanism that uses hash digest comparison to verify whether the decrypted formula file is missing, tampered with, or contains garbled characters, ensuring that the formula used in production is authentic and valid.

[0067] Physical dimension characteristics are a set of parameters that characterize the macroscopic physical properties of materials, including material density, viscosity, surface tension, flowability, volatility coefficient, room temperature state, particle size and other physical property parameters.

[0068] Chemical dimension characteristics are a set of parameters that characterize the chemical properties of materials, including chemical parameters such as material polarity, solubility, acid and base values, component boiling points, oxidative stability, and cross-linking properties.

[0069] The material multidimensional property feature set is a standardized feature dataset formed by integrating the physical and chemical dimensions of all ingredients after normalization, providing input basis for intelligent prediction of feed amount.

[0070] The target feed rate is a precise baseline feed rate predicted by taking into account the theoretical formula dosage, material conveying adhesion loss, room temperature volatilization loss, and pipeline residue compensation.

[0071] Step S200: Match the material holding container of the corresponding specification based on the target feeding amount, drive the material transfer actuator to complete the transfer and positioning of the material holding container, and obtain the container positioning coordinates corresponding to the material holding container.

[0072] Based on the numerical range of the target feeding amount of each material, matching raw material weighing containers with suitable volume, range, and accuracy specifications is used to avoid the problem of overfeeding in small-range containers and insufficient weighing accuracy in large-range containers. Transfer control commands are issued to drive material transfer actuators such as conveyor belts and servo transfer robots to accurately transport empty containers to the weighing monitoring benchmark position of the corresponding batching station. The two-dimensional plane positioning coordinates of the container in the station coordinate system are collected in real time through station laser positioning sensors and vision positioning modules to lock the container position and eliminate weighing errors caused by position offset for subsequent high-precision weight acquisition.

[0073] Material containers are sealed weighing cups and storage tanks used to hold various flavoring raw materials and placed above the weighing module. They are classified into three types according to their capacity: micro-volume containers, regular containers, and large-capacity containers.

[0074] The material transfer actuator is an automated transfer hardware unit consisting of a servo motor, a transfer robot, a conveyor belt, and a limit cylinder, which realizes the automatic transport and limited placement of empty containers from the loading area to each batching station.

[0075] The container positioning coordinates are two-dimensional coordinate parameters of the center point of the weighing container, established in a plane coordinate system based on the origin of the batching station. They are used to determine whether the container is placed within the effective weighing area and to avoid off-center weighing deviations.

[0076] Step S300: Activate the weight monitoring module of the corresponding workstation according to the container positioning coordinates, collect the weight data of the materials already added in the material holding container in real time, and obtain the amount of remaining materials to be added under the current batching condition.

[0077] The container positioning coordinates are matched and verified with the effective weighing area coordinates of each station. If the coordinates fall within the effective range, the high-precision weight monitoring module of the current station is activated. After performing tare calibration on the empty container, the real-time cumulative weight data of the material fed into the container is continuously collected at high frequency. The target feeding amount predicted in step S100 is subtracted from the current cumulative material weight to dynamically calculate the material mass value that still needs to be fed under the current batching condition, i.e., the remaining material to be fed, which serves as the core basis for switching the feeding mode and adjusting the conveying speed.

[0078] The weight monitoring module is a real-time weight detection hardware unit consisting of a high-precision pressure weighing sensor, a signal filtering and acquisition unit, and an analog-to-digital conversion module. It supports tare zeroing and high-frequency time-series weight data output.

[0079] The weight data of the materials already added is a time-series data of the total mass of flavoring materials in the container collected from the start of feeding, reflecting the material feeding progress of the current workstation in real time.

[0080] The remaining amount of material to be added is the difference between the target amount of material to be added and the current cumulative weight already added. It represents the remaining mass of material that still needs to be accurately added at the current workstation and is dynamically reduced and updated as the material is added.

[0081] Step S400: Compare the remaining amount of material to be fed with the preset micro-material weight threshold, and adaptively switch between coarse feeding mode and micro-fine feeding mode based on the comparison result, and generate segmented adjustment instructions for the conveying speed corresponding to the working conditions.

[0082] The system pre-sets a threshold for determining the weight of trace materials based on the precision requirements of the flavoring ingredient process. It then performs a real-time logical comparison between the dynamically updated remaining material to be added and this preset threshold. When the remaining material exceeds the threshold, the system automatically switches to a coarse feeding mode, using high conveying speed to quickly add most of the material and ensure the batching production cycle time. When the remaining material is less than or equal to the preset threshold, the system automatically switches to a fine feeding mode, reducing the material conveying speed and minimizing the nozzle pre-opening degree to suppress overshoot. For each of the two different feeding operation modes, the system generates segmented conveying speed adjustment commands for the corresponding conveying rate, screw speed, and feed valve opening, and sends these commands to the material conveying execution unit.

[0083] The preset trace material weight threshold is a critical remaining weight value used to distinguish between coarse feeding and fine feeding operation modes. It is determined by the batching accuracy level and the minimum weighing division value of the material.

[0084] The coarse feeding mode is a high-speed batch feeding operation mode used to quickly consume most of the remaining materials and improve the overall production efficiency of batching.

[0085] The micro-feeding mode is a high-precision feeding operation mode with low conveying speed, small opening and slow discharge. It is used to accurately replenish materials when the target feeding value is approached to prevent overfeeding.

[0086] The segmented conveying speed adjustment command is a standardized speed control message issued for both coarse and fine feeding modes, which includes speed adjustment parameters such as screw conveyor speed, feed valve opening, and material pumping frequency.

[0087] Step S500: Based on the segmented adjustment command, retrieve the pre-constructed material flow compensation mapping set, and combine the current material property parameters with the remaining amount of material to be added to obtain the real-time flow compensation value.

[0088] Beforehand, through multiple batches of material flow rate calibration tests under different physical properties and conveying rates, a flow deviation database corresponding to multi-dimensional parameters such as material viscosity, volatility, conveying mode, and remaining feed amount is constructed and encapsulated as a material flow rate compensation mapping set. Based on the currently executed segmented conveying speed adjustment command, the current feeding mode is locked. Combining the multi-dimensional physical property parameters corresponding to the material with the real-time remaining feed amount, multi-dimensional index matching is performed in the mapping set, and table lookup fitting is used to obtain the real-time flow compensation value caused by pipeline adhesion, pressure fluctuation, and flow velocity drift under the current operating conditions. This value is used to correct the deviation between the theoretical discharge time and the actual discharge flow rate.

[0089] The material flow compensation mapping set is a structured database built on a large number of material property-flow velocity-flow deviation calibration samples, which stores the offset compensation parameters of the actual flow rate of the material relative to the theoretical standard flow rate under different operating conditions.

[0090] Real-time flow compensation values ​​are quantitative compensation values ​​used to correct instantaneous flow drift deviations in material conveying, compensating for output errors caused by changes in material viscosity, pipeline pressure, and conveying speed.

[0091] Step S600: The basic opening and closing control time of the dispensing nozzle is dynamically corrected and adjusted using real-time flow compensation values, and time compensation is performed in combination with the material volatilization loss characteristics to obtain the actual discharge time adapted to the real-time working conditions.

[0092] The standard basic opening and closing control time of the feeding nozzle is calculated based on the theoretical flow rate of the target single feeding. The real-time flow compensation value obtained in step S500 is used to perform the first dynamic correction on the basic opening and closing time to offset the discharge deviation caused by the drift of the conveying flow rate. Then, the room temperature volatility coefficient in the material properties is retrieved to calculate the equivalent discharge time deviation corresponding to the volatilization loss of the material during the nozzle discharge and pipeline temporary storage stages. The time after the first correction is adjusted for a second volatilization loss compensation. Finally, the actual discharge time that takes into account both flow drift and material volatilization interference factors is obtained, which serves as the control time parameter for the precise opening and closing of the nozzle solenoid valve.

[0093] The basic opening and closing control time of the dispensing nozzle is the theoretical opening and closing time required for dispensing the corresponding mass of material under ideal working conditions with standard rated flow and no loss.

[0094] The dynamic adjustment of duration is a calculation process that uses flow deviation to numerically compensate and correct the theoretical opening and closing time, thereby eliminating the output metering deviation caused by disturbances in the conveying conditions.

[0095] The volatility loss characteristic is the physical property law of flavor materials that are prone to volatility loss during the discharge process at normal temperature and pressure. The volatility loss coefficient is used to quantitatively characterize the proportion of mass loss in the material discharge stage.

[0096] The actual discharge time is the actual opening duration of the solenoid valve, which is ultimately executed by the dispensing nozzle after double correction for flow compensation and evaporation loss.

[0097] Step S700: Complete the single batching and feeding according to the actual feeding time, and collect the feeding deviation data of the current feeding cycle; input the feeding deviation data, physical and chemical characteristics of the material and the station operation parameters into the trained gradient improvement integrated prediction model to complete the closed-loop correction of the feeding amount and obtain the corrected feeding amount for the next feeding cycle; realize the collaborative scheduling of multiple batching stations and high-precision continuous batching control based on the corrected feeding amount.

[0098] The dispensing nozzle completes the single material dispensing operation according to the actual dispensing time after the setting. After the dispensing is completed, the final actual material dispensing mass is read from the weighing module of the workstation. The difference between the actual dispensing amount and the target dispensing amount is calculated to obtain the dispensing deviation data of this dispensing cycle. The dispensing deviation, the multi-dimensional physicochemical characteristics of the material, the conveyor speed of the workstation, the container positioning offset, the ambient temperature and humidity and other workstation operating parameters are combined into the model input feature vector. This vector is imported into a pre-trained converged dispensing deviation gradient improvement integrated prediction model to explore the causes of deviation and solve the deviation correction coefficient. A closed-loop reverse correction is performed on the target dispensing amount of this batch, and the corrected dispensing amount for the next continuous dispensing cycle is output. With the real-time corrected dispensing amount of each workstation, the dispensing completion sequence of the workstation, and the equipment load status as constraints, the start and stop sequence of dispensing at multiple dispensing workstations and the priority of conveying scheduling are uniformly planned to avoid problems such as pipeline pressure imbalance, air source conflict, and material blockage caused by dispensing at multiple workstations at the same time. This realizes multi-group dispensing and multi-workstation collaborative scheduling and continuously carries out high-precision closed-loop automatic dispensing control.

[0099] The discharge deviation data is the difference between the actual mass of ingredients fed in a single batch and the preset target amount of ingredients. It is used to characterize the magnitude and direction of the measurement error in this batching.

[0100] The operating parameters of the workstation include equipment and environmental operating parameters that affect the feeding accuracy, such as the conveyor speed, container positioning offset, ambient temperature and humidity, pipeline pressure, and solenoid valve response delay.

[0101] The gradient boosting integrated prediction model is a machine learning model trained on a massive amount of historical feeding deviation samples. It is used to identify the causes of feeding deviation, solve the feeding correction coefficient, and realize intelligent closed-loop correction of feeding amount.

[0102] The corrected feed rate is the benchmark feed rate for the next cycle after combining the results of the current feed deviation traceability, material properties and workstation conditions for correction and optimization. It is used to continuously iterate and improve the accuracy of batching.

[0103] Multi-feeding station collaborative scheduling is a scheduling and control mechanism that uses the adjustment of feeding quantity, equipment operating load and feeding completion sequence of each station as constraints to uniformly allocate feeding start and stop times and conveying scheduling priorities of each station, so as to realize orderly, balanced and conflict-free continuous feeding operation of multiple stations.

[0104] Furthermore, the high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors provided in this embodiment specifically includes step S100 as follows:

[0105] Step S110: The RSA asymmetric encryption and decryption algorithm is used to perform modular exponentiation decryption on the encrypted material formula ciphertext data and the local preset private key to obtain the decrypted plaintext formula dataset; the data integrity is verified by combining the SHA-256 hash digest operator with the unique timestamp and the original check code, and only the compliant formula data that passes the verification is retained, while the tampered and abnormal data is removed.

[0106] The system uses a locally pre-stored secure RSA private key to perform modular exponentiation on the transmitted ciphertext of the recipe to decrypt it and restore the original plaintext recipe dataset. It then calculates a hash digest based on the SHA-256 hash algorithm for the plaintext recipe and the current unique timestamp. The system compares the locally pre-stored original checksum with the real-time calculated digest. If the digests match, the recipe is deemed unaltered and transmission was undistorted, and the compliant recipe data is retained. If the digests do not match, the recipe is deemed to have been tampered with, missing, or garbled during transmission, and the abnormal recipe data is directly discarded and prohibited from use in ingredient production.

[0107] The RSA asymmetric encryption and decryption algorithm uses a public-key encryption and private-key decryption mechanism to ensure that recipe transmission is protected against theft; the SHA-256 hash digest operator is a one-way irreversible hash verification algorithm used for data integrity and anti-tampering verification; and a unique timestamp binds the time sequence of recipe collection and transmission to prevent replay attacks and bypassing of repeated tampering verification.

[0108] The formula for decrypting formula data is as follows:

[0109] (1)

[0110] In formula (1), It is a plaintext formula dataset, which is a decrypted plaintext structured material formula dataset. It is dimensionless and contains original business data such as the proportions, physicochemical parameters, and process feed limits of various flavor raw materials. It is a reliable data source for subsequent material feature extraction and target feed amount prediction. This is a binary encrypted file containing encrypted material formula data during transmission. It is used to prevent the leakage of core commercial formulas for flavorings. The data is encrypted using RSA public key encryption during the cloud distribution and local area network transmission stages. This is an explicit operator for RSA asymmetric encryption and decryption, using modular exponentiation at the underlying level. Decryption is achieved through completely open and reproducible operational rules, eliminating the presence of black-box encryption and decryption logic; among which... For encrypted numerical values, For private key index, The modulus of the key. The modulo operation (modulo operator) calculates the non-negative remainder when the dividend is divided by the modulus. The RSA private key vector, pre-configured for local security offline, is stored only in the local encrypted security partition of the batching control system. It is not transmitted over the network or uploaded to the cloud, thus avoiding the risk of key leakage from the transmission link. The control logic of formula (1) is that the batching control system receives the encrypted formula ciphertext sent from the cloud. Read the pre-installed private key in the local secure partition. The RSA decryption operator is invoked to decrypt and restore the ciphertext based on explicit modular exponentiation, outputting a structured plaintext recipe dataset. The decrypted plaintext data is directly sent to the subsequent SHA-256 hash integrity verification stage. Decrypted data that fails the verification is discarded and does not enter the subsequent feeding prediction process. This formula (1) relies on the RSA asymmetric decryption algorithm and performs explicit decryption operation on the encrypted fragrance formula ciphertext through a local offline pre-set private key to restore the structured plaintext formula dataset. The private key does not participate in network transmission throughout the process, effectively eliminating the security risks of key leakage and data theft in the transmission link of high-value fragrance commercial formulas, and providing a safe and reliable original data source for subsequent formula integrity verification and material property feature extraction.

[0111] The formula for determining recipe integrity hash verification is:

[0112] (2)

[0113] In formula (2), This is a binary identifier for the legality and validity of the formula data, with a value of only 1 or 0. A value of 1 indicates that the decrypted fragrance formula plaintext has not been tampered with or subjected to a network replay attack, and the data is legal and reliable. A value of 0 indicates that the data is abnormal, and the formula data is directly removed and prohibited from entering the subsequent feeding prediction and ingredient control process. It is an explicit hash digest operator for SHA-256, which is a publicly reproducible one-way hash operation. It can map structured plaintext recipe data of arbitrary length to a hash digest value of fixed length of 256 bits. As long as any byte of the original recipe is modified, the final hash digest result will inevitably change. It is a bitwise XOR operator used to perform obfuscation operations on the recipe hash digest and the unique timestamp bound to the encryption stage, thus resisting replay-type network attacks. It is a globally unique timestamp bound to the cloud when the formula is encrypted and distributed. It is dimensionless, and only one timestamp is assigned to the same batch of formulas, which can accurately distinguish the duplicate distribution of the same formula data packets. The original verification code, generated synchronously in the cloud encryption process and distributed along with the ciphertext, is pre-generated by the encryption end through the same hash and XOR operation, serving as a trusted comparison benchmark for local verification. The control logic of formula (2) is based on the plaintext recipe dataset obtained after RSA decryption. Perform a SHA-256 hash operation to obtain a unique hash digest of the recipe; perform a bitwise XOR obfuscation operation on the hash digest and the unique timestamp bound to the transmission to generate a local verification feature; compare the locally calculated local verification feature with the preset original checksum. Perform an equality comparison: If the two are equal, then mark the validity and legality of the formula data with a binary value. Set to 1 to allow the formula; if not equal, set to 0 to discard the abnormal formula and trigger a security alarm. Formula (2) extracts the fixed digest value of the decrypted formula through the SHA-256 hash operator, combines it with the unique timestamp bitwise XOR obfuscation and compares it with the original preset check code, and determines whether the formula has been tampered with or replayed abnormally through the binary valid identifier; relying on the hash anti-tampering + timestamp anti-replay dual verification mechanism, it accurately filters illegal formula data, ensures the data source security and reliability of the flavor intelligent ingredient system, and avoids the production safety and commercial loss risks caused by malicious tampering of formula.

[0114] Step S120: Extract four core physicochemical parameters—material density, dynamic viscosity, volatility coefficient, and solubility—from the compliant formulation data to construct a structured multidimensional material property feature vector.

[0115] From the verified compliant formula documents, the four core physicochemical indicators corresponding to each flavor ingredient are analyzed in batches: material density, dynamic viscosity, room temperature volatility coefficient, and solvent solubility. The extreme values ​​of the four types of parameters are uniformly normalized to eliminate dimensional differences, and they are arranged in a structured order according to a fixed dimension to generate a material multidimensional physical property feature vector with unified dimensions that can be directly input into machine learning models.

[0116] The structured material multidimensional physical property feature vector is a one-dimensional standardized feature vector formed by the normalization and orderly arrangement of four types of key physicochemical parameters, which serves as the standard input for the regression prediction model.

[0117] The formula for constructing the multidimensional eigenvector of a material is:

[0118] (3)

[0119] In formula (3), The first is a multidimensional physical property feature vector of the material, and the second is a structured one-dimensional feature vector output in this step. These serve as standardized input data for subsequent machine learning feed prediction models, unifying the feature dimension format of various flavor raw materials. The measured density of the fragrance material is given in units of... The density was obtained from on-site testing with a density meter, and the reasonable range of values ​​for the process is as follows: This directly affects the material conveying velocity and the instantaneous discharge rate of the nozzle. The dynamic viscosity of the material, in units Data were acquired using a rotational viscometer, and the range of values ​​was limited to [value missing]. Viscosity differences cause variations in pipeline resistance, which is the core influencing factor on flow deviation when switching between coarse and fine feeding modes. Volatility coefficient of the material, unit The value range was obtained through a calibration experiment on volatilization loss under constant temperature conditions. It is used to characterize the degree of volatilization loss of fragrance during feeding, settling, and transportation, and to provide a physical property basis for compensation of discharge time. The solubility of the material is dimensionless, and the normalization constraint value is within a certain range. The range between these values ​​characterizes the solubility stability of the flavoring raw material in the formulation's mixed solvent system, and is used to predict the risks of transport blockage and feeding errors caused by precipitation and agglomeration during the ingredient preparation process. For vector transpose operator. This formula (3) constructs a structured multidimensional material property feature vector by splicing together four core physicochemical parameters of flavor material, namely measured density, dynamic viscosity, volatility coefficient and solubility, according to a fixed dimension, and sets reasonable process value ranges for each parameter to complete abnormal data filtering; through multi-property joint feature modeling, it comprehensively characterizes the fluid transport characteristics and volatilization loss characteristics of different flavor raw materials, solves the defect of poor generalization ability of single property fitting, and provides standardized and mechanism-oriented model input features for accurate prediction of target feed amount in machine learning.

[0120] Step S130: Use a random forest regression model containing decision trees to perform integrated fitting and prediction on the multidimensional physical property feature vector of the material, take the mean of all decision tree outputs and constrain the results to be non-negative, and obtain the target feed amount corresponding to each ingredient.

[0121] The normalized multidimensional physical property feature vector is input into the pre-trained convergent random forest regression model. Multiple independent decision trees within the model perform nonlinear fitting prediction of physical properties and feeding loss in parallel. The arithmetic mean of the predicted feeding results output by all decision trees is calculated to suppress the overfitting bias of a single decision tree. The maximum value constraint operator is used to limit the prediction results to non-negative values. Finally, the target feeding amount of each ingredient after loss compensation is output.

[0122] The random forest regression model, a machine learning regression model based on multi-decision tree ensemble, is suitable for nonlinear prediction of feed quantity under small batches of multi-property samples; the ensemble fitting prediction reduces prediction variance and improves the robustness of feed quantity prediction by fusing the results of multiple decision trees.

[0123] In this embodiment, random forest regression is selected as the preferred machine learning algorithm. Under the same conditions, it can be replaced by XGBoost regression, LightGBM regression, SVR support vector regression, and MLP multilayer perceptron neural network. All of these algorithms can complete the integrated fitting prediction of the target feed amount based on the multidimensional physical properties of the material, and are equivalent technical means that can be directly replaced by this invention.

[0124] The formula for predicting the target feed amount is:

[0125] (4)

[0126] In formula (4), The target feed amount for flavoring materials is expressed in mg, and the reasonable range of values ​​for the process is as follows: The predicted output of this step serves as the benchmark setting for subsequent batching station container selection and coarse and fine feeding control. The value is always non-negative to avoid invalid negative feeding commands. The total number of independent decision trees in the random forest ensemble model is dimensionless. The optimal calibration is determined through offline grid search based on a multi-batch physical property-feeding history dataset of flavorings. This is used to balance the accuracy of model prediction with the time required for online inference computation. For the first The decision tree uses an explicit piecewise regression mapping function, which is trained under supervision based on historical material multidimensional property characteristics and corresponding actual feed samples. Internal splitting thresholds, branching rules, and leaf node prediction values ​​are all fixed and queryable, making it a completely reproducible explicit operation without any black-box prediction logic. The input is a pre-constructed material multidimensional property feature vector. Output the predicted value of a single feeding operation for each decision tree. To integrate the mean operator, the arithmetic mean of the prediction results of all 120 independent decision trees is calculated. This multi-model voting average method suppresses the overfitting risk of a single decision tree and improves the model's generalization prediction ability for flavor materials with different physical properties. The control logic of formula (4) is to integrate the multi-dimensional physical property feature vectors of the material. The 120 independent decision trees, after being trained and solidified, are input sequentially. Each decision tree performs piecewise regression calculations according to the branching rules obtained from offline training, and outputs its own material feeding prediction results. The prediction results of all decision trees are then ensembled using an arithmetic mean to obtain the initial predicted material feeding amount. Negative abnormal prediction results are filtered out through non-negativity constraints, and value range verification is performed in conjunction with the process material feeding range. Finally, the compliant target material feeding amount is output. Formula (4) is based on the random forest ensemble learning architecture. It uses 120 offline-calibrated explicit decision trees to perform regression prediction on the multidimensional physical properties of the material for feeding amount. The arithmetic mean of the prediction results of all single trees is calculated and the output is constrained to be non-negative to obtain the target feeding amount of the flavoring material. It relies on ensemble learning to suppress the overfitting of single models and accurately fits the nonlinear conveying feeding law of flavoring materials with different viscosities and volatiles, so as to realize the high-precision feeding amount prediction of multi-category flavorings. At the same time, the model inference rules are completely reproducible, ensuring the traceability and compliance management of the ingredient process.

[0127] Preferably, in the high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances provided in this embodiment, the material holding container is a material cup, the material transfer actuator is a transfer robot, and step S200 specifically includes:

[0128] Step S210: Calculate the theoretical filling volume of the material by combining the target feeding amount with the material density, set the safety margin coefficient, select the material cups whose rated volume meets the constraints as matching containers, and establish a one-to-one mapping relationship between feeding amount, material cup number, and material batching station coordinates.

[0129] The target feed mass is converted into the theoretical filling volume of the material using the material density formula, and multiplied by a preset safety margin coefficient to prevent material overflow during filling; the rated volume parameters of each specification of material cup are traversed through the system's pre-stored parameters, and material cups with a rated volume greater than the theoretical filling volume with safety margin are selected as the optimal matching containers; the target feed quantity, the unique number of the material cup, and the coordinates of the corresponding batching station to be assigned for each group are bound and stored, and a structured mapping table with a one-to-one correspondence between the three is constructed.

[0130] Theoretical filling volume is the theoretical volume occupied by the target feed mass at the corresponding material density; safety margin factor is the volume amplification factor used to reserve filling redundancy and prevent material overflow.

[0131] The constraint formula for matching the capacity specification of the material cup is:

[0132] (5)

[0133] In formula (5), For the first The rated nominal volume of the feed cup, in units These are fixed parameters calibrated at the factory for the material cup, used to limit the maximum liquid volume that a single material cup can hold. The safety margin coefficient for filling is dimensionless and is fixed at 1.2 after batch testing under flavor filling conditions. It reserves 20% of the volume as a redundancy to offset the volume increase caused by liquid surface tension, internal air bubbles, and instantaneous flow fluctuations during the filling process of liquid flavor, and to prevent material overflow. The target material input quantity predicted in the previous steps, in units of Divide by 1000 to complete the unit conversion from milligrams to grams, and combine with the measured density of the flavoring material. It can accurately calculate the theoretical filling volume of flavoring materials. The measured density of the fragrance material is given in units of... The multidimensional physical property parameters of the material collected in step S120 are used to realize the accurate conversion from the target feeding mass to the filling volume. The control logic of formula (5) is to use the target feeding amount and material density to convert the mass of the flavoring to be added into the theoretical filling liquid volume; multiply by the filling safety margin coefficient of 1.2 to obtain the minimum required safety volume with overflow protection; traverse all the rated volume parameters of the cups stored in the system, and only select the cups with rated volume greater than or equal to the safety volume as the matching containers; bind a unique number and the corresponding feeding station positioning coordinates to the successfully matched cups, construct a one-to-one mapping data table of feeding amount - cup number - station coordinates, and send it to the transfer robot to perform positioning and transfer. Formula (5) converts the target feeding mass into the theoretical filling volume by combining the material density. It introduces a filling safety margin coefficient of 1.2 to set volume redundancy constraints and automatically selects material cups with a rated volume not less than the safe filling volume. By reserving a safe volume, it effectively prevents material overflow caused by bubbles and liquid surface tension during liquid flavoring filling, realizes intelligent and safe matching of material cup specifications under different feeding ranges, and provides a reliable container selection basis for the manual position scheduling of transfer machinery.

[0134] Step S220: Drive the transfer robot to complete the grasping, shifting and placement of the cup. Collect the pixel coordinates of the cup image through an industrial camera. Use the pre-calibrated camera intrinsic parameter matrix and pixel physical scaling factor to convert the pixel coordinates into the three-dimensional world coordinates of the production line to obtain the precise positioning coordinates of the container.

[0135] The servo control command is issued to drive the transfer robot gripper to grab an empty cup and move it along a preset trajectory to the target station, where it is then smoothly placed and limited. The industrial camera at the station is triggered to capture an image of the cup area, and the pixel coordinates of the cup's center point are extracted through image feature recognition. The pre-calibrated camera intrinsic parameter matrix, pixel scaling factor, and actual physical size scaling factor are called, and the two-dimensional pixel coordinates are mapped to the global three-dimensional world coordinates of the production line through camera perspective transformation calculation, thus obtaining the precise container positioning coordinates of the cup.

[0136] The camera intrinsic parameter matrix is ​​a matrix of inherent parameters such as focal length and principal point obtained from camera calibration, used for the transformation of pixel coordinates to world coordinates; the pixel physical scaling factor is a calibration parameter that represents the actual physical length of the pipeline corresponding to a single pixel.

[0137] The coordinate transformation formula is:

[0138] (6)

[0139] In formula (6), , , These are the three-dimensional physical positioning coordinates of the material cup in the coordinate system of the material dispensing line, in mm. They are the final output of the precise position parameters of the workstation in this step, used for the transfer robot to perform fixed-point gripping and precise placement control. , The horizontal and vertical pixel coordinates of the cup target features acquired by the industrial camera in the image represent the original data acquired at the image level; the third dimension, supplemented by a constant 1, constitutes a homogeneous coordinate format, satisfying the dimension requirements of matrix inverse operations. The inverse matrix of the pre-calibrated camera intrinsic parameter matrix, The calibration is obtained offline using the checkerboard calibration method and then stored. Inverse matrix operations are used to perform reverse correction on the radial and tangential distortions of the lens, eliminating pixel positioning errors introduced by optical distortion. , is the pixel-to-physical distance scaling factor, in mm / pixel, obtained through actual measurement and calibration using a standard calibration block. It is used to convert the corrected normalized pixel coordinates into the actual physical length of the production line, achieving accurate mapping of image coordinates to the three-dimensional world coordinates of the industrial site. The control logic of formula (6) is that the industrial camera acquires real-time images of the material cup and extracts the pixel coordinates corresponding to the center position of the material cup through a feature detection algorithm. The process involves constructing a 3D homogeneous pixel coordinate vector; calling the offline-fixed camera intrinsic parameter inverse matrix to perform inverse distortion correction on the pixel homogeneous coordinates to eliminate positioning offset caused by lens distortion; and multiplying by a pixel-physical scaling factor to convert the normalized pixel coordinates into pipeline 3D physical coordinates. The three-dimensional positioning coordinates are sent to the motion controller of the transfer robot to realize the precise grasping of the cup and the fixed-point placement of the workstation. This formula (6) performs distortion inverse correction on the homogeneous coordinates of the image pixels based on the pre-offline calibrated camera intrinsic parameter inverse matrix, and combines the pixel-physical distance scaling coefficient to accurately convert the two-dimensional image pixel coordinates of the cup into the three-dimensional physical workstation coordinates of the production line; the visual positioning deviation is eliminated by lens distortion correction, realizing high-precision visual positioning of the transfer robot for grasping and placing, and providing a precise position control basis for multi-workstation automated collaborative material preparation.

[0140] Step S230: Calculate the Euclidean distance between the current material cup position and the reference coordinates of each idle preset workstation as the workstation allocation cost, select the workstation number corresponding to the minimum cost value as the target allocation workstation, and complete the dynamic allocation of multiple workstations.

[0141] Traverse all currently idle batching stations and read the preset reference positioning coordinates of each station; calculate the Euclidean space distance between the three-dimensional positioning coordinates of the material cup and the reference coordinates of each idle station, and use the distance value as the station allocation cost function value; select the idle station number with the minimum allocation cost as the final target allocation station for the material cup, so as to achieve optimal dynamic allocation of material cups among multiple stations and reduce the weighing error caused by station off-center loading.

[0142] Euclidean distance, the straight-line distance between two points in multidimensional space, serves as a quantitative cost indicator for optimal workstation allocation; workstation allocation cost characterizes the magnitude of the positional deviation when the material cup is transferred to the corresponding workstation, and the smaller the value, the higher the workstation adaptability.

[0143] The optimal workstation allocation formula is:

[0144] (7)

[0145] In formula (7), For the first The feeding cups are distributed to the first... The space allocation cost of each idle batching station is quantified by using a two-dimensional Euclidean distance to measure the spatial interval between the current position of the material cup and the preset position of the station. The smaller the distance value, the shorter the transfer path of the robot arm. The coordinates of the current material cup in two-dimensional plane are obtained by visual coordinate transformation. They are taken from the station positioning parameters output by formula (6) and the unit is mm. For the first The offline preset reference coordinate vector of each batching station is a fixed station position parameter that is permanently stored in the system. Only idle stations participate in the cost traversal calculation. It is an L2 norm operator used to solve the Euclidean straight-line distance between two points, accurately representing the travel length of the robot arm from the current position of the material cup to the target station. The optimal workstation number is determined by minimizing the spatial allocation cost, and the target workstation command for the transfer robot is issued. The control logic of formula (7) is to traverse all idle batching workstations in the system and sequentially retrieve the preset reference coordinates of each workstation. Based on the real-time two-dimensional positioning coordinates of the cup, the Euclidean space distance between the cup and each idle workstation is calculated using L2 norm operation to obtain the allocation cost corresponding to each group of workstations. The minimum value traversal solution strategy is adopted to select the workstation number corresponding to the minimum allocation cost as the optimal target workstation. The optimal workstation number is bound to the current cup's feeding task, and the transfer robot is sent to perform the nearest transfer and placement. This formula (7) uses the two-dimensional Euclidean distance between the real-time positioning coordinates of the cup and the preset reference coordinates of each idle workstation as the workstation allocation cost, and traverses to solve the workstation number corresponding to the minimum cost as the optimal allocation workstation. By using the nearest dynamic workstation allocation, the robot's transfer stroke is shortened, the equipment waiting time is reduced, and the parallel collaborative scheduling of multiple feeding workstations is realized, which greatly improves the overall production and operation efficiency of batch intelligent feeding of flavorings.

[0146] Furthermore, the high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors provided in this embodiment includes step S300 as follows:

[0147] Step S310: Activate the high-precision weighing sensor at the corresponding workstation based on the container positioning coordinates, and perform sliding mean filtering and noise reduction processing on the original weighing sampling sequence using a fixed sliding time window to eliminate weighing pulse noise caused by mechanical vibration and electromagnetic interference, and obtain the real-time net weight of the material after filtering.

[0148] After verifying that the container's positioning coordinates fall within the effective weighing area of ​​the workstation, the weighing sensor is tare-cleared and initialized. Raw weighing time series data is continuously collected at a fixed sampling frequency. A sliding time series window with a preset fixed duration is selected, and the arithmetic mean of multiple sets of raw weighing sample values ​​within the window is calculated. Instantaneous pulse abnormal sampling points caused by equipment mechanical vibration and industrial electromagnetic interference are filtered out, and stable, noise-free filtered real-time material net weight data is output.

[0149] The original weighing sampling sequence is a continuous weight time series data without noise reduction, collected at high frequency by the sensor; the sliding time series window is a fixed-length continuous sampling interval used for mean noise reduction.

[0150] The formula for sliding window mean filtering noise reduction is:

[0151] (8)

[0152] In formula (8), For the first The real-time net weight of the material at each sampling time, after being processed by the moving average filter, is in mg. It is a stable and reliable weighing data after noise suppression, and is used to calculate the amount of material to be added at the current workstation in real time. For the first The raw weighing sample value directly collected by the weighing sensor at all times is susceptible to high-frequency pulse noise caused by mechanical vibration of the robotic arm, electromagnetic interference on site, and airflow disturbance, resulting in instantaneous value jumps. The total length of sampling points within the sliding time window, after being fixed and calibrated under the condition of rapid addition of trace amounts of flavoring, is [value missing]. The window extracts 15 consecutive sets of raw sampled data from the current time and performs an average calculation, effectively filtering out high-frequency noise while keeping the time lag to a minimum. The sampling time index within the window is defined as the 15 consecutive sampling time points backward from the current time, ensuring that each filtering operation is calculated based on the latest time-series sample of equal length. The control logic of formula (8) is as follows: the weighing sensor continuously collects the original weight sequence of the material in the cup at a fixed sampling frequency, and caches the latest 15 sets of original sampling data in real time; with the current time as the end point of the window, a continuous time-series sampling segment of length 15 is extracted, and the arithmetic mean of all the original weighing values ​​within the window is calculated; the mean result is used as the net weight of the material after filtering at the current time, which is used to update the remaining amount of material to be added in real time; the window slides backward point by point with the sampling time, and the mean filtering operation is performed cyclically. Formula (8) uses a time-series sliding window with a fixed length of 15 to calculate the arithmetic mean of the original sampled weight sequence of the weighing sensor at the workstation window by window, thereby suppressing high-frequency pulse noise such as electromagnetic and mechanical vibration in the industrial field. The short-time sliding window design effectively smooths the weight jump data while avoiding large time lag, adapting to the real-time accurate monitoring of material weight under the condition of rapid feeding of trace amounts of flavoring, and providing a reliable weight data source for adaptive switching of feeding mode.

[0153] Step S320: Subtract the filtered real-time net weight of the material from the target feed amount, and ensure that the value is non-negative by using the maximum value constraint to calculate the remaining amount of material to be fed under the current working conditions.

[0154] The preset target feeding quantity of the material is read and the difference is calculated with the current real-time net weight of the filtered material. In order to avoid the difference being negative due to small fluctuations in the weighing during the feeding process, the difference is limited by a non-negative maximum value constraint operator. When the difference is less than 0, it is forcibly set to 0, and the original difference is retained at other times. Finally, the compliant and valid remaining material quantity to be fed is obtained.

[0155] Non-negative maximum value constraint, a numerical limit calculation rule to prevent the remaining feed amount from becoming negative.

[0156] The formula for calculating the remaining amount to be invested is:

[0157] (9)

[0158] In formula (9), This represents the remaining amount of material to be fed at the current workstation, in mg. Its value is always greater than or equal to 0. It is the core judgment parameter for adaptive switching between coarse feeding and micro-fine feeding modes, and is used to drive the segmented adjustment of the nozzle conveying speed in real time. The total target feed amount of materials predicted in step S130 is expressed in mg, and the theoretical total feed mass of the ingredients for this batch is given. The current real-time net weight of the material is the result of noise reduction processing using the moving average filter of formula (8), representing the total mass of the flavoring material that has been added to the cup. As a maximum value constraint operator, the lower limit of the remaining amount to be fed is forcibly limited to 0, avoiding the abnormal situation of negative value of the remaining amount due to overfeeding or small weighing error. The control logic of formula (9) is to subtract the net weight of the filtered material already fed at the current station from the preset target feeding amount to obtain the theoretical remaining feeding mass difference; the calculation result is compared with 0 through the maximum value function. When the difference is greater than 0, the difference is used as the remaining amount to be fed; if overfeeding or weighing deviation causes the difference to be negative, the remaining amount to be fed is forcibly set to 0; the remaining amount to be fed after non-negative constraint is sent to the feeding speed control unit in real time as the basis for judgment of coarse and fine feeding mode switching. This formula (9) uses the difference calculation between the target feeding amount and the current filtered net weight, combined with the maximum value constraint to forcibly limit the remaining amount of material to be fed to a non-negative number, which can effectively avoid the negative value abnormality caused by overfeeding and weighing error, and provide a stable and reliable judgment parameter for the adaptive switching of coarse feeding and fine feeding modes, ensuring the continuous and stable operation of the segmented feeding control logic.

[0159] Preferably, the high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors provided in this embodiment includes step S400 as follows:

[0160] Step S410: Preset the weight threshold of trace materials, the conveying speed of the coarse feeding gear, and the conveying speed of the fine feeding gear.

[0161] Based on the flavor and fragrance category, ingredient accuracy level, weighing sensor graduation value, and nozzle discharge characteristics, three types of process control parameters are pre-configured locally in the ingredient control system: a critical threshold for determining the weight of trace materials to distinguish between coarse and fine feeding conditions, a high-speed rated conveying speed of the screw conveyor in coarse feeding mode, and a low-speed rated conveying speed in fine feeding mode. These three types of parameters can be configured and stored according to different flavor formulations and different material viscosities, serving as benchmark constraint parameters for switching feeding modes and adjusting conveying speed.

[0162] The trace material weight threshold is the critical mass value of the remaining material to be fed in the two operating conditions of coarse feeding and trace fine feeding. It is the judgment criterion for segmented speed control. The coarse feeding gear conveying speed is the rated high speed of the material conveying screw in the coarse feeding mode, which is used for the rapid conveying and feeding of large quantities of materials. The fine feeding gear conveying speed is the rated low speed of the material conveying screw in the trace fine feeding mode, which is used for the precise replenishment of small doses of materials.

[0163] Step S420: When the remaining amount of material to be fed is greater than the preset micro-material weight threshold, execute the coarse feeding mode and match the high-speed conveyor gear to complete the rapid feeding.

[0164] The remaining amount of material to be added is dynamically updated in real time and compared with the preset micro-material weight threshold in step S410. When the remaining material to be added is greater than the preset micro-material weight threshold, it is determined that the current stage is still in the large-scale material feeding stage. The control system automatically locks the coarse feeding operation mode, calls the preset coarse feeding gear high-speed conveying parameters, and issues screw speed and feed valve opening control commands to drive the conveying mechanism to continuously convey materials at high speed, quickly consume most of the material to be added, and shorten the production time of a single batch of batching.

[0165] The coarse feeding mode is a batch feeding operation mode with high-speed conveying as the core. It is suitable for the early feeding stage when there is a lot of remaining material, and prioritizes the efficiency of batching and production.

[0166] Step S430: When the remaining amount of material to be fed is less than or equal to the preset micro-material weight threshold, switch to micro-precision feeding mode, match the low-speed conveying gear to complete precise feeding, and realize coarse and fine dual-stage adaptive segmented speed regulation control.

[0167] If the real-time remaining amount of material to be fed is less than or equal to the preset micro-material weight threshold, it is determined that the current stage of micro-precision feeding has been entered. The system immediately cuts off the high-speed conveying command for coarse feeding and automatically switches to the micro-precision feeding operation mode. The preset low-speed conveying gear parameters are called to control the screw to run at a low speed, and the nozzle is opened at a small degree to slowly discharge the material. The overshoot error caused by the discharge inertia is reduced by relying on low-speed conveying. The dynamic monitoring of the remaining material amount realizes the seamless adaptive switching between the coarse and fine feeding modes, forming a two-stage segmented speed regulation closed-loop control, which simultaneously realizes high production efficiency and high precision of micro-metering.

[0168] The micro-feeding mode is a low-speed, small-aperture fine feeding operation mode, used for micro-feeding at the end of the feeding process to suppress feeding errors;

[0169] The coarse and fine dual-segment adaptive segmented speed regulation control is a closed-loop speed regulation strategy that automatically switches between high and low speed conveying gears based on the remaining amount to be fed, taking into account both production cycle and batching accuracy.

[0170] The formula for adaptive segmented speed control is:

[0171] (10)

[0172] In formula (10), The real-time conveying line speed of the ingredient conveying pipeline is measured in mm / s. The speed control command of the nozzle feeding actuator directly determines the instantaneous discharge rate of the flavoring. This is the coarse feeding speed setting, a high-speed conveying setting in coarse feeding mode. It is calibrated offline to 35mm / s under batch batching conditions and is used for rapid feeding of large quantities of materials to improve overall batching production efficiency. The conveying speed for the fine feeding gear is the low-speed conveying gear in the micro-fine feeding mode, which is fixed at 6mm / s. When the remaining amount to be fed is close to the target value, the low-speed feeding is used, which can significantly reduce the discharge overshoot error caused by fluid inertia and ensure the accuracy of micro-feeding. For the micro-weight threshold of coarse and fine feed switching, in units After testing and calibration of various viscosity flavoring materials, it was determined to be... This serves as the critical value for determining the two feeding modes. , This is an enable flag for coarse feeding and fine feeding binary modes, taking only 0 or 1 values; when the flag is 1, it indicates that the corresponding feeding mode is activated, which is used to interlock the downstream flow compensation and discharge time correction modules to synchronize and match the control strategy. The non-negative remaining material quantity to be added, calculated by the preceding formula (9), serves as the real-time input parameter for mode switching. The control logic of formula (10) reads the current remaining material quantity to be added in real time. , and the preset trace weight threshold Make a size comparison; if the remaining amount to be fed is greater than 120mg, assign the conveying speed to the high speed level of 35mm / s, set the coarse feeding mode identifier to 1, and execute large-batch rapid feeding; if the remaining amount to be fed is less than or equal to 120mg, automatically switch to low speed of 6mm / s, set the fine feeding mode identifier to 1, suppress fluid inertia overshoot with low flow rate, and carry out micro-precision feeding. This formula (10) uses the calibrated micro-material weight threshold of 120mg as the switching critical condition, and adaptively switches the conveying speed between the two levels of 35mm / s high-speed coarse feeding and 6mm / s low-speed fine feeding according to the real-time remaining amount of material to be fed, and outputs the feeding mode interlock identifier simultaneously; by segmented speed regulation to balance the batch batching production efficiency and micro-feeding control accuracy, it effectively suppresses the discharge overshoot defect caused by fluid inertia, and realizes the dual-objective batching control of high efficiency and high precision of flavoring.

[0173] Furthermore, in the high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances provided in this embodiment, the material flow compensation mapping set construction method in step S500 includes:

[0174] Actual discharge deviation data of different viscosities and volatile flavors under multiple conveying speeds and multiple sets of remaining feed amounts are collected in advance. Discrete lookup table mapping function is constructed using material dynamic viscosity, real-time conveying speed and remaining feed amount as three-dimensional indexes. During online operation, the three sets of parameters of current working condition material dynamic viscosity, real-time conveying speed and remaining feed amount are input into the discrete lookup table mapping function to match and obtain real-time flow compensation value.

[0175] Multiple calibration tests were conducted beforehand to collect actual discharge deviation data between the actual discharge and theoretical flow rates of flavoring materials with different dynamic viscosities and volatility under various conveying speeds and gradients of remaining material to be added. Using the material's dynamic viscosity, the conveying speed corresponding to the real-time conveying speed, and the current remaining material to be added as a three-dimensional index, a discrete lookup table mapping function was constructed and encapsulated as a material flow compensation mapping set. During the online batching operation, the real-time conveying speed was locked according to the current segmented speed adjustment command. Combined with the material viscosity property parameters and the real-time remaining material to be added three-dimensional parameter lookup table index, the real-time flow compensation value used to correct pipeline flow drift under the current operating conditions was obtained.

[0176] The material flow compensation mapping set is a three-dimensional discrete lookup table knowledge base built based on multi-condition calibration samples, which stores the flow deviation compensation values ​​corresponding to different viscosities, conveying speeds, and remaining feed amounts.

[0177] The three-dimensional index discrete lookup table mapping function uses viscosity, conveying speed, and remaining amount to be added as search keys to quickly match the optimal flow compensation parameters.

[0178] Real-time flow compensation values ​​are quantitative correction parameters used to compensate for deviations in discharge flow caused by pipeline pressure fluctuations, material viscosity drift, and screw conveyor slippage.

[0179] The formula for calculating real-time traffic compensation is:

[0180] (11)

[0181] In formula (11), This is a real-time flow compensation value in mg, and the value can be positive or negative. When it is positive, it means that a small amount of additional material needs to be added to compensate for the insufficient output caused by pipeline adhesion and nozzle lag. When it is negative, it means that the nozzle needs to be shut off in advance to counteract the excessive output caused by fluid inertia and correct the inherent output deviation of the conveying system from the perspective of material quantity. The pre-constructed three-dimensional discrete lookup table mapping function is based on a dataset built and stored offline from a large number of flavor batch operating condition tests. It belongs to a fully traceable explicit index lookup table logic and does not involve black-box operations; the dynamic viscosity of flavor materials is used respectively. Real-time conveying speed Remaining amount of materials to be added As a three-dimensional search index, each index is uniquely bound to a set of traffic compensation values ​​that have been measured and calibrated. The dynamic viscosity of the flavoring material is taken from the multidimensional physical property parameters of the material collected in step S120. It determines the fluid flow resistance and inertia, and is the core physical property factor affecting the flow deviation. The real-time conveying line speed of the current segmented speed regulation output directly determines the instantaneous flow rate of the fluid and the inertial delay of nozzle opening and closing. The remaining amount of material to be fed determines the margin range from the target feeding value. Different margins result in different pipeline pressures and fluid flow states, corresponding to different discharge deviation characteristics. The control logic of formula (11) is to retrieve three sets of working condition parameters in real time during the online operation phase: the dynamic viscosity of the current flavoring material, the real-time conveying speed under coarse and fine feeding modes, and the current remaining amount of material to be fed as a three-dimensional retrieval index; perform precise index matching in the offline pre-constructed flow compensation mapping data table, and directly look up the corresponding flow compensation value under this working condition; and add the flow compensation amount to the remaining amount to be fed as the benchmark input for nozzle opening and closing time correction, so as to perform pre-compensation control for discharge deviation in advance. This formula (11) uses the material dynamic viscosity, real-time conveying speed, and remaining amount of material to be added as a three-dimensional retrieval index. It directly matches the flow compensation value under the current working condition through the discrete lookup table mapping function constructed by the pre-offline test. It achieves millisecond-level pre-compensation for discharge deviation based on the offline working condition knowledge base, avoids the real-time calculation overhead of complex fluids, and can accurately correct the flavoring feeding deviation caused by pipeline adhesion and nozzle inertial delay, adapting to high-precision batching scenarios of multi-property materials.

[0182] This embodiment uses a three-dimensional index of viscosity, conveying speed, and remaining material to be added as the basic lookup table architecture. In actual engineering applications, an ambient temperature parameter can be added as a fourth-dimensional index to construct a four-dimensional discrete flow compensation mapping set, which is suitable for high and low temperature flavoring compounding conditions. Both the three-dimensional and four-dimensional index architectures are equivalent implementations within the scope of protection of this invention.

[0183] Preferably, the high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors provided in this embodiment includes step S600 as follows:

[0184] Step S610: Based on the preset basic opening and closing time of the nozzle, the flow deviation of the basic opening and closing time is linearly corrected by combining the flow-time linear compensation coefficient and the real-time flow compensation value to obtain the nozzle opening and closing time after the initial flow correction.

[0185] The system reads the preset nozzle rated basic opening and closing time of the batching process, combines the pre-calibrated flow-time linear compensation coefficient with the real-time flow compensation value obtained from step S500, and performs positive or negative correction on the theoretical basic opening and closing time through linear calculation to offset the discharge metering deviation caused by pipeline flow drift and conveying slippage, and outputs the initial opening and closing time of the nozzle after only flow dimension correction.

[0186] The formula for calculating the nozzle opening and closing time after initial flow correction is as follows:

[0187] (12)

[0188] In formula (12), The initial opening and closing time of the nozzle after linear correction of flow deviation, in seconds, is used to characterize the basic control time after the nozzle reference opening and closing time is compensated for by flow deviation. It is a prerequisite parameter for subsequent secondary correction of volatile loss time. The basic opening and closing time of the nozzle is preset in seconds. It is obtained by offline calibration under standard density and standard viscosity fragrance conditions. It is the theoretical nozzle opening and closing time under ideal and unbiased conditions. The flow-time linear compensation coefficient, in units of s / mg, is calibrated and solidified offline through multiple sets of gradient feeding tests. It characterizes the nozzle opening and closing time correction amount corresponding to a unit feeding deviation and establishes a linear mapping relationship between the discharge quality deviation and the control time at the actuator. The real-time flow compensation value obtained from the three-dimensional lookup table in formula (11) is in mg, representing the discharge deviation caused by pipeline adhesion, fluid inertia, and nozzle delay under the current operating conditions. The control logic of formula (12) is to retrieve the preset basic opening and closing time of the nozzle under standard operating conditions. And the offline calibrated traffic-duration linear compensation coefficient Multiply the real-time flow compensation amount by the linear compensation coefficient to obtain the duration correction amount; add the duration correction amount to the basic opening and closing time to obtain the nozzle opening and closing time after the initial correction of the flow deviation; when the compensation amount is positive, extend the nozzle opening time for supplementary feeding, and when the compensation amount is negative, close the nozzle in advance to suppress overfeeding. This formula (12) takes the basic opening and closing time of the nozzle under standard working conditions as the benchmark, and combines the offline calibrated flow-time linear compensation coefficient to linearly convert the real-time flow feeding deviation into the nozzle duration correction amount, so as to realize the feedforward pre-compensation of the discharge deviation at the execution end; by adjusting the opening and closing time in the positive / reverse direction to offset the steady-state deviation of fluid transportation, the steady-state feeding accuracy of flavor trace ingredients is improved from the control execution level, and a reliable benchmark parameter is provided for the secondary correction of subsequent volatilization loss time.

[0189] Step S620: Combine the material volatility coefficient and the volatilization time compensation coefficient, and calculate the supplementary discharge time required for volatilization loss based on the initial correction time of the flow rate.

[0190] The volatility coefficient of the material in the multidimensional physical properties of the current ingredients is retrieved and matched with the pre-set volatilization time compensation coefficient of the flavoring. The nozzle opening and closing time after the initial flow correction is used as the benchmark. The discharge time that needs to be compensated for due to the loss of material volatilization at room temperature is calculated by proportional calculation. This is used to make up for the loss of actual feeding quality caused by material volatilization during the nozzle discharge process.

[0191] The formula for calculating the time required to replenish the material due to volatile loss is as follows:

[0192] (13)

[0193] In formula (13), The supplementary discharge time, measured in seconds, is used to compensate for the loss due to flavor evaporation. It is used to make secondary compensation for the nozzle opening and closing time after the initial flow correction, making up for the material loss due to evaporation during the process of material settling and waiting for the nozzle to open. The evaporation time compensation coefficient is dimensionless and is calibrated and solidified offline through multiple sets of isothermal gradient evaporation tests. It is used to characterize the proportional benchmark of time compensation under different volatile materials and to establish a quantitative mapping relationship between evaporation rate and nozzle opening and closing time. Volatility coefficient of the material, unit These are physical property parameters obtained from the initial raw material testing, which directly determine the amount of vapor phase volatilization loss of the fragrance per unit time. The larger the value, the more severe the material volatilization loss. The initial opening and closing time of the nozzle after linear correction of flow deviation in formula (12) is used as the reference time scale for compensation of volatilization loss time. The longer the feeding duration, the greater the cumulative volatilization loss. The control logic of formula (13) is to retrieve the volatility coefficient obtained by pre-detection of the material, the volatilization time compensation coefficient calibrated offline, and the reference opening and closing time of the nozzle after correction of flow dimension; the additional discharge time required under the current working condition is obtained by multiplying the three; the stronger the volatility and the longer the feeding duration, the longer the discharge time required to be compensated; the volatilization compensation time is superimposed on the initial correction time to obtain the final nozzle execution opening and closing time. The volatilization loss of the fragrance material is compensated by appropriately extending the nozzle opening time. Formula (13) uses the nozzle opening and closing time after the initial flow correction as the benchmark, and combines the material volatility coefficient and the offline calibrated volatilization time compensation coefficient to calculate the supplementary discharge time used to compensate for gas phase loss. By introducing the volatile properties of fragrance into the time compensation mechanism, the material escape loss in the feeding process of highly volatile fragrances is compensated in a targeted manner, solving the feeding accuracy defect caused by ignoring volatilization loss in traditional batching control, realizing the layered compensation of dual errors of flow and volatilization, and further improving the control accuracy of trace fragrance batching.

[0194] Step S630: Add the flow correction time and the evaporation compensation time together, and constrain the flow rate correction time to the safe time interval of the equipment by using the value range truncation operator to obtain the actual discharge time of the nozzle that is adapted to the real-time operating conditions.

[0195] The initial total discharge time is obtained by summing the initial flow correction time and the evaporation loss compensation time. The upper and lower limit thresholds of the solenoid valve safety timing are called, and the total time is limited by the value range truncation operator. When the total time exceeds the upper limit, the safe maximum value is taken, and when it is lower than the lower limit, the safe minimum value is taken. Finally, the actual discharge time of the nozzle that can be directly issued and executed is output.

[0196] The formula for setting the actual discharge time of the nozzle is:

[0197] (14)

[0198] In formula (14), The final actual discharge opening and closing time of the nozzle, in seconds, is the underlying execution control command output in this step, which is directly sent to the nozzle solenoid valve drive unit. The initial opening and closing time of the nozzle after linear correction for flow deviation; To compensate for the loss of fragrance vapor phase volatilization time, the two are added together to obtain the theoretical total discharge time after two-dimensional error compensation. , The upper and lower limits of nozzle hardware safety timing are constrained by the minimum response characteristics of the solenoid valve and the maximum continuous feeding safety conditions of the pipeline, and the hardware calibration range is fixed as follows: The lower limit prevents the time from being shorter than the solenoid valve's response threshold, resulting in no material output, while the upper limit avoids excessive opening time leading to large-scale material overfeeding, workstation contamination, and raw material scrapping. This is a range truncation operator, belonging to explicit limit logic: it forces output when the total duration is less than the minimum value. Force output when the value is greater than the maximum value. If the time is within the range, the original time after superposition is directly output. The control logic of formula (14) is to arithmetically sum the flow dimension correction time and the volatile loss compensation time to obtain the theoretical nozzle opening and closing time after double error compensation; to perform safety limit processing on the total time through the value range truncation operator, and strictly constrain the execution time within the time safety range allowed by the equipment; and to issue the actual discharge time after the limit as the final control command to the solenoid valve to perform one batching. This formula (14) superimposes the flow deviation correction time and the volatile loss compensation time, and constrains the total opening and closing time within the hardware safety time range through the value range truncation operator. Internally, the actual discharge time that the output nozzle can safely execute; relying on dual-mechanism feedforward compensation combined with hardware boundary safety limits, it can achieve high-precision micro-volume dispensing of multiple types of flavorings while avoiding production safety failures caused by extreme parameters, taking into account control accuracy, material versatility and batch production safety.

[0199] Furthermore, the high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors provided in this embodiment specifically includes step S700 as follows:

[0200] Step S710: Collect the final actual net weight of material fed in this cycle, calculate the absolute deviation and relative deviation of a single feeding, and construct a standardized discharge deviation feature vector.

[0201] After a single batching operation is completed, the final actual net weight of the material fed into the station is read after the weighing module stabilizes. The difference between the actual net weight of the material fed into the station and the target amount of material fed into the current cycle is used to obtain the absolute feeding deviation. The absolute deviation is divided by the target amount of material fed into the station to obtain the relative feeding deviation. The two types of deviation parameters are normalized and arranged in an orderly manner to construct a standardized one-dimensional discharge deviation feature vector.

[0202] Absolute deviation is the mass difference between the actual amount of material fed and the target amount of material fed; relative deviation is the ratio of the absolute deviation to the target amount of material fed, used to characterize the relative error level of material feeding.

[0203] The formulas for calculating absolute deviation and relative deviation are:

[0204] (15)

[0205] In formula (15), The absolute deviation of a single batch of feed is expressed in mg. It is obtained by taking the absolute value of the difference between the final actual net weight of the feed and the target feed amount. It is used to intuitively characterize the actual mass amplitude of the feed error and is suitable for error quantification assessment under large feed range. This represents the final, filtered, and stabilized net weight of the material after the completion of this batching cycle, in mg. It is the true weighing value after the batching loop is closed. The preset target feed amount, in mg, represents the theoretically set feed mass for this batch of ingredients. The relative deviation of material input is dimensionless and has a range of values. It characterizes the proportion of error relative to the target feed amount, and is specifically adapted to the error severity evaluation in the scenario of small feed amount of trace flavorings, which can eliminate the interference of feed amount size on error evaluation. To minimize the zero-prevention constant, a fixed value is taken. This is used to prevent the division-by-zero calculation anomaly caused by the denominator being 0 when the target feed amount infinitely approaches 0, and to ensure the numerical stability of the deviation calculation under the micro-scale extreme working condition. The control logic of formula (15) is to read the final stable weighing data after the feed is completed, make the difference with the target feed amount and take the absolute value to obtain the feed absolute deviation; on the basis of the target feed amount, a minimum zero-prevention constant is superimposed as the denominator, and the absolute deviation is divided by the denominator to obtain the dimensionless relative deviation; the absolute deviation and the relative deviation are combined into a two-dimensional feed deviation feature vector, which is used for subsequent deviation type classification, control parameter iterative optimization and feed quality statistical traceability. This formula (15) calculates the absolute feed deviation and dimensionless relative feed deviation of a single feed respectively, and avoids the division-by-zero anomaly of the denominator under the micro-scale working condition by introducing a minimum zero-prevention constant; relying on the two-dimensional deviation characterization method, it takes into account both the evaluation of the large range error amplitude and the evaluation of the micro range error ratio, and is suitable for the accurate identification of deviation in the flavor wide range feed scenario, providing standardized error feature data for feed quality traceability and control parameter closed-loop iterative optimization.

[0206] The formula for constructing the feature vector of material discharge deviation is:

[0207] (16)

[0208] In formula (16), The discharge deviation feature vector is a standardized discharge deviation feature vector after multi-source fusion. It is a one-dimensional row vector transpose, dimensionless, and serves as the standardized input feature for subsequent deviation root cause classification and adaptive optimization model of control parameters, realizing the quantitative characterization and root cause tracing of feeding errors. The absolute deviation of a single feeding is the mass amplitude of the feeding error, used to reflect the actual material loss due to overfeeding or underfeeding. To address relative feeding deviations, eliminate the influence of feeding range, and accurately characterize the severity of errors in micro-batch scenarios. The volatility coefficient of flavoring materials is an inherent physical property parameter of the material, used to characterize the inherent deviation causes of material volatilization loss in the gas phase. The stirring speed at the batching station is measured in r / min. It is a parameter of the equipment's operating conditions and is used to characterize equipment-related deviations such as fluid uniformity and pipeline transport disturbances. The control logic of formula (16) is to extract four types of data in sequence: absolute feeding deviation, relative feeding deviation, material volatile property parameters, and real-time stirring speed at the station for this batching cycle. The four types of features are then spliced ​​into a one-dimensional structured feature vector in a fixed order and transposed to construct a discharge deviation feature vector. This discharge deviation feature vector is then input into the deviation root cause identification model to complete the classification and location of property-related and equipment-related error causes. It is also used for offline iterative optimization of nozzle flow compensation parameters and volatile compensation coefficients. This formula (16) structurally splices together four types of parameters—absolute feeding deviation, relative feeding deviation, material volatility coefficient, and station stirring speed—into a discharge deviation feature vector, integrating multi-dimensional information such as feeding error, material properties, and equipment operating conditions. It not only provides a standardized model input for closed-loop iterative optimization of control parameters, but also enables precise tracing and positioning of feeding deviation causes related to physical properties and equipment. While improving the long-term steady-state accuracy of batching, it also realizes intelligent root cause analysis of process failures.

[0209] The discharge deviation feature vector given in this formula is a basic simplified 4-dimensional form, which is only used as a standardized and simplest input example. In actual engineering, features such as material density, pipeline pressure, ambient temperature, and nozzle loss coefficient can be superimposed to expand to 7 dimensions and more. The vector construction and matrix transpose operation rules remain unchanged, and the input dimensions can be flexibly expanded. Those skilled in the art can adjust the number of features according to the process accuracy requirements to fully reproduce the solution process of the feeding deviation correction coefficient.

[0210] Step S720: Input the fusion feature vector of the material discharge deviation into the gradient boosting integrated prediction model containing the weak regression tree, and impose a value range truncation constraint on the model output to obtain the material discharge deviation correction coefficient.

[0211] The gradient boosting integrated prediction model mentioned in this step is a two-stage joint architecture. The first layer performs a deviation classification task to identify whether the current feeding is caused by viscosity, volatility, conveying inertia, or weighing noise. The second layer performs regression calculations based on the classification results to solve for the adaptive correction coefficient of the feeding deviation. The gradient boosting integrated prediction model is a general term for the overall two-stage architecture. The calculation of the core correction amount at the bottom layer adopts regression calculation.

[0212] The model input features are fused together with the discharge deviation feature vector, physical and chemical characteristics of the material, station conveyor speed, ambient temperature and humidity, and other station operating parameters. These features are then fed into a pre-trained and converged gradient boosting integrated prediction model. The deviation evolution law is fitted through serial iteration using multiple weak regression trees. The correction coefficients of the original model output are subject to value range truncation constraints to avoid excessively large coefficients that could cause severe fluctuations in feeding. Finally, a stable and reliable feeding deviation correction coefficient is output.

[0213] Weak regression tree, the basic weak learner of the gradient boosting model, is used to fit the nonlinear law of feeding deviation in segments; feeding deviation correction coefficient is used to correct the scaling ratio parameter of the target feeding amount in the next cycle.

[0214] The formula for predicting the feeding deviation correction coefficient is:

[0215] (17)

[0216] In formula (17), This is an adaptive correction coefficient for feeding deviation, dimensionless, and constrained within a safe range. Within the range, adaptive iterative corrections are used for the nozzle flow compensation coefficient and evaporation compensation coefficient in the next batch, achieving closed-loop self-optimization of control parameters. To increase the total number of weak regression decision trees within the gradient boosting ensemble model, the calibration was fixed after optimization based on operating conditions. To balance the model fitting accuracy with the real-time performance of online inference. For the first The explicit feature mapping function of the weak regression decision tree, the branching rules and leaf node outputs of each decision tree are all solidified through offline training of historical multi-condition batching samples. It belongs to a completely reproducible deterministic lookup table mapping logic and does not have a black box uninterpretable structure. The multi-source fusion discharge deviation feature vector constructed in the preceding steps serves as the unified input of the integrated model, carrying multi-dimensional information on feeding deviation, material properties, and equipment operating conditions. The range truncation operator limits the original predicted output of the model to a certain value. To prevent the correction coefficient from being too large or too small under extreme working conditions, which could cause continuous system oscillations or drastic changes in batching parameters and loss of control. The control logic of formula (17) is to input the current batch fusion deviation feature vector into 80 offline trained weak regression decision trees in sequence to obtain the prediction output of each tree; to sum the outputs of all weak learners to obtain the original prediction correction coefficient of the ensemble model; and to constrain the prediction coefficient within a safe range by using the range truncation operator. Within this model, a stable and reliable adaptive correction coefficient for feeding deviation is obtained. This adaptive correction coefficient is then used to iteratively update the flow-time linear compensation coefficient and the evaporation time compensation coefficient to complete the adaptive optimization of control parameters for the next batch. Formula (17) inputs the multi-dimensional discharge deviation feature vector into the gradient boosting ensemble model composed of 80 weak regression decision trees. After accumulating the prediction results of all weak learners, the model is then... The adaptive feeding correction coefficient is obtained by truncating the safety range; by relying on integrated learning to fit the nonlinear feeding error of material and working condition coupling, the closed-loop intelligent iterative optimization of compensation control parameters is realized. Compared with fixed ratio correction, the convergence speed is faster and the adaptive capability of multiple types of materials is stronger. At the same time, the amplitude limit constraint ensures the stability of the industrial control system without oscillation.

[0217] Step S730: The target feeding amount for the current cycle is scaled and corrected using the feeding deviation correction coefficient to obtain the corrected feeding amount for the next feeding cycle; at the same time, the current working condition parameters and discharge deviation data are updated to the material flow compensation mapping set to achieve iterative optimization of the compensation knowledge base.

[0218] The target feed amount for the current cycle is multiplied and scaled with the feed deviation correction coefficient to obtain the corrected feed amount for the next batching cycle after deviation closed-loop correction. The operating condition data such as material viscosity, conveying gear, remaining feed amount, and actual discharge flow deviation corresponding to this batching are used as new calibration samples and added to the material flow compensation mapping set to realize online iterative optimization of the flow compensation knowledge base and continuously improve the accuracy of subsequent flow compensation.

[0219] The closed-loop feeding correction formula is:

[0220] (18)

[0221] In formula (18), The corrected feed amount for the next feeding cycle is the new target feed amount after deviation adaptive correction for the next batching cycle, in mg. It serves as the set feed benchmark value for the next batching task, realizing closed-loop iterative optimization of the feed setting value. This is the original target feed amount for this batching cycle, in mg. This is the feed deviation correction coefficient predicted by the gradient lifting ensemble model and limited by a safety range. It is dimensionless and its value range is [missing information]. This is used to quantify the overall feeding deviation ratio of the system under the current working condition and to perform proportional scaling correction on the target feeding amount. The control logic of formula (18) is to multiply the original target feeding amount of the current cycle by the adaptive feeding deviation correction coefficient, and obtain the target feeding amount after closed-loop optimization in the next cycle through proportional scaling. If the overall feeding is under-fed in this cycle, the correction coefficient is greater than 1 and the set feeding amount for the next cycle is increased; if the feeding is over-fed, the correction coefficient is less than 1 and the set feeding amount is decreased. The dynamic viscosity of the material, conveying speed, remaining amount to be fed, and actual discharge deviation of the current cycle are synchronously stored in the flow compensation discrete mapping knowledge base to complete the incremental iterative update of the offline compensation data table. In the next cycle, the corrected target feeding amount will be used as the set value, and the nozzle duration feedforward compensation control will be performed in combination with the updated flow compensation knowledge base to form a two-layer self-learning control architecture of feedforward lookup table compensation + back-end feeding parameter closed-loop iteration. This formula (18) scales the target feed amount for the current cycle by using the feed deviation correction coefficient to generate the optimized target feed setting value for the next cycle. Simultaneously, it updates the current cycle working conditions and deviation data increments to the three-dimensional flow compensation mapping knowledge base, forming a double closed-loop self-learning architecture of feedforward lookup table compensation and feed parameters. This architecture can effectively adapt to the slow drift problem of flavor raw material batch properties, realize the autonomous iteration and adaptive correction of the control system, and continuously improve the steady-state accuracy of long-term batch micro-batch feeding.

[0222] The estimated completion time of batching at all online batching stations is collected. The goal is to minimize the total deviation by constructing a time-series balanced scheduling objective function with the aim of making the completion time of all stations as consistent as possible. The optimal material feeding start-up sequence and the corrected material feeding amount distribution scheme for each station are solved under the constraints of the rated conveying capacity of each station equipment and the upper and lower limits of the corrected material feeding amount. This minimizes the deviation between the completion time of each station and the global average completion time, thereby achieving balanced and collaborative batching scheduling for multiple stations.

[0223] Step S740: Taking the time balance of the batching end time of all online batching stations as the optimization objective, construct the multi-station collaborative time balance scheduling objective function, solve the optimal correction feeding quantity distribution scheme for each station, and minimize the total deviation between the batching end time of each station and the average end time.

[0224] The time-balanced scheduling objective function is an optimization function used to quantify the degree of imbalance in the production cycle of multiple workstations. The optimal solution corresponds to the most stable collaborative feeding time sequence scheme for each workstation.

[0225] The objective function for multi-station collaborative time-balanced scheduling is:

[0226] (19)

[0227] In formula (19), The total cost function for multi-station time-series balancing is dimensionless and represents the sum of absolute deviations of the end time of all online batching stations relative to the overall average end time. The optimization objective is to find the target material feeding scheme for each station that minimizes the total cost, thereby achieving cycle time balancing for each station on the production line. For the first The estimated end time of this batching cycle for each batching station, in seconds, is estimated based on the current remaining amount to be added at the station, the conveying speed, and the duration of nozzle opening and closing. The arithmetic mean of the completion times of batching at all online batching stations is used as the global standard timing reference for the production line. This represents the total number of batching stations currently in operation within the system. To minimize the optimization operator, we represent solving for the optimal values ​​of the decision variables such that the subsequent objective function... The value reaches its minimum. To minimize the cost function To optimize the objective, the optimal decision is solved. Optimization constraints: The target feeding amount after closed-loop correction of each workstation is used as the decision variable. Within the allowable range of feeding accuracy of a single workstation, the timing of the feeding task of each workstation is adjusted to minimize the total global timing deviation. The control logic of formula (19) is to predict the end time of the current round of feeding of each online batching workstation in real time, count the end time of all workstations and calculate the global average end time; use the sum of the absolute deviations of the relative average of the end times of all workstations as the optimization cost to construct the optimization objective; within the safety constraint range of the feeding accuracy of a single workstation, the target feeding amount after closed-loop correction of each workstation is coordinated and scheduled to solve the optimal distribution scheme and compress the production timing difference between workstations; the feeding task of each workstation is distributed according to the optimized timing scheme to balance the start and end time of feeding of each workstation and avoid some workstations from completing early and being idle for a long time, and some workstations from being continuously congested and waiting in line. Formula (19) uses the sum of the absolute deviations of the end time of all online batching stations relative to the global average end time as the total cost of time-series balancing. It solves the optimal feeding and distribution scheme of each station by minimizing the total cost. Under the constraint of ensuring the accuracy of micro-batch batching at a single station, it balances the production cycle of each station, avoids station congestion or equipment idling, realizes parallel and collaborative production of multiple stations, and significantly improves the overall throughput efficiency and equipment utilization of the entire flavor ingredient production line.

[0228] Please see Figure 2This embodiment provides a high-precision intelligent formulation and ingredient coordination control system for flavorings and fragrances, used to implement the above-mentioned high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances. It includes a target feed quantity prediction module 10, a container positioning coordinate acquisition module 20, a remaining material to be fed acquisition module 30, a conveying speed segmented adjustment instruction generation module 40, a real-time flow compensation value acquisition module 50, an actual discharge time acquisition module 60, and a workstation collaborative scheduling and continuous ingredient control module 70. The target feed quantity prediction module 10 is used to acquire encrypted material formulation data and perform decryption and integrity verification on the encrypted material formulation data. The system extracts the physical and chemical characteristics of each ingredient to construct a multi-dimensional material property feature set. A pre-set machine learning algorithm is used to fit and calculate this feature set, accurately predicting the target feed amount for each material. A container positioning coordinate acquisition module 20 matches the corresponding material container based on the target feed amount, drives the material transfer actuator to complete the transfer and positioning of the material container, and obtains the container positioning coordinates. A remaining material quantity acquisition module 30 activates the corresponding workstation's weight monitoring module based on the container positioning coordinates, collecting real-time data on the amount of material already fed into the material container. The system uses the weight data of the materials to obtain the remaining amount of materials to be added under the current batching condition; the conveying speed segment adjustment command generation module 40 compares the remaining amount of materials to be added with a preset micro-material weight threshold, adaptively switches between coarse feeding mode and micro-fine feeding mode based on the comparison result, and generates a conveying speed segment adjustment command for the corresponding condition; the real-time flow compensation value acquisition module 50 retrieves a pre-constructed material flow compensation mapping set based on the segment adjustment command, and matches it with the current material properties and the remaining amount of materials to be added to obtain the real-time flow compensation value; the actual discharge time acquisition module 60 is used to obtain the actual discharge time using real-time... The flow compensation value dynamically corrects and adjusts the basic opening and closing control time of the dispensing nozzle, and performs time compensation in combination with the material volatilization loss characteristics to obtain the actual discharge time adapted to the real-time working conditions; the station collaborative scheduling and continuous dispensing control module 70 is used to complete the single dispensing according to the actual discharge time and collect the discharge deviation data of the current dispensing cycle; the discharge deviation data, material physicochemical characteristics and station operating parameters are input into the trained gradient improvement integrated prediction model to complete the closed-loop correction of the dispensing amount and obtain the corrected dispensing amount for the next dispensing cycle; based on the corrected dispensing amount, multi-dispensing station collaborative scheduling and high-precision continuous dispensing control are realized.

[0229] The following detailed description, using specific embodiments, illustrates the high-precision intelligent formulation and ingredient coordination control method and system for fragrances and flavors provided by this invention:

[0230] I. Basic operating condition settings for the embodiment (all parameters strictly match the aforementioned formula thresholds, dimensions, and value ranges)

[0231] 1. Test material: lemon oil flavoring

[0232] 2. The calibration table of core process parameters is shown in Table 1:

[0233] Table 1

[0234]

[0235] 3. Measured physical properties of lemon oil flavoring (falling within the preset value range of the formula)

[0236] ;

[0237] density (range Dynamic viscosity (range Volatility coefficient (range ), solubility (dimensionless) ).

[0238] 4. Hardware Configuration

[0239] The multi-station feeding line has a total of 4 parallel feeding stations; optional cup specifications: , , Weighing sensor accuracy Solenoid valve dispensing nozzle.

[0240] II. Step S100: Formula decryption, integrity verification, and target feed quantity prediction (corresponding formulas 1-4)

[0241] Step S110: Recipe Decryption and Integrity Verification

[0242] 1. The host computer obtains the encrypted lemon oil flavoring formula. Using RSA private keys Decryption using the modular exponentiation method according to formula (1) .

[0243] 2. Calculate the SHA-256 hash digest of the decrypted plaintext and bind it to the recipe distribution timestamp. , and the original check code XOR comparison, The formula is complete and unaltered, and is allowed to proceed to the next stage of the process.

[0244] Step S120: Construct multidimensional physical property feature vectors

[0245] Substituting the measured physical property parameters, construct the eigenvector according to formula (3):

[0246] .

[0247] Step S130: Random forest regression to predict target feed quantity

[0248] Based on the ensemble prediction of 120 decision trees, the calculation is performed according to formula (4):

[0249] ;

[0250] ,exist The value range is reasonable, non-negative, and meets the threshold constraint.

[0251] III. Step S200: Cup matching, visual positioning, and workstation allocation (Formulas 5-7)

[0252] Step S210: Matching the material cup specifications

[0253] Theoretical volume of material:

[0254] ;

[0255] Screening with a safety margin according to formula (5):

[0256] ;

[0257] Select the smallest available cup size: This satisfies the volume constraint.

[0258] Step S220: Convert camera pixel coordinates to world physical coordinates

[0259] Visual acquisition of material cup pixel coordinates By pre-calibrating the camera intrinsic parameter matrix and scaling factor The container's positioning coordinates are obtained by conversion using formula (6):

[0260] .

[0261] Step S230 Optimal workstation allocation

[0262] Calculate the Euclidean distance cost between the material cup position and the reference coordinates of the four idle workstations. Solve using formula (7) to find the optimal workstation to be assigned as the No. 2 material preparation workstation.

[0263] IV. Step S300: Weighing and filtering + calculation of remaining amount to be invested (Formulas 8 and 9)

[0264] Step S310: Sliding window mean filtering for noise reduction

[0265] The weighing sensor at station 2 collects 15 sets of raw sample values ​​in real time (window length). ): ;

[0266] Moving mean filtering using formula (8):

[0267] ;

[0268] Step S320: Calculation of remaining material to be added

[0269] Calculate according to the non-negativity constraint in formula (9):

[0270] .

[0271] V. Step S400: Adaptive switching between coarse and fine feeding modes (Formula 10)

[0272] This triggers the coarse feeding mode:

[0273] ;

[0274] The feed is continuously coarsely fed, and when the remaining feed amount drops below 120mg, it automatically switches to fine feeding. .

[0275] This embodiment records the node switching process: when material is fed to... hour, Switch to the fine feeding mode.

[0276] VI. Step S500: 3D Index Traffic Compensation Lookup Table (Formula 11)

[0277] Current operating parameters: Using the three sets of parameters as a three-dimensional index, the pre-built traffic compensation mapping table is queried, and the result is obtained by looking up the table using formula (11): A negative value indicates that the nozzle opening and closing time needs to be shortened to suppress over-discharge deviation caused by fluid inertia.

[0278] VII. Step S600: Dual correction setting of nozzle opening and closing time (Formulas 12, 13, 14)

[0279] Step S610 Flow deviation linear duration correction (Formula 12)

[0280] ;

[0281] Step S620: Compensation for evaporation loss time (Formula 13)

[0282] ;

[0283] Step S630 cuts off the safety interval to obtain the actual discharge time (Formula 14).

[0284] ;

[0285] Falling Within the safe zone, the nozzle will discharge material for the specified duration.

[0286] 8. Step S700: Material feeding closed-loop correction + multi-station collaborative scheduling (Formulas 15-19)

[0287] Step S710: Construction of discharge deviation characteristics

[0288] After a single feeding is completed, the sensor collects the final actual feeding amount: Substitute into formula (15):

[0289] ;

[0290] ;

[0291] Workstation stirring speed The deviation feature vector is obtained by concatenating the features (Formula 16):

[0292] ;

[0293] Step S720: Prediction of deviation correction coefficient (Formula 17)

[0294] Gradient boosting model prediction results are truncated within their range:

[0295] ;

[0296] Step S730: Adjust the feed amount for the next cycle (Formula 18)

[0297] ;

[0298] Simultaneously, the viscosity, conveying speed, remaining feed amount, and discharge deviation data are updated into the three-dimensional flow compensation mapping knowledge base to achieve self-learning iteration.

[0299] Step S740 Multi-station collaborative timing balance scheduling (Formula 19)

[0300] The completion times for this batching process at the four workstations were 326s, 329s, 331s, and 324s, respectively. The average completion time was... .

[0301] With the goal of minimizing the total timing deviation, the corrected material input is evenly distributed to each workstation, ultimately reducing the maximum waiting time at each workstation from the traditional 42.6s to 9.3s, thus balancing the production line cycle time.

[0302] IX. Supporting Evidence from the Test Results and Beneficial Effects of this Embodiment

[0303] 1. Single feeding accuracy results in this embodiment

[0304] Target feed intake: 1260.00 mg;

[0305] Actual feed dosage: 1259.12 mg;

[0306] Absolute feeding error: This is far lower than the average error of ±7.62mg for traditional fixed-time batching;

[0307] The relative error of the feed is 0.0698%, which meets the quality control requirements of the fragrance and flavor industry for high-precision micro-batch formulation.

[0308] 2. Statistical validation through multiple repeated experiments (30 consecutive batches of lemon oil flavoring)

[0309] Table 2 shows a comparison of the beneficial effects of this invention with traditional open-loop formulation methods:

[0310] Table 2

[0311]

[0312] The high-precision intelligent formulation and ingredient coordination control system for flavorings and fragrances provided in this embodiment has the following specific advantages compared with the prior art:

[0313] 1. This embodiment uses a dual security mechanism of RSA asymmetric encryption + SHA-256 hash timestamp to prevent formula leakage, tampering and replay attacks. At the same time, it uses multi-dimensional physical properties such as density, viscosity, volatility and solubility combined with a random forest regression model to predict the target feed amount and quantitatively compensate for the conveying loss of various materials, thus solving the industry pain point of large deviation between formula security and initial feed benchmark from the source.

[0314] 2. This embodiment automatically matches the material cup specifications with the feeding volume, uses robotic arm vision 3D positioning and Euclidean distance for optimal station allocation, eliminates container range mismatch and station off-center weighing errors, improves the reliability of weighing data acquisition, and provides hardware data support for accurate calculation of the remaining amount to be fed.

[0315] 3. This embodiment uses sliding time-average filtering to suppress weighing pulse noise, and the calculation of the remaining amount to be fed is stable and reliable. It can accurately trigger the adaptive switching between coarse feeding and micro-feeding modes, taking into account both batching production efficiency and micro-feeding control accuracy.

[0316] 4. This embodiment constructs a three-dimensional indexed flow compensation lookup table knowledge base for viscosity, conveying speed, and remaining amount to be added, and corrects pipeline flow drift deviation online in real time. At the same time, it combines the material volatility coefficient to perform secondary compensation and adjustment of nozzle opening and closing time, eliminating the two core metering interferences of flow and volatility from two dimensions, and greatly improving the accuracy of micro-dispensing.

[0317] 5. This embodiment uses a gradient boosting ensemble learning model to automatically trace the source of feeding deviation and solve the correction coefficient, completing the closed-loop iterative correction of feeding quantity, and realizing online self-optimization of the flow compensation knowledge base. Through the collaborative scheduling of multi-station time-series balancing objective function, it effectively avoids material blockage, pipeline pressure imbalance, and production rhythm disorder in parallel feeding at multiple stations, and realizes intelligent closed-loop material control of fragrance and flavor with safety, high precision, self-iteration, and multi-station collaboration throughout the entire process.

[0318] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for high-precision intelligent formulation and ingredient coordination control of flavorings and fragrances, characterized in that, Includes the following steps: S100. Obtain encrypted material formula data, perform decryption and integrity verification on the encrypted material formula data, extract the physical and chemical features of each ingredient, and construct a multi-dimensional material property feature set; use a preset machine learning algorithm to fit the multi-dimensional material property feature set to accurately predict the target feed amount of each material. S200: Based on the target feeding amount, match the corresponding specification of the material holding container, drive the material transfer actuator to complete the transfer and positioning of the material holding container, and obtain the container positioning coordinates corresponding to the material holding container; S300. Based on the container positioning coordinates, the weight monitoring module of the corresponding workstation is activated to collect the weight data of the material already added in the material holding container in real time, and to obtain the amount of remaining material to be added under the current batching condition. S400. Compare the remaining amount of material to be fed with the preset micro-material weight threshold, and adaptively switch between coarse feeding mode and micro-fine feeding mode based on the comparison result, and generate a segmented adjustment command for the conveying speed corresponding to the working condition. S500: Based on the segmented adjustment command, retrieve the pre-constructed material flow compensation mapping set, and combine the current material property parameters with the remaining amount of material to be added to obtain the real-time flow compensation value. S600: The basic opening and closing control time of the feeding nozzle is dynamically corrected and adjusted using the real-time flow compensation value, and the time compensation is performed in combination with the material volatilization loss characteristics to obtain the actual discharge time that is adapted to the real-time working conditions. S700: Complete a single batching operation according to the actual discharge time, and collect the discharge deviation data of the current batching cycle; input the discharge deviation data, physical and chemical characteristics of the material, and station operating parameters into the trained gradient boosting integrated prediction model to complete the closed-loop correction of the batching amount and obtain the corrected batching amount for the next batching cycle; realize the collaborative scheduling of multiple batching stations and high-precision continuous batching control based on the corrected batching amount.

2. The high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors according to claim 1, characterized in that, Step S100 specifically includes: S110. The RSA asymmetric encryption and decryption algorithm is used to perform modular exponentiation decryption on the encrypted material formula ciphertext data and the local preset private key to obtain the decrypted plaintext formula dataset. The data integrity is verified by combining the SHA-256 hash digest operator with the unique timestamp and the original check code. Only the compliant formula data that passes the verification is retained, and the tampered abnormal data is removed. S120: Extract four core physicochemical parameters—material density, dynamic viscosity, volatility coefficient, and solubility—from compliant formulation data to construct a structured multidimensional material property feature vector. S130. An integrated fitting and prediction method is used to fit the multidimensional physical property feature vector of the material using a random forest regression model that includes decision trees. The mean of all decision tree outputs is taken and the results are constrained to be non-negative to obtain the target feed amount for each ingredient.

3. The high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors according to claim 1, characterized in that, The material holding container is a material cup, and the material transfer actuator is a transfer robot. Step S200 specifically includes: S210. Combine the target feeding amount with the material density to calculate the theoretical filling volume of the material, set the safety margin coefficient, screen the material cups whose rated volume meets the constraints as matching containers, and establish a one-to-one mapping relationship between feeding amount, material cup number, and material batching station coordinates. S220 drives the transfer robot to complete the grasping, shifting and placement of the cup. It collects the pixel coordinates of the cup image through an industrial camera, and uses the pre-calibrated camera intrinsic parameter matrix and pixel physical scaling coefficient to convert the pixel coordinates into the three-dimensional world coordinates of the production line to obtain the precise positioning coordinates of the container. S230. Calculate the Euclidean distance between the current material cup position and the reference coordinates of each idle preset workstation as the workstation allocation cost, select the workstation number corresponding to the minimum cost value as the target allocation workstation, and complete the dynamic allocation of multiple workstations.

4. The high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances according to claim 1, characterized in that, Step S300 specifically includes: S310. Activate the high-precision weighing sensor at the corresponding workstation based on the container positioning coordinates. Use a fixed sliding time window to perform sliding mean filtering and noise reduction on the original weighing sampling sequence to eliminate weighing pulse noise caused by mechanical vibration and electromagnetic interference, and obtain the real-time net weight of the material after filtering. S320. Subtract the filtered real-time net weight of the material from the target feed amount, and ensure that the value is non-negative by the maximum value constraint, to calculate the remaining amount of material to be fed under the current working condition.

5. The high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances according to claim 4, characterized in that, Step S400 specifically includes: Step S410: Preset the weight threshold of trace materials, the conveying speed of the coarse feeding gear, and the conveying speed of the fine feeding gear; Step S420: When the remaining amount of material to be fed is greater than the preset micro-material weight threshold, execute the coarse feeding mode and match the high-speed conveyor gear to complete the rapid feeding. Step S430: When the remaining amount of material to be fed is less than or equal to the preset micro-material weight threshold, switch to micro-precision feeding mode, match the low-speed conveying gear to complete precise feeding, and realize coarse and fine dual-stage adaptive segmented speed regulation control.

6. The high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances according to claim 5, characterized in that, The method for constructing the material flow compensation mapping set in step S500 includes: Actual discharge deviation data of different viscosities and volatile flavorings under multiple conveying speeds and multiple sets of remaining feed amounts are collected in advance. A discrete lookup table mapping function is constructed using the material dynamic viscosity, real-time conveying speed, and remaining feed amount as a three-dimensional index. During online operation, the three sets of parameters of the material dynamic viscosity, real-time conveying speed, and remaining feed amount under the current working condition are input into the discrete lookup table mapping function to match and obtain the real-time flow compensation value.

7. The high-precision intelligent formulation and ingredient coordination control method for fragrances and flavors according to claim 1, characterized in that, Step S600 specifically includes: S610. Based on the preset basic opening and closing time of the nozzle, the flow rate deviation is linearly corrected by combining the flow rate-time linear compensation coefficient and the real-time flow rate compensation value to obtain the nozzle opening and closing time after the initial flow rate correction. S620. Combining the material volatility coefficient and the volatilization time compensation coefficient, calculate the supplementary discharge time required for volatilization loss based on the initial correction time of the flow rate. S630. Add the flow correction time and the evaporation compensation time, and constrain them within the safe time interval of the equipment by the value range truncation operator to obtain the actual discharge time of the nozzle that is adapted to the real-time working conditions.

8. The high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances according to claim 1, characterized in that, Step S700 specifically includes: S710. Collect the final actual net weight of material fed in this cycle, calculate the absolute and relative deviations of single feeding respectively, and construct a standardized discharge deviation feature vector. The formulas for calculating absolute deviation and relative deviation are: ； in, This refers to the absolute deviation of a single batch of ingredients. This represents the final, filtered, and stabilized net weight of the material after the completion of this batching cycle. To preset the target feed amount, Due to relative deviation in material input, To minimize the zero-prevention constant, a fixed value is taken. ; The formula for constructing the feature vector of material discharge deviation is: ； in, This is the characteristic vector of the material discharge deviation. This represents the absolute deviation of a single feeding operation. The volatility coefficient of the fragrance material. The mixing speed at the ingredient preparation station. It is a vector transpose operator; S720. Input the fused feature vector of the material discharge deviation into the gradient boosting integrated prediction model containing a weak regression tree, and impose a value range truncation constraint on the model output to obtain the material discharge deviation correction coefficient. The formula for predicting the feeding deviation correction coefficient is: ； In formula (17), This is the adaptive correction coefficient for feeding deviation. For the first The explicit feature mapping function of a weak regression decision tree. This is the feature vector of multi-source fusion output deviation. For the range truncation operator, For summation operators; S730: The target feeding amount for the current cycle is scaled and corrected using the feeding deviation correction coefficient to obtain the corrected feeding amount for the next feeding cycle; at the same time, the current working condition parameters and discharge deviation data are updated to the material flow compensation mapping set to achieve iterative optimization of the compensation knowledge base; The closed-loop feeding correction formula is: ； In formula (18), The adjusted feed amount for the next feeding cycle; S740. Taking the time balance of the batching end time of all online batching stations as the optimization objective, construct the multi-station collaborative time balance scheduling objective function, solve the optimal correction feeding quantity distribution scheme of each station, and minimize the total deviation between the batching end time of each station and the average end time.

9. The high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances according to claim 8, characterized in that, In step S740, the objective function for multi-station collaborative time-balanced scheduling is: ； In formula (19), The total cost function for multi-station time-series equalization. For the first The estimated end time of this batching process for each batching station. This is the arithmetic mean of the batching completion times at all online batching stations. This represents the total number of batching stations currently in operation within the system. To minimize the optimization operator, To minimize the cost function To optimize the objective, we need to find the optimal solution.

10. A high-precision intelligent formulation and ingredient coordination control system for flavorings and fragrances, used to implement the high-precision intelligent formulation and ingredient coordination control method for flavorings and fragrances as described in any one of claims 1 to 9, characterized in that, include: The target feed quantity prediction module is used to acquire encrypted material formula data, perform decryption and integrity verification on the encrypted material formula data, extract the physical and chemical dimension features of each ingredient, and construct a multi-dimensional material property feature set; and use a preset machine learning algorithm to fit the multi-dimensional material property feature set to accurately predict the target feed quantity of each material. The container positioning coordinate acquisition module is used to match the material holding container of the corresponding specification based on the target feeding amount, drive the material transfer actuator to complete the transfer and positioning of the material holding container, and acquire the container positioning coordinates corresponding to the material holding container. The remaining material to be added module is used to activate the weight monitoring module of the corresponding workstation based on the container positioning coordinates, collect the weight data of the material already added in the material holding container in real time, and obtain the remaining material to be added under the current batching condition. The conveying speed segment adjustment instruction generation module is used to compare the remaining amount of material to be fed with the preset micro-material weight threshold, adaptively switch between coarse feeding mode and micro-fine feeding mode based on the comparison result, and generate conveying speed segment adjustment instructions for the corresponding working conditions. The real-time flow compensation value acquisition module is used to retrieve a pre-constructed material flow compensation mapping set based on the segmented adjustment command, and match the current material property parameters and the remaining amount of material to be added to obtain the real-time flow compensation value. The actual discharge time acquisition module is used to dynamically correct and adjust the basic opening and closing control time of the batching nozzle using the real-time flow compensation value, and to compensate for the time by combining the material volatilization loss characteristics to obtain the actual discharge time that is adapted to the real-time working conditions. The workstation collaborative scheduling and continuous batching control module is used to complete a single batching operation according to the actual discharge time, and collect the discharge deviation data of the current batching cycle; input the discharge deviation data, physical and chemical characteristics of the material, and workstation operating parameters into the trained gradient boosting integrated prediction model to complete the closed-loop correction of the batching amount and obtain the corrected batching amount for the next batching cycle; and realize multi-batching workstation collaborative scheduling and high-precision continuous batching control based on the corrected batching amount.

Images