A method and system for predicting and closed-loop regulating uniformity of daqu and starter

Abstract

The present application relates to a kind of dregs of fermentation liquor and evenness prediction and closed-loop control method and system.The method includes: the characteristic data in the process of dregs of fermentation liquor airing is collected and input characteristic vector is constructed;Input characteristic vector is input into the prediction model trained in advance to output predicted fermentation time, wherein the prediction model is trained using asymmetric loss function, and the penalty weight when the prediction result is short is higher than the penalty weight when it is long;According to predicted fermentation time and target total, calculate the target flow rate of koji powder;The space transmission delay between measurement position and drop position is obtained, and the delay is used to time align characteristic data and fermentation execution instruction;Based on the flow rate deviation of actual koji powder flow rate and koji powder target flow rate, closed-loop regulation is carried out to dynamically control the koji powder addition rate of fermentation mechanism.The present application effectively eliminates the dynamic pure lag influence of physical space transmission, and avoids the process risk caused by insufficient koji powder addition, realizes the prior prediction and whole-process closed-loop control of fermentation evenness.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology in Baijiu brewing, and in particular to a method and system for predicting and closed-loop controlling the uniformity of mash addition. Background Technology

[0002] Baijiu brewing is a complex process involving microbial growth, metabolism, and material transformation. After the mash is steamed and gelatinized, it needs to be spread out and cooled. Once cooled to a suitable temperature, koji powder is added as a saccharification and fermentation agent, followed by fermentation in a fermentation pit. The uniformity of koji addition is a core technological indicator affecting fermentation quality and alcohol yield. Uneven mixing of koji powder and mash will lead to inconsistent distribution of the micro-ecological environment and fermentation degree within the pit, directly affecting the final yield and flavor quality of the base liquor. With the advancement of automation in brewing processes, various automated koji addition and feeding equipment have been introduced into production sites. For example, CN221956061U discloses an automated baijiu brewing system whose spreading area includes a multi-stage continuous spreading machine and a koji addition mechanism to achieve automated long-line spreading and koji addition cooling. Another example is CN121005232A, which discloses a control method for intelligent brewing feeding equipment. This method obtains a target yield and controls the running speed of the conveyor belt and the rotation speed of the dispersing rollers based on this target yield, aiming to improve the efficiency and uniformity of spreading the fermentation raw materials.

[0003] However, despite improvements in the automation of underlying hardware, the existing koji-adding control logic still suffers from serious open-loop settings and equipment linkage defects in engineering practice. In actual brewing processes such as "steaming and spreading," the thickness of the mash layer is not linearly variable, but the existing segmented koji-adding control is essentially still an open-loop setting, meaning it only sets a fixed total amount of koji and a static koji-adding rate for each batch of mash. In actual operation, the koji-adding machine speed remains constant throughout the entire koji-adding process or is mechanically preset for a specific time period, completely lacking the dynamic response capability to real-time key process parameters such as the dynamic changes in mash volume and the distribution of mash layer thickness. More significantly, the control of existing koji-adding machines mostly uses isolated speed settings, failing to establish effective feedforward coordination with the speed changes of the front-end feeder. When the actual conveying rate of the mash fluctuates, the addition rate of koji powder cannot achieve synchronous adaptive matching. This isolated operation of the equipment further amplifies the process error, resulting in an inherent physical deviation in the uniformity of koji addition between the front and rear sections within the same steaming cycle.

[0004] Furthermore, due to the unique physical properties of the koji powder, the koji-adding process is highly susceptible to nonlinear disturbances. During transport, the koji powder is prone to absorbing moisture, clumping, and even arching, leading to a sudden drop in flow rate. Existing systems generally lack control strategies to address these material characteristics. Even though existing technologies (such as CN109656138A) have introduced intelligent algorithms like fuzzy PID in the temperature control of baijiu brewing cellars, directly applying conventional feedback control to koji-adding flow rate regulation will result in a continuous accumulation of deviation in the integral term once koji powder arching occurs, causing a sharp increase in the controller output. When the arching suddenly disappears, the excessive control output will inevitably cause the koji powder to "rush" out, resulting in a severe exceedance of the instantaneous flow rate. In addition, existing systems generally ignore the physical spatial transmission delay between the mash volume measurement device and the koji-adding machine's discharge point, failing to achieve time-domain alignment between sensor signals and actuator actions, further deteriorating the stability of flow rate control.

[0005] To address the uniformity issue, some current technologies employ visual inspection or laboratory sampling for evaluation. However, these approaches generally suffer from severe lag in their detection mechanisms. Existing online detection systems operate on a post-evaluation model of "adding koji first, then detecting." By the time the system detects substandard koji uniformity, the entire koji-addition process for that batch of mash is often already complete. This significant time lag renders the system obsolete, eliminating the physical conditions and time window for dynamic process control. Effective remedial measures are difficult to implement on-site, easily leading to incomplete saccharification and reduced alcohol yield due to uneven or insufficient koji addition. Therefore, there is an urgent need for a method that can accurately predict koji addition before it begins, overcome spatial time delays and nonlinear material disturbances during operation, and achieve multi-parameter feedforward and feedback composite closed-loop control.

[0006] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the present invention provides a method and system for predicting and controlling the uniformity of fermentation and koji addition, so as to solve at least some of the above-mentioned technical problems.

[0008] This invention discloses a method for predicting and closed-loop controlling the uniformity of fermented mash addition, which includes the following steps: Collect feature data during the fermentation and drying process of the fermented mash, and construct an input feature vector; The input feature vector is input into a pre-trained prediction model, and the output is the predicted curve duration. The prediction model is trained using an asymmetric loss function, which is configured to penalize the prediction result that is too short with a higher penalty weight than the prediction result that is too long. The target flow rate of the yeast powder is calculated based on the predicted yeast addition time and the target total amount of yeast added. The spatial transmission delay between the physical measurement location of the feature data acquisition and the bending and unloading location is used to time-align the acquired feature data with the bending execution command. The actual flow rate of the koji powder is collected, and closed-loop adjustment is performed based on the flow rate deviation between the actual flow rate and the target flow rate of the koji powder to dynamically control the koji powder addition rate of the koji adding mechanism.

[0009] Existing technologies typically employ a post-processing evaluation model in the koji-adding stage of baijiu brewing. By the time non-uniformity is detected, the process is already complete, making intervention difficult. This invention proposes a method for predicting and controlling the uniformity of koji addition in the mash. First, an input feature vector encompassing operating conditions and material parameters is constructed. Using a prediction model trained with an asymmetric loss function, a predicted koji-adding time with a safety margin is output before the koji-adding action, effectively mitigating process risks caused by insufficient koji powder. Second, spatial transmission delay compensation is introduced to eliminate control lag caused by physical distance. Finally, closed-loop adjustment is implemented based on the deviation between the target flow rate and the actual flow rate. This method organically combines proactive pre-prediction, precise spatiotemporal alignment, and real-time dynamic feedback, significantly improving the uniformity and system controllability of the koji-adding process.

[0010] According to a preferred embodiment, constructing the input feature vector includes: The moisture content of the fermented mash was obtained and numerical values ​​were assigned using an ordered dictionary encoding method to quantify the viscosity and flowability physical properties of the fermented mash. The real-time speed of the feeder and the assigned moisture content of the mash are weighted and combined to generate a combined feature code that integrates two-dimensional information of conveying rate and material characteristics. The combined feature code is then used as one of the dimensional elements of the input feature vector.

[0011] To address the quantification of the moisture content characteristic of fermented grains, an ordered dictionary encoding method is employed to assign values ​​to the moisture content. Compared to conventional one-heat encoding, this approach effectively avoids feature dimensionality expansion under high cardinality categories. Furthermore, by weighting and combining the moisture content—a physical property characterizing the viscosity and flowability of the fermented grains—with the real-time speed of the feeder, the final feature encoding comprehensively maps both the macroscopic conveying state and microscopic physical properties of the material. This provides the prediction model with a more representative input dimension, improving the model's adaptability to complex operating conditions.

[0012] According to a preferred embodiment, the calculation logic of the asymmetric loss function and the steps of outputting the predicted curve duration include: When the model predicts a duration greater than or equal to the actual curve duration, the asymmetric loss function assigns a first penalty weight; when the model predicts a duration less than the actual curve duration, the asymmetric loss function assigns a second penalty weight greater than the first penalty weight, in order to drive the prediction model to tend to make conservative predictions. The predictive model first outputs the effective cooling time of the mash through reasoning and calculation. The effective cooling time is then subtracted from the response time calibrated by the field equipment to obtain the final predicted curing time, in order to avoid the impact of the mechanical response lag of the equipment.

[0013] At the model training level, conventional loss functions tend to cause prediction results to fluctuate around the mean. This solution employs an asymmetric loss function, assigning a higher penalty weight when the predicted duration is too short. This weight bias guides the model to output a more conservative curing duration, thereby reducing the risk of incomplete saccharification at the process level. Furthermore, in the output stage, the predicted effective cooling time is subtracted from the equipment's physical response time, eliminating time errors caused by mechanical inertia during the start-up and shutdown phases of the curing machine. This ensures the accurate implementation and sufficient execution of the prediction algorithm on actual physical equipment.

[0014] According to a preferred embodiment, time alignment of acquired feature data with curve-adding execution instructions using spatial transmission delay includes: The compensation time is dynamically calculated based on the physical chain plate length from the physical measurement position to the material feeding and unloading position, and the real-time rotation speed of the spreading chain plate. When executing control commands at the current moment, the feature data collected at the historical moment corresponding to the current moment minus the compensation time is called to eliminate the dynamic pure lag caused by physical space transmission.

[0015] There is an objective physical distance between the location for collecting characteristics of the fermented mash at the spreading and drying machine and the location for adding koji and discharging the material. This solution dynamically calculates the time compensation amount based on the physical chain plate length and real-time chain plate speed, enabling the control system to accurately retrieve the real characteristic data of the fermented mash arriving at the discharging point at historical moments when generating the current control command. This mechanism eliminates spatial misalignment interference during continuous material conveying, ensuring precise spatiotemporal matching between the amount of koji powder fed and the corresponding amount of fermented mash in a local area.

[0016] According to a preferred embodiment, closed-loop regulation is performed based on the flow rate deviation between the actual koji powder flow rate and the target koji powder flow rate, including: A nonlinear proportional-integral-derivative controller employing an integral separation strategy for the flow velocity deviation input; The relationship between the absolute value of the flow velocity deviation and the set integral separation threshold is determined. When the absolute value is greater than the integral separation threshold, the integral action is cut off. When the absolute value is less than or equal to the integral separation threshold, the integral action is activated to eliminate the steady-state error. The flow rate adjustment is output based on the calculation results of the nonlinear proportional-integral-derivative controller to drive the operation of the bending mechanism.

[0017] In actual conveying, koji powder is prone to moisture absorption and clumping, causing sudden changes in flow rate. If conventional proportional-integral-derivative (PID) control is used, the continuous accumulation of the integral term during abnormal flow rates can easily lead to control saturation, resulting in material surge after clumping is eliminated. This solution introduces an integral separation strategy: when the absolute value of the flow rate deviation exceeds a set threshold, the integral action is paused, relying solely on the proportional and derivative components to quickly respond to the deviation; once the flow rate deviation converges within the threshold range, the integral action is resumed to eliminate steady-state error. This mechanism effectively suppresses severe system oscillations caused by nonlinear physical disturbances, enhancing the stability of flow rate regulation.

[0018] According to a preferred embodiment, the method further includes an online correction step to compensate for accumulated errors in the curvature process: The real-time cumulative amount of koji added is obtained by integrating the actual koji powder flow rate, and the cumulative deviation between the real-time cumulative amount of koji added and the target amount of koji added is calculated. When the cumulative deviation is determined to be greater than the set deviation threshold, the set value is dynamically calculated and corrected based on the remaining curvature time; The generated corrected setpoint is fed back along the control loop, replacing the original target flow rate of the powder and updating it as the input setpoint of the nonlinear proportional-integral-derivative controller for subsequent time periods.

[0019] While transient closed-loop flow rate regulation can mitigate short-term disturbances, it struggles to eliminate cumulative deviations within a single operation cycle. This solution integrates the actual flow rate to obtain the real-time cumulative curing amount for the steaming operation. When the cumulative deviation from the target curing amount exceeds the limit, the setpoint is dynamically recalculated and corrected using the remaining curing time. This control mechanism, which smoothly distributes historical cumulative errors to subsequent operation periods, balances the stability of instantaneous flow rate with the conservation of the total macroscopic curing amount, ensuring the overall physical accuracy of a single batch of curing operations.

[0020] According to a preferred embodiment, the method further includes a composite decoupling step that combines coordinated adjustment of the feeder speed: In response to the change signal of the feeder conveyor belt speed, the material level height of the mash on the spreading chain plate is collected, and the feedforward flow velocity reference is calculated in real time based on the real-time speed of the feeder, the real-time bulk density of the mash, and the effective cross-sectional area of ​​the mash calculated from the material level height and the width of the chain plate. The feedforward velocity reference is superimposed and fused with the velocity adjustment output of the nonlinear proportional-integral-derivative controller to generate the final velocity setpoint, so as to perform composite decoupled control on the bending mechanism.

[0021] Existing koji-adding and feeding equipment mostly have independent speed settings, making it difficult for the koji-adding machine to adaptively adjust when the thickness of the mash layer and the conveying rate fluctuate. This solution constructs a feedforward flow velocity benchmark based on the real-time speed of the feeder, the real-time bulk density of the mash, and the effective cross-sectional area, thus physically reproducing the real-time mass flow rate of the mash. By superimposing feedforward coarse adjustment with deviation-based feedback fine adjustment, the koji-adding mechanism can respond in advance to fluctuations in the mass flow rate of the upstream material, effectively overcoming the koji-adding imbalance caused by uneven material layer distribution under traditional speed-following control.

[0022] According to a preferred embodiment, the method further includes adaptive iterative updates to the prediction model: After a single batch of koji addition is completed, the fermentation process monitoring data and the results of the cellar exit test are used as lagging indicators to reverse-label the koji addition time for historical data. Based on preset rules, production data samples containing outliers are automatically removed, and samples that are confirmed to meet the uniformity standard are added to the training set as positive samples. The prediction model is then incrementally learned or retrained according to a preset period.

[0023] Considering the difficulty in obtaining real-time online data on the uniformity of koji addition, this solution incorporates relatively delayed production testing results, such as monitoring fermentation temperature upon entering the fermentation pit and measuring physicochemical indicators upon exiting the pit, to reverse-label the koji addition time of previously collected samples. Combined with filtering and cleaning rules for outlier data, the system can periodically incrementally learn the prediction model using valid samples validated over a long period of time. This adaptive iterative mechanism reduces the model's dependence on the initial training set, enabling it to dynamically adjust to seasonal environmental temperature changes and batch variations in raw grains, thus improving the long-term robustness of the prediction model in industrial settings.

[0024] Furthermore, compared with existing technologies that only disclose rough matching by adjusting the conveying speed through frequency converters (such as CN103087870A), this invention overcomes the objective process defects of ignoring the dynamic pure lag of long-distance conveying and the easy absorption of moisture and agglomeration of koji powder. This invention adopts an adaptive closed-loop architecture that combines spatial-temporal alignment with a nonlinear integral separation strategy: under steady-state continuous conveying conditions, the control system performs precise timestamp alignment matching based on the physical chain plate calibration length L1; under the condition that the koji powder causes transient arching due to moisture absorption, resulting in a sudden drop in actual flow rate, the control system detects that the absolute value of the flow rate deviation exceeds the integral separation threshold, instantly cuts off the integral action, and relies only on the proportional derivative term for high-frequency rapid drive correction; when the arching is eliminated and the flow rate deviation converges back to the threshold range, the system re-activates the integral action to eliminate steady-state static error. This system design not only eliminates the dynamic pure lag blind zone caused by long-distance transmission in physical space, but also effectively avoids the risk of severe system oscillation and material runaway caused by nonlinear physical disturbances of powder, achieving dynamic adaptive flexible control that cannot be achieved by simple speed ratio matching.

[0025] This invention also discloses a system for predicting and controlling the uniformity of fermented mash addition, comprising: A data acquisition network, deployed along the fermented mash conveying link, is used to collect characteristic data during the fermented mash spreading process and construct input feature vectors; The control system is connected to a data acquisition network. It has a pre-trained prediction model built-in and is configured to: input the input feature vector into the prediction model and output the predicted curvature duration; the prediction model is trained using an asymmetric loss function, which is configured to penalize a shorter prediction result with a higher penalty weight than a longer prediction result; and the target curvature flow rate is calculated based on the predicted curvature duration and the target curvature total amount. The flow rate measurement actuator is located at the koji feeding end and is connected to the control system to collect the actual koji powder flow rate and perform the koji feeding action; The control system is also configured to: acquire the spatial transmission delay between the physical measurement position of the data acquisition network and the material feeding position of the flow rate measurement actuator; use the spatial transmission delay to time-align the acquired feature data with the material feeding execution command; and perform closed-loop adjustment based on the flow rate deviation between the actual material flow rate and the target material flow rate to dynamically control the material feeding rate of the flow rate measurement actuator.

[0026] The system provided by this invention achieves collaborative design in hardware layout and logical architecture. It acquires multi-dimensional operating condition data through a deployed field data acquisition network, utilizes the predictive model built into the control system to achieve safe prediction and target flow velocity conversion based on asymmetric losses, and relies on a flow velocity measurement actuator to complete the closed-loop feeding action. This system architecture incorporates an alignment module for spatial transmission delay and a dynamic adjustment module for flow velocity deviation at the control layer, changing the traditional single-machine open-loop mechanical mode of koji-adding equipment. It provides an integrated control platform for the koji-adding process with pre-prediction, process disturbance immunity, and precise feedback capabilities.

[0027] According to a preferred embodiment, the flow velocity measurement actuator includes a weightless weighing system and a variable frequency drive device. The loss-in-weight weighing system includes a powder storage bin and a feeding mechanism, and the powder storage bin and the entire feeding mechanism are configured as a single weighing unit. The control system is further configured to acquire the weight signal of the entire scale body at high frequency when the material flows out, calculate the weight reduction per unit time, and filter to obtain the actual flow rate of the koji powder.

[0028] Conventional koji powder flow rate detection devices are susceptible to material bridging, adhesion to walls, and localized density unevenness, leading to distortion of instantaneous flow rate signals. This solution integrates the koji powder storage silo and its associated feeding mechanism into a loss-in-weight weighing system, obtaining the microscopic instantaneous koji powder flow rate from the macroscopic overall mass decay rate. This integrated weighing system configuration reduces the direct interference of internal material flow pattern changes on flow rate detection accuracy. Combined with high-frequency acquisition and filtering of the weight signal, it can output a high signal-to-noise ratio actual flow rate feedback value, providing more reliable underlying data support for accurate calculations by the upstream controller. Attached Figure Description

[0029] Figure 1 This is a main flowchart of a method for predicting and controlling the uniformity of fermented mash with added yeast, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the spatial layout structure of the physical equipment for the spreading and curing process provided in an embodiment of the present invention; Figure 3 This is a diagram showing the field hardware environment and sensor communication architecture topology provided in an embodiment of the present invention. Figure 4 This is a diagram illustrating the core closed-loop control logic and functional module architecture provided in the embodiments of the present invention. Figure 5 The flowchart illustrates the training and adaptive incremental learning process of the curve duration prediction model provided in this embodiment of the invention.

[0030] List of reference numerals 101: Spreading machine; 102: Volume detection unit; 103: Feeding chain speed detection unit; 104: First material level status detection unit; 105: Spreading machine drive motor; 106: Spreading chain speed detection unit; 107: Spreading and dispersing device; 108: Spreading bed; 109: Fan unit; 110: Distillate moisture detection unit; 111: Second material level status detection unit; 112: Spreading bed drive motor; 113: Adding yeast quantitative feeder; 114: Feeding monitoring unit; 115: Mixing machine; 116: Mother mash receiving hopper. Detailed Implementation

[0031] The following is a detailed explanation with reference to the accompanying drawings.

[0032] Combination Figure 1 The main flowchart of the method shown is as follows: Figure 4 The core control logic framework shown in this embodiment provides a method for predicting and controlling the uniformity of koji addition in fermented mash based on a gradient boosting tree. Addressing the shortcomings of existing open-loop setting modes in spreading and koji addition systems, which lack responsiveness and cannot achieve process control through post-event detection, this embodiment constructs a closed-loop control strategy that sequentially includes control links such as on-site status monitoring, koji addition duration prediction, flow rate dynamic adjustment, and equipment feedforward coordination, to achieve uniform mixing of koji powder and fermented mash. Figure 1 and Figure 4 As shown in the data flow and hierarchical progression, the method comprises six core steps: data acquisition and multi-dimensional feature construction, gradient boosting tree-based prediction of fermentation time, calculation of target koji powder flow rate, closed-loop control of the fermentation process, feedforward and feedback composite control of the feeder speed and the fermentation machine, and iterative model update based on new samples. More preferably, the method includes the following steps: Collect feature data during the fermentation and drying process of the fermented mash, and construct an input feature vector; The input feature vector is input into a pre-trained prediction model, and the output is the predicted curve duration. The prediction model is trained using an asymmetric loss function, which is configured to penalize the prediction result that is too short with a higher penalty weight than the prediction result that is too long. The target flow rate of the yeast powder is calculated based on the predicted yeast addition time and the target total amount of yeast added. The spatial transmission delay between the acquired feature data and the material addition / discharge position is used to time-align the acquired feature data with the material addition execution command. The actual flow rate of the koji powder is collected, and closed-loop adjustment is performed based on the flow rate deviation between the actual flow rate and the target flow rate of the koji powder to dynamically control the koji powder addition rate of the koji adding mechanism.

[0033] To implement the above method, this embodiment provides a highly integrated field physical hardware architecture. Combined with... Figure 2The diagram shows the spatial layout of the field physics equipment. The "upstream / downstream" designation is based on the main material transport direction, with upstream being near the inlet and downstream near the outlet. Along the material transport direction, the mash spreader 101, driven by the spreader drive motor 105, performs a front-end feeding action. Its transport status is monitored in real-time by the feeding chain speed detection unit 103 and the first material level detection unit 104. A volume detection unit 102 is deployed above the spreader 101. Downstream of the spreader 101 is a spreading bed 108, driven by a spreading bed drive motor 112. The chain speed of the spreading bed is collected by the spreading chain speed detection unit 106. A spreading and dispersing device 107 is configured above the conveying surface of the spreading bed 108, and fan units 109 are distributed along its lower edge to perform mash cooling operations. On the downstream outlet side of the spreading bed 108, a mash moisture detection unit 110 and a second material level detection unit 111 are installed. A koji-adding quantitative feeder 113 is mounted directly above its discharge end. Feed monitoring units 114 are arranged on both sides of the koji-adding quantitative feeder 113 to form a loss-in-weight weighing structure. Below the discharge end, a mixing machine 115 and a mother mash receiving hopper 116 are vertically mounted in sequence. Based on this objective spatial assembly relationship, a horizontal projection distance is formed between the installation centerline of the volume detection unit 102 and the discharge centerline of the koji-adding quantitative feeder 113. This distance defines the physical chain plate length for subsequent dynamic time delay compensation. .

[0034] Based on the above spatial layout, combined with Figure 3 The illustrated on-site hardware environment and sensor architecture topology diagram shows that this embodiment deploys multiple sensors and electrical control links with actuators on relevant equipment to collect on-site data and execute control commands. Specifically, the aforementioned first material level status detection unit 104 and second material level status detection unit 111 both use inductive sensors to identify the start and end positions of adding yeast to each batch of mash; the speed sensor is an encoder-type speed sensor, installed on the shaft end of the chain drive motor, used to collect the real-time speed of the spreading chain; the mash weighing stage uses a pressure-type weighing sensor with a range of 0-2000 kg and a system accuracy of ±0.5%, installed below the hopper for real-time monitoring of mash weight. Combined with... Figure 1 The hardware topology reveals that the mash weighing system specifically includes a sorghum weighing system, a rice husk weighing system, and a mother mash weighing system, each connected to an independent weighing sensor. A laser level gauge is installed above the still as a level sensor, with a measurement range of 0–5m and an accuracy of ±5mm, used to measure the height of the material in the still. A mash moisture detection device and a mash volume detection device connected via TCP / IP protocol are also deployed at the front end of the site.

[0035] To detect the flow rate of koji powder with high frequency and high precision, the koji powder storage silo and the entire feeding mechanism are integrated into a single weighing unit, forming a loss-in-weight weighing system (loss-in-weight scale). The load cell of this system is installed at the bottom of the koji powder storage silo, and the weighing instrument with loss-in-weight function is installed inside the control cabinet. Its range is 0–500 kg, and the system accuracy reaches ±0.2%. During the feeding process, as the material flows out of the loss-in-weight scale, the control system acquires the weight signal from the scale at high speed and high frequency, calculates the weight reduction of the material in the weighing hopper per unit time, and obtains the actual koji powder flow rate after filtering and optimization. .like Figure 2 As shown in the physical spatial layout, the vertical physical drop L2 from the installation position of the weighing sensor to the material drop point of the feeder is 0.5m.

[0036] At the control and execution end, the system is equipped with a PLC controller (preferably Siemens S7-1500 series) and multiple frequency converters (preferably Siemens G120 series). The PLC controller's internal modules are divided into digital input channels, high-speed counting channels, analog input channels, communication modules, and digital output channels. Specifically, the digital input channels receive signals for feeding trolley positioning detection, feeding conveyor presence / absence detection, and yeast addition presence / absence detection; the high-speed counting channels receive high-frequency pulse signals from the encoders of the spreading conveyor chain speed and the feeding trolley conveyor speed; the analog input channels receive signals from the sorghum / rice husk / mother mash weighing system and the mash moisture detection signal via 4-20mA signals; and the digital output channels connect to and control the spreading and leveling device and the yeast mixing device. Each frequency converter drives the spreading conveyor chain motor, the feeding trolley conveyor chain motor, and the yeast feeder motor, respectively. At the communication protocol level, the PLC controller's communication module connects to the aforementioned frequency converters, powder weighing system, and HMI via PROFINET industrial Ethernet; the industrial control server and the switch, and the switch and the PLC controller's communication module are connected via TCP / IP protocol, thereby establishing a complete field signal interaction and control execution link.

[0037] In the feature fusion and target prediction stage, the data acquisition layer acquires signals from the sorghum weighing system, rice husk weighing system, and mother mash weighing system, and simultaneously acquires the moisture content of the fermented mash, the volume of the fermented mash exiting the fermentation chamber, the speed of the feeding trolley conveyor belt, the speed of the spreading conveyor belt, and the mash source type code as ordered encoding input. The above basic data is unidirectionally input to the feature fusion layer for multi-dimensional feature vector construction, which includes a combined encoding based on the feeder speed and mash source type. The feature fusion layer outputs the fused feature vectors to the curve duration prediction layer, which incorporates a gradient boosting tree prediction model. The predicted duration of the curve is output through reasoning and calculation. Subsequently, the target velocity conversion layer performs the conversion formula based on the predicted curvature duration. The target powder flow rate is calculated and then input into the PID closed-loop control layer.

[0038] In the flow rate regulation and closed-loop execution phase, the feedforward coordination layer receives the speed change signal of the feeder conveyor belt and performs feedforward compensation calculations. The calculated feedforward compensation is then directly injected into the PID closed-loop control layer. The PID closed-loop control layer incorporates a nonlinear PID controller employing an integral control strategy, whose feedforward execution deviation calculation... Assuming the deviation is within normal limits, the nonlinear PID controller outputs a speed adjustment. At the execution layer, the inverter of the koji-adding machine, the motor of the koji-adding machine, and the koji-adding screw feeder are driven sequentially. The physical movement of the koji-adding screw feeder causes a weight change in the loss-in-weight weighing system of the koji powder temporary storage bin. The real-time feedback measurement layer collects and outputs the actual flow rate based on this data. Returning to the PID closed-loop control layer, thus forming the bottom-level flow velocity loop.

[0039] In the cumulative compensation and model update stages, the real-time feedback measurement layer synchronously calculates the real-time cumulative curve addition. It is then independently output to the online correction layer for cumulative quantities. The online correction layer for cumulative quantities internally executes cumulative deviation detection logic, the trigger condition of which is set as follows: When the cumulative deviation is determined to exceed the limit, the system triggers the correction formula. A corrected setpoint is generated and input as the corrected target flow rate into the PID closed-loop control layer to replace the original target koji powder flow rate. In addition, a feedback mechanism is triggered after a single batch of koji addition is completed, transmitting the actual operating characteristics and hysteresis detection results back to the koji addition duration prediction layer to update the internal parameters of the prediction model.

[0040] In the feature fusion layer, the control unit performs multi-dimensional feature vector construction to form the input feature vector. .in, To adjust the speed of the chain plate for spreading out, The total volume of the fermented mash. This refers to the total weight of the fermented mash. A combined encoding method was used to determine the relationship between the feeding trolley speed and the moisture content of the mash. In the brewing process of strong-aroma baijiu, the moisture content of the mash entering the fermentation pit is controlled between 52% and 56%. To quantify the characteristics of the mash material, an ordered dictionary encoding method was used to assign values ​​to the moisture content. Specifically, the source type encoding values ​​were assigned as 1, 2, 3, 4, and 5 respectively, based on the moisture content from low to high: 52%, 53%, 54%, 55%, and 56%. Based on this setting, mash with higher moisture content (e.g., 56%) corresponds to physical characteristics such as higher viscosity, stronger adhesion of koji powder, and poor flowability, and is given a higher feature weight in the model; mash with lower moisture content (e.g., 52%) corresponds to physical characteristics such as higher looseness, easier penetration of koji powder, and weaker adhesion, and is given a lower feature weight. This encoding method retains the ordered physical meaning of the mash type while avoiding the dimensionality explosion problem caused by one-hot encoding in high-cardinality category scenarios. Combined Feature Encoding The construction formula is set as ,in, For the real-time speed of the feeding trolley, For the ordered encoded values ​​of the aforementioned fermented mash type, the normalization coefficients are respectively taken as follows: and The selection of this normalization coefficient ensures that the contributions of the two indicators, velocity and moisture content, are of roughly the same magnitude, thereby integrating the two-dimensional information of the mash conveying rate and the mash material characteristics.

[0041] After multidimensional feature construction, the input feature vector The input is fed into the curve duration prediction layer. This prediction layer uses a gradient boosting tree as the basic prediction model, and its prediction expression is set as follows: ,in For the number of decision trees, For the first The prediction function of the decision tree. The prediction model uses asymmetric mean squared error as the loss function during offline training, and its loss function formula is defined as follows: In the baijiu brewing process, an insufficient amount of yeast added due to a predicted shorter yeast addition time directly leads to incomplete saccharification and a reduced alcohol yield. The resulting technological harm is significantly greater than the slightly excessive amount of yeast added due to a predicted longer yeast addition time. Based on this assessment of the asymmetric harm, an asymmetric parameter is introduced into the loss function formula. The penalty weight allocation mechanism is configured as follows: when the prediction duration is too long... At that time, weight When the prediction period is too short, At that time, weight In mathematical calculations, when When this happens, the loss function degenerates into the standard mean square error; by setting... This causes the model to be penalized with a larger weight when the predicted time is too short. This asymmetric weight penalty distribution drives the model to tend to make conservative predictions during the online inference phase. The model first outputs the effective cooling time of the fermented mash after inference calculation. And then through the formula The final predicted duration of the added curve was calculated. .in, The response time required for the koji-adding machine to reach a stable flow rate from startup is calibrated based on the physical characteristics of the equipment on-site, and is typically set to 2–5 seconds. The above compensation calculation effectively mitigates the risk of decreased alcohol yield caused by insufficient koji powder addition.

[0042] The target flow rate conversion layer is based on the predicted curvature duration. Total amount of koji added as set by the process According to the formula The target flow rate of the powder was calculated. This calculated value serves as the basis for setting the rotational speed of the koji-adding machine. Since the mash volume measuring device is installed at the front end of the spreading chain plate, while the koji-adding machine is installed at the rear end, there is a physical spatial transmission delay in the transport of the mash from the front to the rear. To address this spatial delay, a dynamic time delay compensation mechanism based on the chain plate speed is introduced, and its compensation time calculation formula is as follows: ,in The physical chain length from the measuring device to the material drop point. This refers to the real-time rotational speed of the drying chain. The control system executes alignment logic when acquiring feature values. That is, the adding machine at the current moment When executing control commands, the function called is... Historical sensor data is collected continuously. The aforementioned compensation mechanism ensures that feature acquisition and control execution are timestamped, eliminating the impact of dynamic time delay caused by physical space transmission on control accuracy.

[0043] In the flow rate regulation stage, the upstream node of the PID closed-loop regulation layer calculates the flow rate deviation between the actual powder flow rate and the target flow rate. This deviation is then input into a nonlinear PID controller. This controller employs an integral separation strategy, and its control law equation is: .in, , , These are the proportional, integral, and differential coefficients, respectively. This is the integral separation coefficient. When the absolute value of the flow velocity deviation... When triggered Thus cutting off the integral action; when When triggered To eliminate steady-state error by enabling integral action, the integral separation threshold is optimized. In actual conveying conditions, koji powder is highly susceptible to moisture absorption, clumping, and arching, leading to a sudden drop in flow rate. This integral separation strategy temporarily cuts off integration when the deviation is too large, preventing the integral term from continuously accumulating to saturation, and resumes integration when the deviation falls back below the threshold. This effectively prevents the material rushing phenomenon caused by a sharp increase in output after the arching is eliminated, and controls the deviation between the actual koji powder flow rate and the target flow rate within a certain range. Within this range. When the nonlinear PID controller determines the deviation to be normal, it outputs a speed regulation amount. Drive the frequency converter to perform adjustment actions.

[0044] To compensate for the accumulated error during the curve addition process, the cumulative amount online correction layer receives the real-time cumulative curve addition amount from the real-time feedback measurement layer. This value is obtained through the integral formula. The calculation is as follows. The cumulative amount undergoes cumulative deviation detection within the online correction layer. When the cumulative amount is determined... With target amount of curvature The deviation satisfies When this occurs, an online correction command is triggered. The system dynamically adjusts the flow rate setpoint based on the remaining curing time. The formula for calculating the correction setpoint is as follows: ,in This represents the time elapsed since the addition of the song until the current moment. This is the remaining curve addition time. After generating the correction setting, this correction value... Feedback is sent along the control loop to the PID closed-loop regulation layer, replacing the original target flow rate. This serves as the PID setpoint for subsequent time periods. Through the aforementioned flow rate recalculation and target value overwrite logic, the total amount of yeast added in the entire steamer is ensured to be consistent with the target value, thus controlling the overall yeast addition deviation within a specified range. Within the range.

[0045] In the coordinated adjustment of the koji powder flow rate in the koji feeder and the feeder speed, the feedforward coordination layer receives the speed change signal of the feeder conveyor belt and performs feedforward compensation calculations. When the feeder speed changes, the control system follows the feedforward compensation formula. Synchronously adjust the set value of the koji powder flow rate of the koji adding machine. Among them, This refers to the real-time speed of the feeder. The bulk density of the fermented mash is determined by the total weight of the fermented mash. With total volume The ratio is calculated in real time, that is ; The effective cross-sectional area of ​​the fermented mash on the spreading chain plate is determined by the height of the upper steamer. With the width of the chain plate The product is calculated as follows: ; The feedforward scaling factor, calibrated from historical data, has a set value range of [value range missing]. The establishment of this feedforward control path ensures that the rate of addition of koji powder is correlated with the mass flow rate of the fermented mash (i.e., the flow rate of the mash). The amount of yeast powder added per unit mass of mash is strictly proportional to the amount of yeast powder added, thus physically ensuring that the amount of yeast powder added per unit mass of mash remains constant.

[0046] To address the issue that the thickness of the mash layer in actual brewing processes, such as "steam exploration and gentle spreading," does not change linearly, the system, based on the aforementioned feedforward control, superimposes a feedback correction path based on the actual flow rate deviation. The calculation logic for the final flow rate setpoint is as follows: ,in This refers to the adjustment value output by the nonlinear PID controller. By combining a composite control strategy of feedforward coarse adjustment and feedback fine adjustment, the koji-adding machine not only follows the speed changes of the feeder but also decouples the changes in the mash level height and bulk density parameters. This setting ensures real-time matching between the koji powder addition rate and the actual conveying mass flow rate of the mash, solving the engineering problem of deviations in the uniformity of koji addition between the front and rear sections within a single steaming cycle.

[0047] Furthermore, addressing the nonlinear flow characteristics caused by material slippage and local redistribution that are unavoidable in actual production, this embodiment provides a fault-tolerant optimization mechanism through nonlinear fusion of multidimensional composite features and a dynamic correction loop for the remaining curvature time. Specifically, when transient mechanical slippage occurs on the spreading chain plate or local redistribution of the mash is caused by the overturning of the spreading and dispersing device 107, the prediction model uses deep fusion to comprehensively reflect the combined feature encoding of macroscopic conveying rate and microscopic material characteristics. It can tolerate local random velocity fluctuations and output a safe feedforward flow rate benchmark. Simultaneously, as the initial high-temperature, high-humidity mash undergoes rapid changes in thermodynamic temperature and dynamic viscosity during long-distance transport, the cumulative amount is monitored in real-time with high precision by the online correction layer, tracking the integral cumulative addition of koji within a single operating cycle. When the cumulative deviation exceeds a set threshold, the system dynamically recalculates and corrects the setpoint based on the remaining koji addition time, and feeds this corrected setpoint back along the bottom-level control loop to replace the original target flow rate. Through the combined mapping of microscopic multidimensional feedforward features and the synergistic effect of the macroscopic remaining time cumulative deviation absolute correction loop, the theoretical error between the uniform idealization assumption and the actual thermodynamic evolution is eliminated from the control engineering perspective, ensuring the achievement of the total koji addition accuracy for a single batch under variable and nonlinear real brewing conditions.

[0048] Combination Figure 5 The model training and incremental learning process shown involves the control system triggering an adaptive iterative update mechanism after the batch curve addition operation is completed. To ensure the effectiveness of the training samples, the system first automatically identifies and removes dirty data from the production site using preset rules. Specific cleaning rules include: performing integrity checks and removing data with missing or excessive values ​​in the input feature vector. Samples with outliers within the specified range; perform operational stability verification, removing samples where equipment failure or human intervention occurred; perform uniformity label verification, retaining only those confirmed to meet uniformity standards (i.e., coefficient of variation) through online testing or laboratory sampling. (sample).

[0049] Since uniformity cannot be tested immediately after adding koji, the system uses lag indicators from the later stages of production to automatically label the historical data with the optimal koji addition time. In the reverse annotation logic, the standard deviation of the multiple temperature measurements taken when the product is placed in the fermentation pit after being fermented is calculated. ℃, and the temperature deviation at different locations inside the cellar 7 and 15 days after the cells were placed in the cellar. Fermentation process monitoring data at ℃ were used as supplementary labeling data. Upon completion of fermentation and removal from the fermentation pit, the residual sugar and alcohol content of each layer of mash were measured, and the coefficient of variation was recorded. The results of the fermentation test are used as the final basis for determining whether the uniformity of the curing process meets the standard. When the uniformity is confirmed to meet the standard, the system stores the input features of that batch and the actual curing time as positive samples in the database and adds them to the training set; when the uniformity does not meet the standard, the sample is not included in the training set or is used as a negative sample for the regularization constraint of the prediction model. The control system incrementally learns or retrains the gradient boosting tree model using the newly added samples on a daily or batch-by-batch basis, and outputs the optimized new model. This incremental learning mechanism enables the prediction model to adapt to batch variations in raw materials and fluctuations in process parameters caused by seasonal environmental changes, thus maintaining prediction accuracy.

[0050] Based on the cooling and fermentation process in a strong-aroma baijiu brewing workshop, this implementation method provides a complete example of actual production sequence application. When a batch of fermented mash begins to cool, the feeding trolley's material presence sensor detects... A trigger signal is generated, and the system synchronously collects the following characteristic data: the rotational speed of the spreading chain plate is... The total volume of the fermented mash is The total weight of the fermented mash is The average height of the upper steamer material level is The feeding trolley speed is The type of grain residue in this batch is grain residue with a moisture content controlled between 52% and 56%, corresponding to an ordered code value of 5. The control system calculates the combined code based on this. And thus construct the input feature vector. The physical chain plate calibrates the length between the front-end volume detection unit 102 and the material drop point of the feeder 113. The system calculates the transmission delay from the front-end measurement position to the material feeding and unloading position as follows: The control command executed by the accelerator at 14:12:00 is based on a feature vector aligned with the data collected by the front-end sensor at 14:08:30.

[0051] The aforementioned feature vectors Input a pre-trained gradient boosting tree model, the number of decision trees in which the model is set. Maximum depth is 6, learning rate is 0.1, and asymmetric parameters are used. The model predicts that the effective cooling time for the fermented mash is 15 minutes (the start and end times of the cooling process are...). ), and output the predicted duration of the added song. The target total amount of added yeast based on the process settings. The system calculates the target flow rate of the powder. The process of adding music is underway. Initially, the loss-in-weight weighing system collected the initial actual flow velocity. Approximately Nonlinear PID controller calculates flow velocity deviation. ,because The system enables the points system (i.e.) The PID controller (setting parameters) , , The speed of the mixing machine is dynamically adjusted based on this deviation. After approximately 30 seconds of closed-loop adjustment, the actual flow rate quickly stabilizes at... Within the range, the steady-state deviation is controlled within During the addition of yeast, when the feeder speed changes from... Down to At the same time, the feedforward control path calculates the bulk density of the mash in real time. and effective cross-sectional area (Among them, the height of the upper steamer material position) Chain plate width And calculate the feedforward velocity reference. The controller will combine the feedforward flow rate with the PID fine-tuning amount. The final set value is generated by superposition. When the music was added up to the 10th minute, the system monitored the cumulative amount of music added. , and the target cumulative amount The deviation is Because it did not exceed the absolute value The threshold was reached, therefore online correction was not triggered. The curve addition operation was performed at... Completed, final cumulative number of tracks added: , and the total target value The deviation is ,satisfy The batch of mash was tested approximately 45 days later, confirming that its uniformity met the standards (multiple sampling points). The system then uses the input feature The actual added duration of 12 minutes and the uniformity standard were added as positive samples to the training set, and the prediction model was incrementally learned within the daily cycle.

[0052] Without altering the core prediction and closed-loop control mechanisms of this implementation, feasible equivalent alternatives exist for some functional modules and algorithms in the above scheme. Regarding the core prediction model, in addition to gradient boosting trees, variants of gradient boosting frameworks such as XGBoost, LightGBM, and CatBoost can be used, or integrated algorithms such as random forests and extreme random trees can be replaced. The substitution condition is that the prediction mapping between multi-dimensional feature inputs and curve duration can still be achieved. Regarding the flow physical measurement hardware, the loss-in-weight weighing system can be replaced by a Coriolis mass flow meter, a percussion plate flow meter, or a screw weighing feeder. The equivalence criterion is the ability to maintain high-frequency, continuous, and high-precision acquisition of the flow velocity signal. For the time-delay compensation algorithm, the calculation mechanism based on chain plate speed can be replaced by a Smith predictor, internal model control (IMC), or a compensation algorithm based on model predictive control (MPC), all of which can achieve time-domain alignment between sensor signals and actuator actions. Furthermore, considering the nonlinear characteristics of koji powder, the nonlinear PID control strategy can be replaced by variable speed integral PID, fuzzy PID, adaptive PID, or sliding mode control. As long as it can suppress koji powder bridging and prevent material shoveling, it is an equivalent alternative structure of this invention.

[0053] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A method for predicting and closed-loop controlling the uniformity of fermented mash addition, characterized in that, Includes the following steps: Collect feature data during the fermentation and drying process of the fermented mash, and construct an input feature vector; The input feature vector is input into a pre-trained prediction model, which outputs the predicted curve duration; wherein the prediction model is trained using an asymmetric loss function, and the asymmetric loss function is configured such that the penalty weight for a prediction result that is too short is higher than the penalty weight for a prediction result that is too long. Based on the predicted addition time and the target addition amount, the target flow rate of the yeast powder is calculated; The spatial transmission delay between the physical measurement position of the feature data and the bending material dropping position is obtained, and the spatial transmission delay is used to time-align the collected feature data with the bending execution command; The actual flow rate of the koji powder is collected, and a closed-loop adjustment is performed based on the flow rate deviation between the actual flow rate and the target flow rate of the koji powder to dynamically control the koji powder addition rate of the koji adding mechanism.

2. The method according to claim 1, characterized in that, The construction of the input feature vector includes: The moisture content of the fermented mash was obtained and numerical values ​​were assigned using an ordered dictionary encoding method to quantify the viscosity and flowability physical properties of the fermented mash. The real-time speed of the feeder and the assigned moisture content of the mash are weighted and combined to generate a combined feature code that integrates two-dimensional information of conveying rate and material characteristics. The combined feature code is then used as one of the dimensional elements of the input feature vector.

3. The method according to claim 1 or 2, characterized in that, The calculation logic of the asymmetric loss function and the steps of predicting the curve duration in the output include: When the model predicts a duration greater than or equal to the actual curve duration, the asymmetric loss function assigns a first penalty weight; when the model predicts a duration less than the actual curve duration, the asymmetric loss function assigns a second penalty weight greater than the first penalty weight, so as to drive the prediction model to tend to make conservative predictions. The prediction model first outputs the effective cooling time of the mash through reasoning and calculation. The effective cooling time is then subtracted from the response time calibrated by the field equipment to obtain the final predicted cooling time, so as to avoid the impact of the mechanical response lag of the equipment.

4. The method according to any one of claims 1 to 3, characterized in that, The step of time-aligning the acquired feature data with the curve-adding execution command using the spatial transmission delay includes: The compensation time is dynamically calculated based on the physical chain plate length from the physical measurement position to the material feeding position and the real-time rotation speed of the spreading chain plate. When executing control commands at the current moment, the feature data collected at the historical moment corresponding to the current moment minus the compensation time is invoked to eliminate the dynamic pure lag caused by physical space transmission.

5. The method according to any one of claims 1 to 4, characterized in that, The closed-loop adjustment based on the flow rate deviation between the actual koji powder flow rate and the target koji powder flow rate includes: The flow velocity deviation input is adopted by a nonlinear proportional-integral-derivative controller with an integral separation strategy. The relationship between the absolute value of the flow velocity deviation and the set integral separation threshold is determined. When the absolute value is greater than the integral separation threshold, the integral action is cut off. When the absolute value is less than or equal to the integral separation threshold, the integral action is activated to eliminate steady-state error. The flow rate adjustment is output based on the calculation results of the nonlinear proportional-integral-derivative controller to drive the operation of the bending mechanism.

6. The method according to any one of claims 1 to 5, characterized in that, The method also includes an online correction step to compensate for accumulated errors during the curvature addition process: The real-time cumulative amount of koji added is obtained by integrating the actual koji powder flow rate, and the cumulative deviation between the real-time cumulative amount of koji added and the target amount of koji added is calculated. When the cumulative deviation is determined to be greater than the set deviation threshold, the set value is dynamically calculated and corrected based on the remaining curvature time; The generated corrected setpoint is fed back along the control loop, replacing the original target flow rate of the powder and updating it as the input setpoint of the nonlinear proportional-integral-derivative controller for subsequent time periods.

7. The method according to any one of claims 1 to 6, characterized in that, The method also includes a composite decoupling step that combines coordinated adjustment of the feeder speed: In response to the change signal of the conveyor belt speed of the feeder, the material level height of the mash on the spreading chain plate is collected, and the feedforward flow velocity reference is calculated in real time based on the real-time speed of the feeder, the real-time bulk density of the mash, and the effective cross-sectional area of ​​the mash calculated from the material level height and the width of the chain plate. The feedforward flow velocity reference is superimposed and fused with the flow velocity adjustment output by the nonlinear proportional-integral-derivative controller to generate the final flow velocity setpoint, so as to perform composite decoupled control on the bending mechanism.

8. The method according to any one of claims 1 to 7, characterized in that, The method also includes adaptive iterative updates to the prediction model: After a single batch of koji addition is completed, the fermentation process monitoring data and the results of the cellar exit test are used as lagging indicators to reverse-label the koji addition time for historical data. Production data samples containing outliers are automatically removed based on preset rules, and samples that are confirmed to meet the uniformity standard are added to the training set as positive samples. The prediction model is then incrementally learned or retrained according to a preset period.

9. A system for predicting and controlling the uniformity of fermented mash addition, characterized in that, include: A data acquisition network, deployed along the fermented mash conveying link, is used to collect characteristic data during the fermented mash spreading process and construct input feature vectors; The control system is communicatively connected to the data acquisition network. The control system has a pre-trained prediction model built-in. The control system is configured to: input the input feature vector into the prediction model and output the predicted curvature duration; wherein the prediction model is trained using an asymmetric loss function, and the asymmetric loss function is configured to penalize a shorter prediction result with a higher penalty weight than a longer prediction result; and calculate the target curvature powder flow rate based on the predicted curvature duration and the target curvature total amount. A flow rate measurement actuator is installed at the koji feeding end and is communicatively connected to the control system. It is used to collect the actual koji powder flow rate and perform the koji feeding action. The control system is further configured to: acquire the spatial transmission delay between the physical measurement position of the data acquisition network and the material feeding position of the flow rate measurement actuator; use the spatial transmission delay to time-align the acquired feature data with the material feeding execution command; and perform closed-loop adjustment based on the flow rate deviation between the actual material flow rate and the target material flow rate to dynamically control the material feeding rate of the flow rate measurement actuator.

10. The system according to claim 9, characterized in that, The flow velocity measurement actuator includes a weightless weighing system and a variable frequency drive device. The loss-in-weight weighing system includes a koji powder temporary storage bin and a feeding mechanism, and the koji powder temporary storage bin and the entire feeding mechanism are configured as an integral scale body; The control system is further configured to acquire the weight signal of the overall scale body at high frequency when the material flows out, calculate the weight reduction per unit time, and filter to obtain the actual powder flow rate.

Images