Food industrial production control method and system based on artificial intelligence

Abstract

The invention relates to the technical field of food industrial production, and discloses a food industrial production control method and system based on artificial intelligence, and the method comprises the steps: A, collecting the original attribute information of a to-be-processed raw material through a raw material data collection module before a food processing production line is started; b, in the food processing process, physical state parameters of the dough are obtained in real time; c, synchronously inputting the data obtained in the step A and the step B into a control system in which a multi-model fusion mechanism is constructed; d, the control system carries out dynamic recognition, difference comparison and trend evaluation on the current batch of raw materials and the process state, and process adjustment suggestions of corresponding stages are output based on feature weights learned in the training data; and E, converting the process adjustment suggestion into an instruction set, issuing the instruction set to a control executor, and integrating an advanced control method and system with intelligent prediction and adaptive control capabilities of multi-source information so as to realize the stabilization of the quality of the baked product.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of food industrial production, and particularly relates to a food industrial production control system based on artificial intelligence. BACKGROUND

[0002] Baked goods, such as bread, cakes, cookies, etc., are staple and leisure foods widely consumed globally. Its production process usually includes multiple key stages such as raw material mixing, kneading, fermentation, shaping, baking, etc. The final quality of the product, such as specific volume, internal structure (porosity), external color and luster, taste (hardness, resilience), etc., is influenced by a combination of factors such as raw material properties, process parameters, environmental conditions, and equipment state, etc.

[0003] Traditional baking production, especially in small-scale or artisanal workshops, highly relies on experienced bakers. They manually adjust parameters such as kneading time, fermentation temperature and humidity, baking temperature, etc. by relying on their experience and sensory judgment of the dough (such as touch, appearance, smell) and perception of the environment. Although this method can produce high-quality products, its production capacity is limited, and the stability and consistency of product quality are highly dependent on the skill level and state of the operator, making it difficult to achieve large-scale standardized production.

[0004] With the development of industrial automation technology, modern baking production lines generally use automated equipment and control systems, such as programmable logic controllers (PLC) or distributed control systems (DCS). These systems can automatically control key variables in the production process (such as the speed and time of the kneading machine, the temperature and humidity of the fermentation cabinet, the temperature and time of the baking oven) according to pre-set process recipes and parameter curves. This automated production method significantly improves production efficiency and scale.

[0005] However, existing automated control systems are mainly based on fixed process recipes and parameter settings, or use simple feedback control (such as PID control), making it difficult to effectively respond to various dynamic fluctuations and uncertainties in actual production processes. These fluctuations mainly come from: Raw material batch differences: Even if the same grade of flour is provided by the same supplier, there are small but significant differences in the physicochemical properties of different batches, such as protein content, wet gluten index, water absorption, enzyme activity, etc.; the initial activity and stability of yeast also vary with batch and storage conditions; the trace components or physical form of other auxiliary materials may also affect the dough characteristics. The inherent variability of these raw materials directly affects the formation, fermentation and baking characteristics of the dough.

[0006] Environmental fluctuations: The environmental temperature, humidity, air pressure, etc. in the production workshop will fluctuate with the season, weather changes or the running state of the air conditioning system. These environmental changes directly affect the fermentation speed, moisture balance of the dough, and the heat transfer efficiency during baking.

[0007] Equipment status changes: Production equipment wears out and ages over long periods of operation. For example, a worn kneader's stirring paddle can affect the energy and shear force applied to the dough. A proofer's temperature and humidity sensors or actuators may experience control accuracy drift. Aging heating elements or uneven thermal fields in ovens can affect baking results. These changes in equipment status can lead to deviations between actual process conditions and setpoints.

[0008] Cumulative effect during the process: The baking process is a continuous, multi-stage process. Minor deviations in the previous stage (such as insufficient or excessive kneading) may be amplified in the subsequent fermentation or baking stages, causing the dough state to deviate seriously from the ideal trajectory, ultimately affecting the quality of the finished product.

[0009] The combined influence of these uncontrollable factors causes the dough's state (such as viscosity, elasticity, extensibility, fermentation volume, and internal structure formation) to deviate from the ideal trajectory during key stages such as kneading, fermentation, and baking, ultimately causing fluctuations in key quality indicators such as specific volume, porosity, color, and taste of the finished product, affecting product consistency and stability, increasing scrap rate and production costs, and limiting the company's ability to continuously supply high-quality, standardized products.

[0010] Existing simple feedback control systems, such as those based on PID, typically only adjust a single process parameter (such as temperature or humidity) and are lagging, meaning corrections are only made after a parameter deviates. This control approach is unable to predict and compensate for the combined impact of complex factors such as raw material variation and environmental fluctuations on dough state and final quality. Furthermore, existing systems often lack in-depth understanding of the initial properties of raw materials, and their real-time perception of process status (especially the physical rheological properties and fermentation state of the dough) is limited and incomplete. Furthermore, they lack the ability to predict final quality based on real-time conditions and make proactive, adaptive adjustments. Summary of the Invention

[0011] The present invention provides an artificial intelligence-based food industrial production control system, which aims to solve the problems mentioned in the above-mentioned prior art.

[0012] To achieve the above object, the present invention provides the following technical solutions: A food industrial production control method based on artificial intelligence, the method comprising: Step A: Before the food processing production line is started, the raw material data collection module collects the original attribute information of the raw materials to be processed, including but not limited to the protein content, moisture content, yeast type and activity, additive ratio, initial humidity, and temperature data of the flour; Step B: During the food processing process, obtaining physical state parameters of the dough in real time, including any combination of at least one or more parameters: viscosity, elasticity, extensibility, surface rebound rate, specific volume, humidity, and temperature; Step C: synchronously inputting the data obtained in step A and step B into a control system having a multi-model fusion mechanism, wherein the control system includes a quality prediction model, a parameter correction model, and a control strategy selection module; Step D: The control system dynamically identifies, compares differences, and evaluates trends in the current batch of raw materials and processing status. Based on the feature weights learned from the training data, it outputs process adjustment suggestions for the corresponding stages, including control parameters for the kneading, fermentation, and baking stages. Step E: Convert the process adjustment suggestions into an instruction set and send it to the control actuator to dynamically adjust the target equipment to compensate for product quality fluctuations caused by changes in raw materials, environment, or equipment status; Step F: Obtain the quality indicators of the end product through the finished product quality acquisition module and send them back to the control system for adaptive optimization and model iteration.

[0013] Preferably, the acquisition of the physical state parameters of the dough in step B is a graded acquisition method, the first level is the basic parameter acquisition, including viscosity and temperature and humidity data; the second level is the supplementary feature acquisition, which is enabled when the model judges an abnormal trend, including elasticity and extensibility; the third level is visual recognition assisted acquisition, which evaluates surface wrinkles, bulging morphology and rebound behavior through an image recognition module to improve modeling accuracy.

[0014] Preferably, the control system is constructed based on an end-to-end deep neural network structure, specifically including: Long short-term memory neural network (LSTM) for time series modeling to capture periodic fluctuations in the process; Feedforward neural network model for quality prediction, predicting key quality indicators of finished products based on current status and raw material information; Reinforcement learning control agent uses actual control feedback data as a reward function to perform adaptive policy optimization and achieve closed-loop optimization.

[0015] Preferably, the process adjustment suggestions include: Speed, time and cycle segmented control during kneading stage; Temperature and humidity settings, fermentation duration, and ambient air flow rate during the fermentation stage; Layered temperature zone setting, heating timing curve, dehumidification window opening and wind speed adjustment during the baking stage; Steam content control and equipment motor power output range adjustment throughout the entire processing process.

[0016] Preferably, the method further comprises an abnormal triggering mechanism: When any parameter shows a deviation or mutation trend that exceeds the set range of the model's expected value, a multiple discrimination process is initiated, including parameter comparison, historical sample review, and image-assisted judgment. The offset trend is adjusted and corrected through the regression correction module, and a new parameter compensation prediction is output.

[0017] The present invention also discloses an artificial intelligence-based food industrial production control system, which includes: Raw material attribute collection module, used to obtain raw material formula composition and batch characteristics before production; Physical state detection module, used to collect dough viscosity, elasticity, extensibility, temperature and humidity parameters in real time during processing; Image-assisted recognition module, used to identify dough surface structure, bubble generation and rebound status; Control unit, used to integrate multi-source data and establish quality prediction, process optimization and control decision models; The control instruction generation and sending module is used to convert the control unit output results into device executable control commands; Actuators, including kneading equipment, fermentation chamber, baking furnace and their supporting drive units; The finished product quality monitoring module is used to collect the specific volume, porosity, color index, and rebound rate data of the final product in real time, and feed it back to the control unit to form a closed-loop learning mechanism.

[0018] Preferably, the control unit includes a model fusion engine, which is constructed as a multi-path control strategy output structure, including: First path: output static optimal control instructions based on current raw materials and process data; Second path: When there is a quality deviation or abnormal trend, a dynamic adjustment strategy is generated based on the reinforcement learning prediction results; The third path: Fuse image information to perform secondary confirmation on abnormal identification and correct the conflicting results output by the first and second paths.

[0019] Preferably, the image-assisted recognition module adopts a convolutional neural network (CNN) structure, and the training samples include high-resolution dough images and their corresponding physical test values. The model can predict viscosity trend values ​​and elasticity trend values ​​based on actual images, and compare them with physical test values, and correct or complete missing measurement data.

[0020] Preferably, the system is provided with a quality target correction mechanism based on feedback values ​​of multiple process stages. The mechanism analyzes the actual performance of the current batch of products in terms of target specific volume, structural uniformity, and color consistency indicators, and actively inputs error terms into the control unit, prompting it to dynamically correct weight values ​​and strategy structures.

[0021] As preferred, the actuator is connected with the control unit through an intermediate instruction interpretation layer, which converts the control unit output parameter instruction into PLC control language or specific motor control signal Technical effects and advantages of the present application: 1. The present application can effectively compensate for product quality fluctuations caused by changes in raw materials, environment or equipment state, thereby significantly improving the batch-to-batch and intra-batch consistency of key quality indicators of the end product, by real-time collection of multi-source data such as raw material properties, process physical state and image information, and dynamic identification, difference comparison and trend evaluation based on a control system with multi-model fusion mechanism, output of process adjustment suggestions and issuance of instructions for dynamic adjustment.

[0022] 2. The present application is provided with an abnormal triggering mechanism, which starts a multiple discrimination process and performs regression correction when any parameter deviates from the model expected value setting interval or shows a sudden trend, outputting a new parameter compensation prediction. This enables the system to actively identify and intelligently respond to unexpected fluctuations and abnormal situations in the production process, improving the robustness of the production process.

[0023] 3. The present application based on artificial intelligence control system for dynamic identification and evaluation of the current batch of raw materials and process state, output the corresponding stage (kneading, fermentation, baking) control parameter adjustment suggestion, and convert it into instruction set to target equipment for dynamic adjustment. This realizes the upgrading of traditional control method relying on experience or fixed parameters, makes the adjustment of process parameters more fine, intelligent and real-time.

[0024] 4. The present application integrates multi-source heterogeneous data such as raw material properties, process physical state (including hierarchical collection), image assisted identification, and synchronously inputs them into the control system. The comprehensive utilization of multi-dimensional data, combined with multi-model fusion mechanism, enables the control system to more comprehensively understand the production process, make more accurate state identification, quality prediction and control decision. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The present application is a system overall architecture schematic diagram; Figure 2 The control unit structure diagram of the present application; Figure 3 The system overall operation flow timing diagram of the present application. DETAILED DESCRIPTION

[0026] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0027] Traditional food baking production relies on preset fixed process recipes and parameter curves. However, the actual production environment is complex and changeable, with many influencing factors: Inherent variability of raw materials: Even for flour of the same grade supplied by the same supplier, there are still slight differences in physical and chemical indicators such as protein content, wet gluten index, water absorption rate, enzyme activity, and ash content between different batches. The initial activity and stability of yeast also vary with batches and storage conditions. The trace components or physical form (such as particle size) of other auxiliary ingredients (such as sugar, oil, salt, and improvers) may also affect dough properties.

[0028] Fluctuations in environmental factors: The ambient temperature, humidity, and air pressure in the production workshop will fluctuate with the seasons, weather changes, or the operating status of the air-conditioning system, directly affecting the fermentation speed and water loss of the dough.

[0029] Equipment status changes: Mechanical wear of the kneading machine will affect the energy and shear force applied to the dough; the temperature and humidity control system of the fermentation box may have local temperature differences or control accuracy drift; aging of the heating elements of the oven or uneven thermal field will affect the baking effect.

[0030] Cumulative effects during the process: Minor deviations in the kneading stage (such as insufficient gluten formation or excessive disruption) will be magnified during the fermentation stage, affecting the dough's air holding capacity and volume; abnormal fermentation conditions will directly lead to product collapse or rough structure during baking.

[0031] These uncontrollable factors lead to fluctuations in the final product's specific volume, internal structure (porosity), external color, and taste (hardness, springback), making it difficult to consistently produce high-quality, consistent products, increasing scrap rates and making quality control more difficult. This invention aims to provide a solution that can sense, intelligently predict, and dynamically compensate for these fluctuations in real time.

[0032] This invention proposes an artificial intelligence-based food industrial production control method, which is applied to bread production lines. Figure 1-Figure 3 , including the following steps: Step A: Collection of raw material attribute information This step is performed before a baking production batch is started. By deploying a raw material data acquisition module in the raw material receiving or preparation area, a comprehensive and quantitative characterization of all raw materials about to be put into production is performed.

[0033] Collection methods: Various techniques may be used, including but not limited to: Laboratory test data input: The analysis report provided by the supplier or the physical and chemical test results of flour, yeast, etc. conducted by the internal laboratory (such as protein content, wet gluten, ash content, moisture, yeast fermentation ability, etc.) are entered into the system through the data interface or manual entry.

[0034] Online / rapid detection sensors: Use near-infrared (NIR) spectrometers to quickly analyze flour components; use moisture meters and thermometers to obtain the initial temperature and humidity of raw materials; and use specialized yeast activity testers to assess the initial fermentation capacity of yeast batches.

[0035] Recipe management system integration: Get the precise formula composition of the current batch from the recipe management database, including the types and proportions of various excipients.

[0036] Data content: The collected data forms a structured raw material feature vector, including but not limited to: flour type, brand, batch number, protein content (%), moisture content (%), wet gluten index, ash content (%), water absorption rate (%), yeast type, batch number, initial fermentation activity (such as CO2 gas production rate or volume), the precise ratio of auxiliary materials such as sugar / oil / salt / improving agents (%), and the ambient temperature and humidity when the raw materials are stored.

[0037] Raw material attribute information forms the basis for subsequent control system decisions, helping the system pre-identify key variables that may affect product quality. Compared to traditional production methods that rely solely on empirically determined process parameters, this invention precisely captures raw material attributes, providing a data foundation for subsequent intelligent control.

[0038] Step B: Real-time acquisition of physical state parameters during processing (grading acquisition) This step is performed continuously throughout the dough processing stages (kneading, proofing, fermentation, shaping, etc.). Physical state detection modules deployed on the production line continuously and in real time acquire key physical state parameters of the dough. To optimize sensor deployment and data processing load, a hierarchical acquisition approach is adopted: Level 1 (basic parameter collection): This is carried out continuously throughout the production process and includes: Viscosity / Consistency: A torque sensor or power sensor installed on the kneading machine's mixing shaft indirectly measures the changes in the dough's resistance during kneading, reflecting the dough's viscosity and gluten formation. Later, the current / torque changes in the dough conveying or handling equipment can be monitored.

[0039] Temperature and Humidity: Use non-contact infrared temperature sensors and humidity sensors (or contact probes) to measure the dough surface and / or core temperature, as well as the ambient humidity. These parameters directly affect enzyme activity and fermentation rate.

[0040] Data frequency: High-frequency acquisition (e.g., multiple times per second) provides continuous process status information.

[0041] Second level (supplementary feature collection): When the control system determines that the dough state may show abnormal trends or requires more precise judgment based on the first level data, the system triggers the use of more advanced sensors to perform supplementary feature collection: Elasticity and extensibility can be measured using an online dough rheometer or a small tensile / compression tester at a specific workstation (e.g., before shaping). For example, a robotic arm can be controlled to stretch or compress the dough, and the force-displacement curve can be measured to extract parameters such as elastic modulus and elongation at break. Alternatively, non-contact methods such as laser displacement sensors can be used to measure the deformation and recovery of dough under specific external forces (e.g., airflow).

[0042] Surface rebound rate: A small pneumatic device is used to apply a brief, constant pressure pulse to the dough surface, and then a high-speed camera or laser displacement sensor is used to measure the speed and extent to which the surface recovers to its original shape.

[0043] Data frequency: triggered on demand, the acquisition frequency is lower than the first level, but it provides richer mechanical property information.

[0044] Level 3 (visual recognition-assisted acquisition): Combined with the image-assisted recognition module, high-resolution industrial cameras are used to capture images of the dough at key workstations (such as the end of kneading, in the middle of fermentation, at the end of fermentation, and after shaping).

[0045] Image Analysis: Images are analyzed using a pre-trained convolutional neural network (CNN) model to assess the dough’s surface texture (smoothness, wrinkles), bubble generation (size, distribution, density), overall morphology (collapse, swelling), and rebound behavior after touch (via continuous frame analysis).

[0046] Data usage: Image data provides non-contact macroscopic and microscopic state information for: Auxiliary judgment: Verify physical sensor data. For example, the image shows that the dough surface is smooth and tensile, which is consistent with the viscosity indicated by the torque sensor.

[0047] Complementary information: Provides features that are difficult to measure directly with physical sensors, such as bubble structure.

[0048] Correction / calibration: Use image features (such as the correlation between surface wrinkle level and viscosity) to predict viscosity or elasticity trend values ​​and compare them with physical measurements to calibrate sensors or provide alternative judgments when sensor data is abnormal.

[0049] Trigger mode: can be collected at fixed time intervals, or triggered by the control system when it determines that visual confirmation is required.

[0050] The hierarchical collection method not only improves the system's operating efficiency and reduces computing resource consumption, but also provides more comprehensive data support at critical moments and enhances the system's ability to identify abnormal situations.

[0051] Step C: Synchronous input of multi-source data and control system processing The initial raw material attribute data collected in step A and the physical state parameters of each dough grade acquired in real time in step B (including sensor data and image analysis results) are synchronously input into the control unit. The control unit is the core of the present invention and is equipped with an advanced multi-model fusion mechanism.

[0052] Data synchronization and preprocessing: All data streams are timestamped and synchronized. Data cleaning, standardization, feature engineering, and other preprocessing are performed to form input vectors in a unified format.

[0053] Control unit architecture: includes the following core components: Data fusion engine: Integrates data from different sensors and different modalities (numerical, image) to generate a comprehensive representation of the current dough state.

[0054] Quality Prediction Model: This model is built using a feedforward neural network (FNN) or deep learning architecture. Inputs include raw material characteristics, historical process parameters, and current and past physical parameters of the dough. The model output is a prediction of the key quality indicators that the final product will achieve given the current dough state and the current process.

[0055] Parameter Correction Model: This model is built based on a regression model or neural network. Inputs include the output of the quality prediction model, the set target quality indicators, and the current dough state. The model calculates the necessary adjustments to subsequent process parameters to bring the predicted finished product quality closer to the target value.

[0056] Control strategy selection module: Serves as the decision-making center, receives quality prediction results and parameter correction suggestions, and combines them with the output of the reinforcement learning (RL) agent to generate the final control instruction set.

[0057] Step D: Dynamic identification, evaluation and process adjustment suggestion output The control unit continuously and dynamically analyzes and processes the input real-time data stream.

[0058] State Identification and Discrepancy Comparison: The system compares the current batch's ingredient characteristics and real-time process state (dough parameters) with the normal production patterns learned from training data and historical successful batch data to identify if there is a deviation from the expected trajectory. For example, is the dough stickiness decreasing too fast, or is the fermentation volume growing slower than expected?

[0059] Trend Assessment: Based on time series models like LSTM, the system predicts the future development trend of the dough state and assesses the potential impact of the current deviation on subsequent stages and final quality.

[0060] Process Adjustment Suggestion Generation: Based on the learned influence weights of different ingredient characteristics and process parameter changes on final quality in the training data, the system calculates the optimal process adjustment suggestions. These suggestions are for key control parameters in the current and subsequent stages, such as: Kneading Stage: Suggest adjusting the speed of the kneading machine (RPM), kneading time, or the duration and speed of each stage of segmented kneading.

[0061] Fermentation Stage: Suggest adjusting the temperature and humidity settings of the fermentation cabinet, total fermentation time, or the flow rate of ambient air.

[0062] Baking Stage: Suggest adjusting the layered temperature zone settings of the oven (such as the top / bottom temperature of the oven chamber), heating timing curve (such as initial high temperature shaping, subsequent low temperature slow baking), and the opening and air speed of the exhaust window.

[0063] Full Process: Suggest adjusting the amount of steam injection, power output range of equipment motors (such as mixing, conveying).

[0064] Step E: Instruction Conversion and Equipment Dynamic Adjustment Convert the process adjustment suggestions output in Step D into a set of executable instructions for the equipment.

[0065] The instruction interpretation layer is a key intermediate layer. The control unit outputs high-level, abstract control suggestions (such as "increase fermentation energy" or "adjust the baking curve type"). The instruction interpretation layer translates these suggestions into low-level control signals specific to the equipment. For example, "increase fermentation energy" may be interpreted as "increase the fermentation cabinet temperature setpoint by 1.5°C and increase the humidity setpoint by 3%RH"; "kneading energy level = 0.7" may be interpreted as "set the kneading machine main motor torque target to 70% of the rated torque and maintain this torque until the dough stickiness reaches the threshold."

[0066] Instruction Issuance: The interpreted instructions are issued to the corresponding control executors (such as PLC, DCS, or independent motion controllers) through standard industrial communication protocols.

[0067] Equipment Adjustment: The actuator receives instructions and drives the corresponding equipment unit to make dynamic adjustments, such as changing motor speed, adjusting heater power, controlling valve opening, and adjusting fan speed. These adjustments are made in real time to compensate for detected fluctuations and guide the dough state back to the desired process trajectory.

[0068] Step F: Finished product quality collection and closed-loop optimization After production is completed, the key quality indicators of the end product are obtained through the finished product quality collection module deployed at the end of the production line.

[0069] Collection methods: Various detection equipment can be used, including but not limited to: Visual inspection system: Performs image analysis of the external shape, color (using Lab color space), and surface features (cracks, burnt spots) of finished bread / pastry.

[0070] Specific volume meter: measures the volume and weight of the product, calculates the specific volume (volume / weight), and reflects the degree of expansion.

[0071] Texture analyzer: performs compression and shear tests on products to measure taste-related indicators such as hardness, elasticity, chewiness, and rebound rate.

[0072] Internal structure analysis: Perform image analysis on sliced ​​products to evaluate the size, distribution, and uniformity of pores.

[0073] Data transmission and usage: The collected finished product quality data (associated with the batch ID) is transmitted back to the control unit.

[0074] Reinforcement Learning Rewards: Product quality indicators (such as the deviation of specific volume from the target value and the pore uniformity score) serve as reward signals for the reinforcement learning agent. Good product quality results in a high reward, while poor quality results in a low reward. This drives the RL agent to adjust its policy, learning control action sequences that yield higher rewards (i.e., better product quality) in different states.

[0075] Model retraining and optimization: Accumulated raw material data, process data, control system decision instructions, equipment execution records, and final product quality data constitute a complete production sample. These samples are used to trigger offline or online retraining and optimization of quality prediction models and parameter correction models, either periodically or when performance degrades, to improve the model's prediction accuracy and the generalization of control effects.

[0076] Quality Target Correction Mechanism: The system analyzes the error between the actual performance of the current batch of products and the set ideal targets for indicators such as target specific volume, structural uniformity, and color consistency. This error term is actively input into the control unit (for example, as part of the loss function or as an additional input feature), prompting the model to dynamically correct its internal weights and strategy structure, so that the control output more directly serves the actual quality goal, rather than simply stabilizing process parameters.

[0077] Abnormal trigger mechanism (preferred implementation): In order to enhance the robustness and security of the system, an abnormal trigger mechanism is set. When any acquisition parameter (raw material properties, process physical state, image feature analysis results) exceeds the preset threshold, deviates from the model expected value range, or shows a sudden change trend, the system immediately initiates a multiple discrimination process: Parameter comparison: Cross-compare the current value and rate of change of the abnormal parameter with the historical normal range and other relevant parameters of the same batch (for example, if the temperature is abnormal, check whether the heater power and ambient temperature are also abnormal).

[0078] Historical sample review: Automatically search the historical database to find the production records, control system strategies and final product results when similar parameter anomalies occurred in the past as a basis for judgment.

[0079] Image-assisted judgment: If the anomaly is related to the dough state, the image-assisted recognition module is forcibly triggered to collect and analyze data, and the nature and severity of the anomaly are assisted by visual information (such as whether the dough has collapsed or whether the surface is dry).

[0080] Regression Correction and Compensation Prediction: Based on historical data and current multi-source information, a dedicated regression correction module or anomaly handling sub-model models and adjusts for deviation trends. The system predicts how the dough state will evolve if the current anomaly persists or develops, and outputs new parameter compensation predictions and control recommendations that account for the anomaly. These recommendations may include more significant parameter adjustments or the initiation of pre-defined emergency response procedures.

[0081] System Implementation The system of the present invention comprises the following main modules, which are interconnected via an industrial network and a data bus: Raw material attribute acquisition module: includes various sensors (NIR, moisture meter, etc.), data interfaces (LIMS, database) and possible image scanning equipment to obtain comprehensive characteristic data of raw materials before production.

[0082] Physical state detection module: This module includes various sensors deployed on the production line (torque sensors, temperature and humidity probes, laser displacement sensors, force sensors, etc.). It collects physical state parameters such as dough viscosity, elasticity, extensibility, temperature and humidity, and specific volume in real time, continuously, or on demand. It supports hierarchical data collection logic.

[0083] The image-assisted recognition module includes a high-resolution industrial camera, a light source, and an image processing unit (either with a built-in GPU or connected to an external GPU server). It utilizes a convolutional neural network (CNN) architecture, with training samples consisting of high-resolution dough images and their corresponding physical test values. Based on actual images, the model can predict viscosity and elasticity trends, assess bubble structure, and surface conditions. The results are then output to the control unit for comparison, correction, or completion of missing measurement data.

[0084] Control unit: The core computing and decision-making unit of the system. Typically composed of a high-performance industrial PC or server, equipped with acceleration hardware such as a GPU / TPU.

[0085] Model fusion engine: realizes the fusion of multi-source data and the collaborative work of multiple models. Constructed into a multi-path control strategy output structure: The first path (baseline control): Based on the current raw material and process data, the trained FNN and historical data, output static or prediction-based conventional optimal control instructions.

[0086] Second path (dynamic compensation): When the quality prediction model determines that there is a quality deviation or abnormal trend, a dynamic adjustment strategy is generated based on the real-time prediction results of the reinforcement learning (RL) agent to perform more active parameter compensation.

[0087] The third path (anomaly correction / verification) integrates image information (from the image-assisted recognition module) to provide secondary confirmation of anomaly identification. When the outputs of the first and second paths conflict, uncertainty is high, or a serious anomaly is detected, the third path provides correction suggestions based on visual judgment or acts as an arbitrator to ensure the robustness and safety of control decisions.

[0088] Quality target correction mechanism: Integrated in the control unit, it dynamically adjusts the RL reward function or the model’s training objective by analyzing the deviation between the finished product quality data and the target value, allowing the control strategy to more directly optimize the final product quality.

[0089] The control instruction generation and transmission module includes instruction interpretation layer software and industrial communication interface hardware. It converts high-level control suggestions output by the control unit (such as "increase the fermentation temperature by 2°C") into low-level control signals (such as PLC register values, analog outputs, and digital switch signals) that can be recognized by specific actuators.

[0090] Actuator: A device unit that receives control instructions and performs physical operations. This includes, but is not limited to: variable frequency motor drives for kneading equipment, temperature and humidity controllers for fermentation chambers, air circulation fans, steam generators, heating power regulators for baking ovens, and dehumidification damper actuators.

[0091] Finished product quality monitoring module: Located at the end of the production line, it includes a visual inspection system, a volume analyzer, a texture analyzer and other equipment. It is used to measure key quality indicators of the final product and transmit the data back to the control unit to form a closed-loop learning mechanism.

[0092] Example 1: Intelligent control system for wheat flour bread production line The present invention was implemented on a wheat bread production line with a daily output of 10 tons. The specific configuration of the control system is as follows: Raw material attribute collection: Flour testing: A near-infrared analyzer was used to collect protein content (12.5±0.3%), wet gluten (28±2%), ash (0.55±0.05%), and moisture (13.8±0.5%).

[0093] Yeast testing: Use a dedicated fermentation tester to measure the CO2 gas production rate (120±10 ml / h).

[0094] Data collection frequency: Each batch of raw materials is tested once upon entering the factory.

[0095] Physical status detection: First-level sensors: kneading machine torque sensor (0-50Nm), temperature sensor (-10-100℃), humidity sensor (0-100%RH).

[0096] Second level sensor: dough tensile test device, measuring elongation at break (80-150mm).

[0097] Third-level vision system: 4K industrial camera, frame rate 30fps, equipped with ring LED light source.

[0098] Control unit: Hardware: Industrial servers equipped with Intel processors and NVIDIA GPUs.

[0099] Software: Includes CNN image recognition module, LSTM time series prediction module and reinforcement learning agent based on PPO algorithm.

[0100] Model training: Initial training is performed using 500 batches of historical production data, including normal production and exception handling cases.

[0101] Control strategy implementation: When a high protein content in flour is detected (>13%), the system automatically adjusts the kneading parameters: the kneading intensity is reduced (torque setting is reduced by 10%) and the kneading time is extended (increased by 2 minutes) to avoid dough tightness caused by excessive gluten development.

[0102] When the dough temperature rises too quickly (>1.5°C / minute), the system triggers the second-level sensor to detect dough elasticity and adjusts the kneading speed or activates the cooling system based on the results.

[0103] During the fermentation stage, the system dynamically adjusts the temperature (within the range of 32-36°C) and humidity (within the range of 75-85%RH) of the fermentation box based on real-time data of temperature, humidity and dough volume growth rate.

[0104] During the baking stage, the system can adjust the oven temperature curve based on the predicted state of the dough before entering the oven: for dough that is not fermented enough, the initial temperature is increased (230℃) and the high-temperature stage is extended (3 minutes); for dough that is fully fermented, the initial temperature is lowered (210℃) and the high-temperature stage is shortened (1 minute).

[0105] System performance evaluation: After implementing this system, the consistency of finished products increased by 15% (measured by the standard deviation of specific volume).

[0106] Energy efficiency improved by 8% (reduced energy consumption for the same output).

[0107] The adaptability to raw material fluctuations is enhanced, and even if the flour batch changes, the quality fluctuation of the finished product is controlled within ±5%.

[0108] The exception handling success rate reached 92%, a significant improvement over the 65% of traditional control systems.

[0109] Example 2: Actual case of dealing with abnormal fluctuations in raw materials During one production run, the system detected that the protein content of a new batch of flour was abnormally low (10.8%, lower than the normal range of 12.2-12.8%). The control system then made the following adjustments: Raw material evaluation stage: The system compares the current flour data with the historical database and identifies it as "low-protein flour anomaly".

[0110] The prediction model outputs a warning: Using standard parameters may result in insufficient dough strength, a small final product volume, and a hard texture.

[0111] Automatic parameter adjustment: Recipe adjustment: The system recommends increasing the amount of active gluten added (from 0% to 1.5%).

[0112] Adjustment of kneading parameters: Extending the kneading time (from 12 minutes to 15 minutes), increasing the proportion of high-speed kneading stage (from 30% to 40%).

[0113] Fermentation adjustment: Extend the fermentation time (from 60 minutes to 75 minutes) and lower the fermentation temperature (from 35℃ to 33℃).

[0114] Real-time monitoring and fine-tuning: During the kneading process, the system discovered through the torque curve that the dough development speed was still lower than expected, and automatically triggered image acquisition.

[0115] Image analysis showed that the dough surface texture was rough, so the system further extended the kneading time (an additional 2 minutes).

[0116] During the fermentation stage, the system detected that the dough volume growth rate was 10% slower than expected and automatically increased the fermentation humidity (from 78% RH to 83% RH).

[0117] Outcome evaluation: The specific volume of the final product reaches 4.2 ml / g, which is only 6.7% lower than that of the standard product (4.5 ml / g).

[0118] The control group produced using the traditional fixed parameter method had a 18.9% decrease in specific volume.

[0119] The system stores the adjustment strategy and results in the database for future optimization of similar situations.

[0120] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0121] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A food industrial production control method based on artificial intelligence, characterized in that: Methods include: Step A: Before the food processing production line is started, the raw material data collection module collects the original attribute information of the raw materials to be processed, including but not limited to the protein content, moisture content, yeast type and activity, additive ratio, initial humidity, and temperature data of the flour; Step B: During the food processing process, obtaining physical state parameters of the dough in real time, including any combination of at least one or more parameters: viscosity, elasticity, extensibility, surface rebound rate, specific volume, humidity, and temperature; Step C: synchronously inputting the data obtained in step A and step B into a control system having a multi-model fusion mechanism, wherein the control system includes a quality prediction model, a parameter correction model, and a control strategy selection module; Step D: The control system dynamically identifies, compares differences, and evaluates trends in the current batch of raw materials and processing status. Based on the feature weights learned from the training data, it outputs process adjustment suggestions for the corresponding stages, including control parameters for the kneading, fermentation, and baking stages. Step E: Convert the process adjustment suggestions into an instruction set and send it to the control actuator to dynamically adjust the target equipment to compensate for product quality fluctuations caused by changes in raw materials, environment, or equipment status; Step F: Obtain the quality indicators of the end product through the finished product quality acquisition module and transmit them back to the control system for adaptive optimization and model iteration.

2. The artificial intelligence-based food industrial production control method according to claim 1, characterized in that: The acquisition of the dough physical state parameters in step B is a graded acquisition method. The first level is the basic parameter acquisition, including viscosity and temperature and humidity data; the second level is the supplementary feature acquisition, which is enabled when the model determines an abnormal trend, including elasticity and extensibility; the third level is visual recognition assisted acquisition, which uses the image recognition module to evaluate surface wrinkles, bulging morphology and rebound behavior to supplement the modeling accuracy.

3. The artificial intelligence-based food industrial production control method according to claim 1, characterized in that: The control system is built on an end-to-end deep neural network structure, specifically including: Long short-term memory neural network (LSTM) for time series modeling to capture periodic fluctuations in the process; Feedforward neural network model for quality prediction, predicting key quality indicators of finished products based on current status and raw material information; Reinforcement learning control agent uses actual control feedback data as a reward function to perform adaptive policy optimization and achieve closed-loop optimization.

4. The artificial intelligence-based food industrial production control method according to any one of claims 1 to 3, characterized in that: The process adjustment recommendations include: Speed, time and cycle segmented control during kneading stage; Temperature and humidity settings, fermentation duration, and ambient air flow rate during the fermentation stage; Layered temperature zone setting, heating timing curve, dehumidification window opening and wind speed adjustment during the baking stage; Control steam content and adjust equipment motor power output range throughout the entire processing process.

5. The artificial intelligence-based food industrial production control method according to claim 1, characterized in that: The method further includes an exception triggering mechanism: When any parameter shows a deviation or mutation trend that exceeds the set range of the model's expected value, a multiple discrimination process is initiated, including parameter comparison, historical sample review, and image-assisted judgment. The offset trend is adjusted and corrected through the regression correction module, and a new parameter compensation prediction is output.

6. An artificial intelligence-based food industrial production control system, characterized in that: The system includes: Raw material attribute collection module, used to obtain raw material formula composition and batch characteristics before production; Physical state detection module, used to collect dough viscosity, elasticity, extensibility, temperature and humidity parameters in real time during processing; Image-assisted recognition module, used to identify dough surface structure, bubble generation and rebound status; Control unit, used to integrate multi-source data and establish quality prediction, process optimization and control decision models; The control instruction generation and sending module is used to convert the control unit output results into device executable control commands; Actuators, including kneading equipment, fermentation chamber, baking furnace and their supporting drive units; The finished product quality monitoring module is used to collect the specific volume, porosity, color index, and rebound rate data of the final product in real time, and feed it back to the control unit to form a closed-loop learning mechanism.

7. The artificial intelligence-based food industrial production control system according to claim 6 is characterized in that: The control unit includes a model fusion engine, which is constructed as a multi-path control strategy output structure, including: First path: output static optimal control instructions based on current raw materials and process data; Second path: When there is a quality deviation or abnormal trend, a dynamic adjustment strategy is generated based on the reinforcement learning prediction results; The third path: Fuse image information to perform secondary confirmation on abnormal identification and correct the conflicting results output by the first and second paths.

8. The artificial intelligence-based food industrial production control system according to claim 7 is characterized in that: The image-assisted recognition module adopts a convolutional neural network (CNN) structure. The training samples include high-resolution dough images and their corresponding physical test values. The model can predict viscosity trend values ​​and elasticity trend values ​​based on actual images, and compare them with physical test values ​​to correct or complete missing measurement data.

9. The artificial intelligence-based food industrial production control system according to claim 8 is characterized in that: The system is equipped with a quality target correction mechanism based on feedback values ​​from multiple process stages. The mechanism analyzes the actual performance of the current batch of products in terms of target specific volume, structural uniformity, and color consistency indicators, and actively inputs error terms into the control unit, prompting it to dynamically correct weight values ​​and strategy structures.

10. The artificial intelligence-based food industrial production control system according to claim 9 is characterized in that: The actuator and the control unit are connected via an intermediate instruction interpretation layer, which converts the parameter instructions output by the control unit into PLC control language or specific motor control signals.

Images