Real-time tracing quality control system for nutritional ingredients of yak milk powder based on NIR spectrum and AI

Abstract

The invention relates to the technical field of intelligent manufacturing of dairy products, and discloses a real-time tracing quality control system for nutritional ingredients of yak milk powder based on NIR spectrum and AI. The system comprises a historical data and model library module for storing data and training a spectrum calibration model; the multi-link spectrum monitoring module is used for collecting near infrared spectrums in real time in key links of production; the dynamic reference set generation module is used for quantifying the spectral data and matching historical cases to form a dynamic reference set; the double-loop normalization and slope decision module is used for calculating a normalization coefficient based on the set and generating a multi-link cooperative control slope; the intelligent compensation and closed-loop control module outputs compensation parameters through a prediction model according to the coefficient and the slope and drives a production line to adjust the process; and the block chain traceability and feedback optimization module is used for realizing full-link data chaining, query and feedback learning. According to the invention, real-time accurate control and whole-process credible traceability of nutritional ingredients in the production process of yak milk powder are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology for dairy products, and in particular to a real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI. Background Technology

[0002] Yak milk powder, as a high-value-added specialty dairy product, owes its core value to preserving the naturally high levels of protein, conjugated linoleic acid, and lactoferrin, among other active nutrients, found in yak milk. Currently, the industry primarily relies on intermittent laboratory chemical testing methods (such as the Kjeldahl method and high-performance liquid chromatography) to monitor the quality of raw materials, semi-finished products, and finished products. These methods are time-consuming, destructive to samples, and their results lag significantly behind the actual production process, failing to provide real-time data to support immediate adjustments to key process parameters such as pasteurization and spray drying. Therefore, the existing quality control model is essentially a passive management approach based on post-production testing and offline judgment, making it difficult to achieve proactive and precise control of the production process based on real-time component analysis. This easily leads to irreversible loss of heat-sensitive active ingredients and batch-to-batch instability in the nutritional content of the final product, hindering the uniformity of product quality and the yield of high-quality products.

[0003] Furthermore, the core deficiency of existing technologies lies in the lack of a systematic solution capable of deeply coupling and closed-loop linkage among the three core links of yak milk production: component perception, process decision-making, and quality control. Because the natural fluctuations in the nutritional components of raw milk and the complex nonlinear effects of subsequent processing are treated separately, existing methods cannot construct a perception-decision-execution closed loop that can dynamically adapt to raw material differences, predict process impacts in real time, and make globally optimal decisions. This results in a control system that remains essentially open-loop and experience-driven, unable to guarantee the stable retention rate of key active nutrients under varying production conditions, and lacking the ability to self-learn and continuously optimize based on real production data and final quality results. Simultaneously, the broken causal chain between production process data, control decisions, and final product quality prevents quality traceability from reaching a basic information level, failing to provide a reliable data loop for process optimization. Therefore, there is an urgent need in this field for an integrated intelligent quality control system capable of achieving dynamic perception throughout the entire process, data-driven decision-making, real-time closed-loop control, and reliable traceability through self-evolution, to fundamentally overcome the technical bottleneck of high-quality and stable yak milk powder production.

[0004] Therefore, this invention proposes a real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI. Summary of the Invention

[0005] This invention provides a real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI. By establishing a systematic method that integrates multi-stage dynamic spectral sensing, case matching analysis, adaptive decision-making, and real-time closed-loop control, it overcomes the problems of detection lag, data fragmentation, static decision-making, and lack of traceability in traditional quality control, and achieves accurate, stable, adaptive, intelligent control and reliable traceability of yak milk powder nutrients throughout the entire process.

[0006] This invention provides a real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI, comprising: The historical data and model library module is used to store historical yak milk production case data and to train a yak milk-specific near-infrared spectral calibration model based on the historical yak milk production case data. A multi-stage spectral monitoring module is used to collect near-infrared spectra of flowing materials in real time at key production stages of the yak milk powder production line. The dynamic reference set generation module is used to quantify real-time near-infrared spectra based on a yak milk-specific near-infrared spectral calibration model, generate the content values ​​of key nutrients in the current production batch, and match historical production case data to form a dynamic reference case set. The dual-loop normalization and slope decision module is used to calculate key normalization coefficients based on a dynamic reference case set, construct historical case decision slopes, and generate multi-stage collaborative control slopes for the current batch. The intelligent compensation and closed-loop control module is used to output process compensation parameters based on key normalization coefficients and multi-stage collaborative control slopes through a pre-trained sequence prediction model. The process compensation parameters are then converted into control commands and sent to the production line actuators to achieve dynamic closed-loop control. The blockchain traceability and feedback optimization module is used to realize on-chain traceability of production data, provide query services, and perform feedback learning to optimize historical data and model library modules.

[0007] Preferably, in the historical data and model library module, historical yak milk production case data includes near-infrared spectra of milk source acceptance, pasteurization, and spray drying; protein content of milk source acceptance; lactoferrin content of pasteurization; conjugated linoleic acid content of spray drying; final product quality evaluation; dairy farm environmental data; acquisition season information; and key equipment parameters used in production. The near-infrared spectral calibration model for yak milk is a partial least squares regression model or support vector machine regression model optimized for the high protein and high fat characteristics of yak milk.

[0008] Preferably, the multi-stage spectral monitoring module includes a milk source acceptance spectral monitoring unit, a pasteurization process spectral monitoring unit, and a spray drying process spectral monitoring unit; The milk source acceptance spectral monitoring unit corresponds to the milk source acceptance stage of the production line and is installed on the conveying pipeline of the milk collection station or the milk collection tank in the factory. The pasteurization process spectral monitoring unit corresponds to the pasteurization stage of the production line. It is integrated into the heat preservation section of the pasteurizer, and the spectral probe of the pasteurization process spectral monitoring unit directly contacts the flowing dairy products through the aseptic process connector. The spectral monitoring unit for the spray drying process corresponds to the spray drying stage of the production line and is installed near the finished product outlet of the spray drying tower.

[0009] Preferably, the spectral probe of the pasteurization process spectral monitoring unit is equipped with a protective sleeve with a viewing window, and the spectral probe of the spray drying process spectral monitoring unit is equipped with a protective sleeve with a viewing window; the protective sleeve is circulated with a cooling medium to cool and protect the spectral probe.

[0010] Preferably, the dynamic reference set generation module includes: The real-time quantization unit is used to call the yak milk-specific near-infrared spectral calibration model to analyze the real-time near-infrared spectrum and output the current key nutrient content values ​​of the current production batch. The case matching and filtering unit is used to compare the current key nutrient content value with the historical key nutrient content values ​​of each historical production case data in the corresponding stage, which are pre-stored in the historical data and model library module. If the absolute difference between the average value of any dimension of the current key nutrient content value and the average value of the historical key nutrient content value of the corresponding dimension of a historical production case data is less than the preset matching threshold for the corresponding dimension, then the corresponding historical production case data is added to the dynamic reference case set.

[0011] Preferably, the dual-loop normalization and slope decision module includes: The normalization coefficient calculation submodule is used to calculate the first and second normalization coefficients of the current production batch based on a dynamic reference case set. The first normalization coefficient is the ratio of the content of the first type of component in the current critical nutrient content of the current production batch at the milk source acceptance stage to the sum of the content of the first type of component in the critical nutrient content of all historical production case data at the milk source acceptance stage in the dynamic reference case set. The second normalization coefficient is the ratio of the retention rate of the second type of component in the current critical nutrient content value of the current production batch during the pasteurization process to the sum of the retention rates of the second type of component in the critical nutrient content values ​​of all historical production cases in the dynamic reference case set during the pasteurization process. The case decision slope construction unit is used to determine the first decision point for each historical production case data in the dynamic reference case set, with the first normalized coefficient of the historical production case data as the x-axis and the normalized value of the key nutrient content of the final product of the historical production case data as the y-axis. The second decision point is determined with the second normalized coefficient of the historical production case data as the x-axis and the normalized value of the key nutrient content of the final product of the historical production case data as the y-axis. The first decision point and the second decision point are connected to form a decision line segment. The slope of the line connecting the midpoint of the decision line segment and the origin of the coordinate system is calculated as the decision slope of the historical production case data. The multi-stage collaborative control slope fusion unit is used to calculate the weighted average of the decision slopes of all historical production case data in the dynamic reference case set to obtain the multi-stage collaborative control slope.

[0012] Preferably, the intelligent compensation and closed-loop control module includes: The prediction model unit contains a pre-trained sequence prediction model, which is a long short-term memory neural network model. The sequence prediction model takes the first normalization coefficient, the second normalization coefficient, and the multi-stage collaborative control slope as input features, and outputs the predicted values ​​of temperature compensation parameters for the material flowing into the subsequent pasteurization stage and the predicted values ​​of time compensation parameters for the spray drying stage. The parameter verification and instruction conversion unit is used to perform range verification and smoothing on the predicted values ​​of temperature compensation parameters and time compensation parameters output by the prediction model unit, and convert the processed temperature compensation parameters and time compensation parameters into standard industrial control signals. The control command issuing unit is used to send industrial control signals to the temperature controller of the pasteurization equipment and the time program controller of the spray drying equipment through the communication interface.

[0013] Preferably, the training data for the long short-term memory neural network model comes from historical data and successful production cases stored in the model library module; the input of each training sample is the first normalized coefficient, the second normalized coefficient, and the multi-stage collaborative control slope of the successful production case; the training label is the pasteurization temperature compensation amount and spray drying time compensation amount obtained by expert system or optimization algorithm, which can make the final product quality of the successful production case reach the optimal level.

[0014] Preferably, the blockchain traceability and feedback optimization module includes: The end-to-end data encapsulation and on-chain unit is used to generate a unique traceability code for each production batch. It binds the key data generated during the operation of the dynamic reference set generation module, the dual-loop normalization and slope decision module, and the intelligent compensation and closed-loop control module with the unique traceability code to generate a data packet, and stores the hash value of the data packet on the blockchain network. The query service unit provides a data query interface based on a unique traceability code, responds to external query requests and returns the corresponding visual quality control report; The feedback learning unit is used to monitor and collect product market quality feedback based on unique traceability codes. When the feedback data indicates that a batch of products is of abnormal quality, the corresponding full-chain data packet is extracted, and after data cleaning and labeling, it is fed back as a new sample to the historical data and model library module.

[0015] Preferably, the data cleaning and labeling operations performed by the feedback learning unit include: removing obvious abnormal data points from the end-to-end data packets, and labeling new samples with the corresponding successful production process parameter range or quality problem labels based on expert knowledge or re-inspection results.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By constructing a closed-loop control system covering the entire chain—from real-time spectral acquisition, dynamic historical case matching, dual-loop normalization and collaborative slope decision-making, to intelligent predictive compensation—it deeply couples the three major links of component perception, process decision-making, and quality control. It not only solves the problems of detection lag and data silos, but more importantly, it achieves adaptive adaptation to raw material fluctuations through dynamic reference analysis and global consideration of the impact of multiple links through collaborative slope decision-making. Ultimately, it achieves precise and stable retention of heat-sensitive nutrients, real-time intelligent adjustment of the production process, and continuous self-optimization based on a data-closed loop, fundamentally improving product quality consistency and production efficiency.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is the core architecture and data flow diagram of the real-time traceability and quality control system for yak milk powder nutritional components based on NIR spectroscopy and AI in this embodiment of the invention. Figure 2 This is a schematic diagram of the core process of dynamic reference and collaborative decision-making in an embodiment of the present invention; Figure 3 This is a schematic diagram of the cyber-physical system closed-loop and optimization in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 As shown, this invention provides an embodiment of a real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI, comprising: The historical data and model library module is used to store historical yak milk production case data and to train a yak milk-specific near-infrared spectral calibration model based on the historical yak milk production case data. A multi-stage spectral monitoring module is used to collect near-infrared spectra of flowing materials in real time at key production stages of the yak milk powder production line. The dynamic reference set generation module is used to quantify real-time near-infrared spectra based on a yak milk-specific near-infrared spectral calibration model, generate the content values ​​of key nutrients in the current production batch, and match historical production case data to form a dynamic reference case set. The dual-loop normalization and slope decision module is used to calculate key normalization coefficients based on a dynamic reference case set, construct historical case decision slopes, and generate multi-stage collaborative control slopes for the current batch. The intelligent compensation and closed-loop control module is used to output process compensation parameters based on key normalization coefficients and multi-stage collaborative control slopes through a pre-trained sequence prediction model. The process compensation parameters are then converted into control commands and sent to the production line actuators to achieve dynamic closed-loop control. The blockchain traceability and feedback optimization module is used to realize on-chain traceability of production data, provide query services, and perform feedback learning to optimize historical data and model library modules.

[0022] In this embodiment, historical yak milk production case data is a set of records stored in a database in a structured form. Each record corresponds to a completed production batch and includes at least the near-infrared spectral data collected at each key production stage of the batch, the corresponding nutrient content values ​​measured by standard chemical methods, and the production process parameters used.

[0023] In this embodiment, the yak milk-specific near-infrared spectral calibration model is a mathematical relationship model established by chemometrics. This model is trained using a large number of yak milk samples with known chemical composition contents and their near-infrared spectra, thereby being able to predict the content values ​​of specific components such as protein, lactoferrin, and conjugated linoleic acid based on the input spectrum of unknown samples.

[0024] In this embodiment, real-time acquisition of the near-infrared spectrum of flowing materials refers to continuously acquiring the optical signal of the dairy product or milk powder flowing through a specific location on the production pipeline using an online near-infrared spectral probe integrated at that location without interrupting production or contacting the material. This signal reflects the molecular vibration information of the material.

[0025] In this embodiment, generating the key nutrient content value of the current production batch means that the system calls the established yak milk-specific near-infrared spectral calibration model to automatically analyze and calculate the spectral data collected in real time, and convert the spectral features into numerical prediction results of specific nutrients that represent the current state of the material.

[0026] In this embodiment, matching historical production case data to form a dynamic reference case set is an automated screening process. The system compares the key nutrient content values ​​calculated for the current batch with the historical content values ​​of each case in the corresponding stage in the historical database. Based on a preset similarity criterion, it selects several historical cases that are closest to the current state, forming a temporary subset of reference cases for subsequent in-depth analysis.

[0027] In this embodiment, calculating the key normalization coefficient based on the dynamic reference case set means dividing the content value of a key nutrient component in the current batch by the sum of the content values ​​of the corresponding components in all historical cases in the dynamic reference case set, thereby obtaining a proportionality coefficient that represents the current batch's level relative to similar historical groups in terms of that component.

[0028] In this embodiment, constructing the historical case decision slope is a method to condense multidimensional data relationships into a single control feature. For each historical case in the dynamic reference case set, the system generates a slope value that characterizes the strength of the synergistic effect of multiple factors within the case, based on its multiple normalized coefficients and final quality indicators, through specific geometric or mathematical operations.

[0029] In this embodiment, the pre-trained sequence prediction model refers to an artificial intelligence model that has been trained using historical time-series data. This model is able to learn the complex dynamic relationships in the production parameter sequence and predict the adjustment amount to be applied to subsequent process steps in order to make the output result tend to the target based on the characteristics of the new input sequence.

[0030] In this embodiment, converting process compensation parameters into control commands means that the system encapsulates the abstract adjustment quantities (such as temperature compensation values ​​and time compensation values) output by the prediction model into specific control signals that can be recognized and executed by actuators (such as temperature control valves and timers) in accordance with the format specified by the production line programmable logic controller or industrial bus communication protocol.

[0031] In this embodiment, achieving full-chain data traceability in production means that the system generates a unique identifier for each production batch and generates a digital fingerprint from the key status data, decision data, and execution data generated in the entire process from spectral acquisition to the issuance of control commands. This fingerprint is then stored in the blockchain distributed ledger to achieve data immutability and trustworthy traceability.

[0032] In a preferred embodiment of the present invention, the establishment of the yak milk-specific near-infrared spectroscopy calibration model requires the collection of no fewer than 500 batches of raw milk samples covering major yak-producing areas and all four seasons. During modeling, partial least squares regression is preferentially used, and the number of principal components in the model is typically between 8 and 12, determined through cross-validation. The established model should control the prediction error for protein content within ±0.1 g / 100g and the prediction error for lactoferrin within ±3 mg / L.

[0033] In the dual-loop normalization and slope decision module, when assigning decision slope weights to each historical case in the dynamic reference case set, two main factors are considered: first, the quality grade of the final product of that case (e.g., superior cases are assigned higher weights, and first-grade cases are assigned standard weights); and second, the overall similarity between that case and the current batch in terms of key nutrient content. The closer the historical case is to the current batch and the higher its quality, the greater the weight of its decision slope in the weighted average calculation.

[0034] The Long Short-Term Memory (LSTM) neural network prediction model employs a network structure with two hidden layers. It is recommended that the first hidden layer have 64 neurons and the second hidden layer have 32 neurons. Model training uses historical successful case data, with the aforementioned first and second normalized coefficients and the slope of multi-stage collaborative control as inputs, and learns from the optimal process compensation amount confirmed by an expert system. Training continues until the prediction error stabilizes, typically requiring an average prediction error of less than 0.5 degrees Celsius for temperature compensation and less than 5 seconds for time compensation.

[0035] To ensure the robustness of the system, its operational logic under abnormal conditions should be further explained.

[0036] System anomaly handling and safeguards To ensure reliable operation in complex industrial environments, this system includes the following exception handling mechanisms: Handling missing reference cases: If the current batch of data is special and a sufficient number of similar cases cannot be matched from the historical database (i.e., the dynamic reference case set is empty or too small), the system will automatically activate the global model trained based on all historical data and the preset benchmark process parameters for control, and issue a prompt, suggesting that operators pay attention and can use this batch as a new case for subsequent learning.

[0037] Handling Data Anomalies: The system continuously monitors the quality of the spectral signal. When an abnormal increase in signal noise is detected, or if the probe may be contaminated, the data within that time period will be automatically marked as suspicious and its use in critical decision-making will be temporarily suspended. Simultaneously, the system will trigger a maintenance alarm, prompting the inspection or cleaning of the spectral probe. Control decisions will rely on valid data from the preceding period.

[0038] Handling of Predicted Parameter Exceeding Limits: Process adjustment parameters output by the intelligent compensation module must be verified through a preset process safety window before being sent to production equipment. For example, the pasteurization temperature compensation value is limited to an increase or decrease of 2 degrees Celsius. If the model's predicted value exceeds this range, the system will automatically adopt the safety boundary value as the final instruction and record this event for subsequent model optimization and case analysis.

[0039] The following embodiments define the specific composition of historical production case data and optimize the calibration model to provide an accurate and reliable data foundation and quantitative tools for the system's analysis and decision-making. The historical data and model library module proposes that the historical yak milk production case data includes near-infrared spectra of milk source acceptance, pasteurization, and spray drying, protein content values ​​of milk source acceptance, lactoferrin content values ​​of pasteurization, conjugated linoleic acid content values ​​of spray drying, final product quality evaluation, dairy farm environmental data, acquisition season information, and key equipment parameters used in production. The near-infrared spectral calibration model for yak milk is a partial least squares regression model or support vector machine regression model optimized for the high protein and high milk fat characteristics of yak milk.

[0040] In this embodiment, the final product quality evaluation is a comprehensive quality judgment result. It can be a comprehensive score calculated by weighting multiple physicochemical indicators (such as total protein, active protein retention rate, lipid oxidation value, etc.) and sensory indicators, or it can be a grade label based on product standards (such as superior grade, first grade). This evaluation serves as the result label for each historical case and is an important basis for the system to judge the success or failure of the case and optimize decision-making strategies during feedback learning.

[0041] In this embodiment, the environmental data of the dairy farm includes information such as the farm's altitude, average annual temperature, and main forage species. Acquisition season information refers to the specific time of year (spring, summer, autumn, or winter) during which the milk is acquired. These two types of information are incorporated into the database as potential variables affecting the basic components of yak milk. This allows the system to indirectly consider the background differences in raw materials when performing case matching and status analysis, improving the rationality of the reference set selection and the adaptability of the decision-making process.

[0042] In this embodiment, the key equipment parameters used in production refer to the main equipment settings or operating states that are directly related to product quality during specific historical production processes. Examples include the specific sterilization temperature and holding time combination of a pasteurizer, and the inlet air temperature and atomization pressure of a spray drying tower. Recording these parameters helps establish a complete correlation chain between process conditions, process states, and final quality during feedback learning.

[0043] In this embodiment, the partial least squares regression model optimized for the high protein and high fat characteristics of yak milk refers to the model being constructed by specifically using sample spectra covering the actual high-value range of yak milk for calibration. During the modeling process, variable selection or weighting methods are used to enhance the model's ability to resolve characteristic spectral bands representing protein and fat chemical bonds, thereby ensuring higher prediction accuracy and stronger resistance to interference within the unique high concentration range of yak milk. The optimization of the support vector machine regression model follows a similar approach, focusing on the selection of kernel functions and parameters to better fit the nonlinear relationship between yak milk spectra and components.

[0044] The following embodiments achieve real-time, uninterrupted component monitoring of raw materials, intermediate products and finished products throughout the entire process by deploying specific spectral monitoring units at each key production stage. A multi-stage spectral monitoring module is proposed, including a milk source acceptance spectral monitoring unit, a pasteurization process spectral monitoring unit and a spray drying process spectral monitoring unit. The milk source acceptance spectral monitoring unit corresponds to the milk source acceptance stage of the production line and is installed on the conveying pipeline of the milk collection station or the milk collection tank in the factory. The pasteurization process spectral monitoring unit corresponds to the pasteurization stage of the production line. It is integrated into the heat preservation section of the pasteurizer, and the spectral probe of the pasteurization process spectral monitoring unit directly contacts the flowing dairy products through the aseptic process connector. The spectral monitoring unit for the spray drying process corresponds to the spray drying stage of the production line and is installed near the finished product outlet of the spray drying tower.

[0045] In this embodiment, the milk source acceptance stage is the first quality control point for raw milk entering the production system. Installing the milk source acceptance spectral monitoring unit on the conveying pipeline of the milk receiving station or the milk receiving tank within the plant means selecting a representative pipe section with stable flow rate, uniform mixing, and representative characteristics along the flow path of the raw milk during unloading, receiving, and pumping to the storage tank. This ensures that the system can quickly and online evaluate the basic components of the raw milk (such as protein and fat) at the beginning of production, providing an initial basis for subsequent process adjustments.

[0046] In this embodiment, the heat preservation section of the pasteurizer is the key area where dairy products are kept at the set pasteurization temperature for a specified time to complete the pasteurization process. Integrating a pasteurization process spectral monitoring unit into this section of the pipeline, and using an aseptic process connector to allow the spectral probe to directly contact the flowing dairy products, enables in-situ, real-time monitoring of changes in dairy product components during the pasteurization process. Direct contact measurement obtains the strongest spectral signal, which is beneficial for accurately monitoring the dynamic changes of heat-sensitive components (such as lactoferrin) during heat treatment, providing direct data support for real-time fine-tuning of pasteurization parameters.

[0047] In this embodiment, the area near the finished product outlet of the spray drying tower refers to the location before the dried yak milk powder, collected by the cyclone separator, enters the finished product conveying pipeline or temporary storage silo. Installing a spectral monitoring unit for the spray drying process at this location allows for online detection of key components (such as conjugated linoleic acid and protein) in the final powder product. This is equivalent to real-time quality inspection of the final output of the production process. The detection results not only verify the effectiveness of preceding process control but also serve as a key quality record for that batch of products, directly linked to the final quality evaluation and traceability information.

[0048] The following embodiments, by adding a cooling and protection structure to the spectral probe in a high-temperature environment, ensure the long-term stable operation and data reliability of the monitoring unit under harsh process conditions. It is proposed that the spectral probe of the pasteurization process spectral monitoring unit is equipped with a protective sleeve with a viewing window, and the spectral probe of the spray drying process spectral monitoring unit is equipped with a protective sleeve with a viewing window; the protective sleeve is circulated with a cooling medium to cool and protect the spectral probe.

[0049] In this embodiment, the protective sleeve with a viewing window is a sealed tubular housing fitted over the optical front end of the spectral probe. The end in contact with the analyte has a viewing window made of a high-temperature resistant, corrosion-resistant material with high near-infrared light transmittance, such as sapphire or special quartz glass. This structure physically isolates the core optical components of the probe from the high-temperature analyte, preventing direct contamination, coking, or mechanical wear of the probe mirror by dairy products or milk powder, while ensuring unobstructed measurement optical path.

[0050] In this embodiment, the protective sleeve is circulated with a cooling medium, which means that a fluid with a temperature lower than the probe's tolerance limit is continuously introduced into the sleeve's interlayer or built-in flow channel. During its flow, this cooling medium carries away heat transferred from the high-temperature material through the viewing window to the protective sleeve and probe body via heat exchange. This establishes a localized temperature-controlled zone around the probe's sensing head, ensuring that critical components such as electronic devices, optical components, and fiber optic connectors inside the probe can continuously operate within the manufacturer-specified safe ambient temperature range. This prevents performance drift, signal attenuation, or permanent damage caused by overheating, thus guaranteeing the long-term stability and accuracy of the monitoring data.

[0051] The following embodiments, by setting quantification and matching rules, enable the automatic and accurate selection of a set of reference cases most relevant to the current production status from massive historical data, and propose a dynamic reference set generation module, including: The real-time quantization unit is used to call the yak milk-specific near-infrared spectral calibration model to analyze the real-time near-infrared spectrum and output the current key nutrient content values ​​of the current production batch. The case matching and filtering unit is used to compare the current key nutrient content value with the historical key nutrient content values ​​of each historical production case data in the corresponding stage, which are pre-stored in the historical data and model library module. If the absolute difference between the average value of any dimension of the current key nutrient content value and the average value of the historical key nutrient content value of the corresponding dimension of a historical production case data is less than the preset matching threshold for the corresponding dimension, then the corresponding historical production case data is added to the dynamic reference case set.

[0052] In this embodiment, the average value of any dimension of the current key nutrient content refers to a sequence of key nutrient (e.g., protein) content values ​​continuously collected and quantified in real time at a specific production stage (e.g., milk source acceptance) for the current production batch. Calculating the average value of this sequence is to use a statistically representative central trend value to summarize the overall level of the current batch at that stage and for that nutrient, allowing for stable comparison with historical cases and avoiding mismatches caused by single instantaneous fluctuations.

[0053] In this embodiment, the average value of historical key nutrient content for a specific historical production case refers to the content value of the same key nutrient (such as protein) recorded at the same production stage (such as milk source acceptance) for each case in the historical database, obtained through standard method testing or backtesting verification using a calibrated model. Typically, a case may have one or more recorded values ​​for a single nutrient at a particular stage, and calculating the average value is also to obtain the representativeness level of that historical case in that dimension.

[0054] In this embodiment, the preset matching threshold for the corresponding dimension is a pre-defined tolerance range for judging similarity, independently set for each key nutrient dimension to be matched (e.g., protein, lactoferrin). This threshold is typically determined based on the natural fluctuation range of the component during normal production, the accuracy of the detection method, and the requirements of process control. Its function is to quantify the standard of similarity. For example, if the matching threshold for protein content is set to 0.2g / 100g, then the two are considered similar in this dimension only if the absolute value of the difference between the average protein content of the current batch and the average protein content of a historical case is less than 0.2g. Different components can have different thresholds set due to their varying importance and volatility.

[0055] In this embodiment, adding corresponding historical production case data to the dynamic reference case set means that historical cases that meet the above similarity conditions (successful matching in any dimension) will be automatically selected by the system and temporarily stored in a data set specifically created for the current batch. This set is dynamically generated, and its content depends entirely on the status of the current batch. Furthermore, the cases in the set all share a common characteristic: they are highly similar to the current batch in at least one key component indicator. This multi-dimensional, threshold-based filtering logic is the core rule for quickly and automatically focusing on relevant cases from massive amounts of data.

[0056] like Figure 2 As shown, the following embodiments, through a dual-loop normalization and geometric slope decision algorithm, transform the complex multi-stage synergistic influence relationship into a quantifiable and interpretable single control parameter, providing a key input for intelligent regulation. A dual-loop normalization and slope decision module is proposed, including: The normalization coefficient calculation submodule is used to calculate the first and second normalization coefficients of the current production batch based on a dynamic reference case set. The first normalization coefficient is the ratio of the content of the first type of component in the current critical nutrient content of the current production batch at the milk source acceptance stage to the sum of the content of the first type of component in the critical nutrient content of all historical production case data at the milk source acceptance stage in the dynamic reference case set. The second normalization coefficient is the ratio of the retention rate of the second type of component in the current critical nutrient content value of the current production batch during the pasteurization process to the sum of the retention rates of the second type of component in the critical nutrient content values ​​of all historical production cases in the dynamic reference case set during the pasteurization process. The case decision slope construction unit is used to determine the first decision point for each historical production case data in the dynamic reference case set, with the first normalized coefficient of the historical production case data as the x-axis and the normalized value of the key nutrient content of the final product of the historical production case data as the y-axis. The second decision point is determined with the second normalized coefficient of the historical production case data as the x-axis and the normalized value of the key nutrient content of the final product of the historical production case data as the y-axis. The first decision point and the second decision point are connected to form a decision line segment. The slope of the line connecting the midpoint of the decision line segment and the origin of the coordinate system is calculated as the decision slope of the historical production case data. The multi-stage collaborative control slope fusion unit is used to calculate the weighted average of the decision slopes of all historical production case data in the dynamic reference case set to obtain the multi-stage collaborative control slope.

[0057] In this embodiment, the specific calculation process of the first normalization coefficient is as follows: The system first extracts the content values ​​of the first type of component (usually referring to protein) recorded in all historical cases during the milk source acceptance stage from the dynamic reference case set, and sums these values ​​to obtain a total. Then, the content value of the first type of component obtained by real-time monitoring of the current production batch during the milk source acceptance stage is divided by the total of these historical cases. This ratio is the first normalization coefficient, which characterizes the relative strength of the protein level of the current raw material compared to the average level of a similar group of past raw materials.

[0058] In this embodiment, the calculation of the second normalization coefficient introduces the concept of retention rate. The second category of components typically refers to heat-sensitive active substances such as lactoferrin. Its retention rate is calculated as a percentage by comparing the content of this component after pasteurization with its initial content during milk source acceptance. The system extracts and sums the lactoferrin retention rate values ​​calculated during pasteurization from all historical cases in the dynamic reference case set. Dividing the current batch's real-time lactoferrin retention rate value calculated during pasteurization by this historical sum yields the second normalization coefficient. This coefficient characterizes the protective effect of the current sterilization process on the active ingredient, relative to the average level of a similar historical case group.

[0059] In this embodiment, the normalized value of the key nutrient content in the final product is the result of normalizing a certain key quality indicator (usually conjugated linoleic acid content) for the final product of each historical case in the dynamic reference case set. The processing method can be to divide the conjugated linoleic acid content value of the case by the maximum conjugated linoleic acid content of all cases in the dynamic reference case set, or by their sum. The purpose is to eliminate the dimensions of absolute values, allowing them to be calculated and compared on the same scale as the first and second normalized coefficients.

[0060] In this embodiment, a decision line segment is formed by connecting the first decision point and the second decision point. Geometrically, this segment establishes a correlation path for each historical case within its own state space. The two endpoints of this line segment reflect the single-point relationships between the raw material basis (first normalization coefficient) and the process effect (second normalization coefficient) of the case, respectively, and the final product quality (normalized value of conjugate linoleic acid). The entire line segment can be viewed as a virtual trajectory from the raw material starting point, through the process midpoint, to the quality endpoint.

[0061] In this embodiment, the slope of the line connecting the midpoint of the decision segment and the origin of the coordinate system is used as the historical case decision slope. This is a method to compress two-dimensional line segment information into a one-dimensional scalar. The origin represents a theoretical state of zero raw materials, zero process, and zero quality. Connecting the origin and the midpoint of the line segment, the slope comprehensively reflects the degree and direction of the joint contribution of raw materials and process to the final quality in this case. The larger the absolute value of the slope, the more significant the combined effect of raw materials and process on the final quality in this case. The sign of the slope indicates whether this combined effect is positively enhancing or negatively weakening. This slope comprehensively characterizes the synergistic contribution strength of raw material characteristics (first normalization coefficient) and process effect (second normalization coefficient) to the final quality in historical cases, providing a quantitative basis for multi-stage synergistic adjustment in the current batch.

[0062] In this embodiment, the midpoint of the decision segment is chosen to calculate the slope because it simultaneously contains balanced information on both the raw material basis (first decision point) and the process effect (second decision point), avoiding the one-sidedness of a single endpoint reflecting only a certain stage. Compared to the endpoint slope, which only focuses on the end effect of the process, and the starting point slope, which only focuses on the initial state of the raw materials, the midpoint slope can more accurately quantify the correlation between the raw material-process synergy and the final quality. For example, the larger the absolute value of the slope, the higher the matching degree between the raw material characteristics and process parameters in that historical case, and the better the final quality. By integrating the midpoint slopes of multiple similar historical cases, the resulting multi-stage synergistic control slope can directly guide the process compensation strategy for the current batch, enabling raw material fluctuations and process adjustments to form a closed-loop adaptation. Compared to traditional single-stage control, this can reduce batch fluctuations in the retention rate of key nutrients by more than 15%, significantly improving quality stability.

[0063] In this embodiment, calculating a weighted average of the decision slopes to obtain a multi-stage collaborative control slope is a key step in integrating collective wisdom. Weighted averaging means that not all historical case decision slopes are treated equally. Weights can be allocated based on factors such as the overall similarity of the case to the current batch and the case's own final quality. Through weighted averaging, the system ultimately obtains a single control parameter that embodies the consensus of similar historical group experiences. This parameter quantifies the overall strategic direction of how raw materials and processes should be coordinated and adjusted to achieve the expected quality target under the current state.

[0064] The following embodiments, through specific predictive models and industrial signal conversion mechanisms, transform abstract decision characteristics into precisely executable equipment control commands, completing the final link from intelligent decision-making to physical execution. An intelligent compensation and closed-loop control module is proposed, including: The prediction model unit contains a pre-trained sequence prediction model, which is a long short-term memory neural network model. The sequence prediction model takes the first normalization coefficient, the second normalization coefficient, and the multi-stage collaborative control slope as input features, and outputs the predicted values ​​of temperature compensation parameters for the material flowing into the subsequent pasteurization stage and the predicted values ​​of time compensation parameters for the spray drying stage. The parameter verification and instruction conversion unit is used to perform range verification and smoothing on the predicted values ​​of temperature compensation parameters and time compensation parameters output by the prediction model unit, and convert the processed temperature compensation parameters and time compensation parameters into standard industrial control signals. The control command issuing unit is used to send industrial control signals to the temperature controller of the pasteurization equipment and the time program controller of the spray drying equipment through the communication interface.

[0065] In this embodiment, the Long Short-Term Memory (LSTM) neural network model is a special type of recurrent neural network that includes memory units and gating mechanisms (input gate, forget gate, output gate), enabling it to effectively learn and remember long-term dependencies and temporal patterns in the input data sequence. In this application, the model is trained to learn the complex nonlinear mapping relationship between a sequence pattern composed of three features—a first normalization coefficient, a second normalization coefficient, and a multi-stage collaborative control slope—arranged according to the production timeline, and the compensation sequence required to adjust the subsequent pasteurization temperature and spray drying time to ensure the final product meets standards.

[0066] In this embodiment, the predicted values ​​of the temperature compensation parameter and the time compensation parameter are adjustment amounts with specific numerical values ​​and physical units, directly output by the prediction model. For example, the temperature compensation parameter might be a value in degrees Celsius, such as +0.5℃ or -1.2℃, indicating the suggested increase or decrease in temperature from the original setpoint. Similarly, the time compensation parameter is a value in seconds. These are process parameter fine-tuning suggestions derived from the model based on the current production status characteristics, aimed at optimizing the final result.

[0067] In this embodiment, range verification refers to the logical process by which the system compares the compensation parameter values ​​predicted by the model with pre-set safety process boundary values. These boundary values ​​are determined based on equipment capabilities, product process specifications, and food safety requirements; for example, the pasteurization temperature compensation range may be limited to [-2℃, +2℃]. If the predicted value exceeds this range, the system will automatically truncate it to the boundary value to prevent the issuance of unsafe or unreasonable control commands.

[0068] In this embodiment, smoothing typically refers to using digital filtering algorithms (such as moving average filtering or first-order lag filtering) to process the predicted compensation parameter sequence. The purpose is to eliminate random noise or occasional sharp fluctuations in the model output, making the final command value change curve more gradual and stable, avoiding unnecessary disturbances to the production line caused by step changes in commands, and improving the smoothness of the control process.

[0069] In this embodiment, converting to a standard industrial control signal refers to encoding and encapsulating the compensation parameters, which have undergone verification and smoothing and are expressed as physical quantities, according to a specific protocol and format that can be recognized by the existing programmable logic controller or distributed control system on the production line. For example, the temperature compensation value is converted into a 4-20mA current signal setpoint for the corresponding analog output module, or the time compensation value is converted into a specific data message sent to the equipment controller via industrial Ethernet. The control command issuing unit is responsible for transmitting this standard signal reliably and in real time to the actuator (such as the PID adjustment module in the temperature controller or the timing setting unit of the time program controller) through the corresponding physical communication interface (such as an analog output card or an industrial Ethernet port).

[0070] In this embodiment, the spectral monitoring data of the pasteurization process is used to analyze the actual effect of the current sterilization process. The corresponding temperature compensation parameters are not applied to the same batch of materials that have been monitored. Instead, they are sent to the temperature controller in real time through closed-loop control logic to adjust the heating parameters of the next batch or subsequent batches of materials flowing into the sterilization zone, so as to ensure the consistency of process stability and nutrient retention rate throughout the entire production batch.

[0071] The following embodiments define a model training method based on historical success cases, enabling the intelligent prediction model to learn the optimal process adjustment strategy and continuously output high-quality compensation parameters. The training data of the long short-term memory neural network model comes from historical data and successful production cases stored in the model library module. The input of each training sample is the first normalized coefficient, the second normalized coefficient, and the multi-stage collaborative control slope of the successful production case. The training labels are the pasteurization temperature compensation amount and spray drying time compensation amount obtained by expert system or optimization algorithm, which can make the final product quality of the successful production case reach the optimal level.

[0072] In this embodiment, a successful production case refers to a historical batch record in the historical data and model library module where the final product quality evaluation is marked as excellent or reaches a preset high standard threshold, and the entire production process data is complete and reliable. These cases represent best practices that have been validated in past production practices.

[0073] In this embodiment, the input for each training sample is constructed by extracting and calculating three key features from the production data of each selected successful production case: the first normalized coefficient, the second normalized coefficient, and the slope of multi-stage collaborative control. These three features are arranged in the actual production time sequence or logical sequence to form a feature sequence characterizing the collaborative state of the case from raw materials to intermediate processes, which serves as the historical state input for model learning.

[0074] In this embodiment, the training labels are not constructed by directly using the original process parameters (such as sterilization temperature and drying time) actually used in the historical records of successful cases. Instead, they are calculated through reverse engineering or optimization algorithms. Specifically, assuming that the case ultimately achieves high-quality results, a set of pasteurization temperature and spray drying time adjustments are theoretically found through back-engineering or iterative calculations using an expert system rule base or numerical optimization algorithms (such as reverse simulation or gradient descent) to achieve better or more stable final quality indicators for the case. These calculated adjustments, as the optimal compensation amounts, constitute the target output labels for the training samples. This method aims to train the model to learn an adjustment strategy that pursues extreme or robust optimization, rather than simply reproducing historical operations.

[0075] like Figure 3 As shown, the following embodiments construct a blockchain-based end-to-end data storage and query service system, realizing trusted and transparent traceability of the production process and convenient access to quality data. A blockchain traceability and feedback optimization module is proposed, including: The end-to-end data encapsulation and on-chain unit is used to generate a unique traceability code for each production batch. It binds the key data generated during the operation of the dynamic reference set generation module, the dual-loop normalization and slope decision module, and the intelligent compensation and closed-loop control module with the unique traceability code to generate a data packet, and stores the hash value of the data packet on the blockchain network. The query service unit provides a data query interface based on a unique traceability code, responds to external query requests and returns the corresponding visual quality control report; The feedback learning unit is used to monitor and collect product market quality feedback based on unique traceability codes. When the feedback data indicates that a batch of products is of abnormal quality, the corresponding full-chain data packet is extracted, and after data cleaning and labeling, it is fed back as a new sample to the historical data and model library module.

[0076] In this embodiment, generating a unique traceability code for each production batch means that at the start of each batch's production, the system automatically generates a globally unique string identifier using an algorithm (such as a UUID generation algorithm or a combination of timestamp and production line number encoding rules). This traceability code will serve as a unique identifier for that batch across all information systems and physical packaging, throughout its entire lifecycle.

[0077] In this embodiment, storing the hash value of the data packet on the blockchain network is the core operation of blockchain notarization. The specific process is as follows: The system first calculates a fixed-length and unique digital fingerprint, i.e., a hash value, by using an encrypted hash function (such as the SHA-256 algorithm) on the packaged data packet containing a unique traceability code and key data across the entire chain. Then, by calling the blockchain network's application programming interface (API), this hash value, along with information such as a timestamp, is packaged into a transaction and sent to the blockchain node. After verification by node consensus, the transaction is recorded in a new block and appended to the chain, thus achieving immutable and permanent notarization of the hash value. During this process, the original data packet itself is typically stored in a trusted local database or distributed file system, and its integrity can be verified by recalculating the hash value and comparing it with the value stored on the chain at any time.

[0078] In this embodiment, providing a data query interface based on a unique traceability code means that the system develops and exposes a standard application programming interface (e.g., a RESTful API). When external users (such as consumers scanning a code or regulatory personnel accessing the backend) initiate a query request, they must provide a valid batch unique traceability code. After receiving the code, the query service unit first verifies the integrity and authenticity of the data hash value record associated with the code on the blockchain. Then, based on the traceability code, it retrieves the corresponding end-to-end data packet from local storage and organizes the key information (such as component values ​​at each stage, decision parameters, and process adjustment records) into easily understandable charts, curves, and text reports according to a preset template. Finally, this visualized quality control report is returned to the requester through the interface.

[0079] In this embodiment, monitoring and collecting product market quality feedback based on unique traceability codes means that the system has established a feedback channel linked to the market. For example, this can be achieved by linking online feedback forms to QR codes on product packaging, or by periodically extracting customer reviews, complaints, or return data associated with specific traceability codes from enterprise customer relationship management systems and after-sales service systems. The system continuously monitors these channels and automatically or semi-automatically matches the collected unstructured feedback information with specific production batch traceability codes to form structured batch-feedback records.

[0080] The following embodiments demonstrate how a data cleaning and labeling feedback mechanism enables the system to continuously optimize its internal knowledge base using market quality feedback, thereby achieving self-learning and performance improvement. The data cleaning and labeling operations performed by the feedback learning unit include: removing obvious abnormal data points from the entire data package and, based on expert knowledge or re-inspection results, labeling new samples with the corresponding successful production process parameter range or quality problem labels.

[0081] In this embodiment, removing obviously abnormal data points from the end-to-end data packet refers to a data quality check performed before adding end-to-end data indicating abnormal batches from market feedback as new samples to the knowledge base. Specifically, the system automatically scans all time-series data in the data packet based on statistical rules (such as the Three Sigma rule) or physical constraints (such as equipment operating limits and reasonable ranges of component concentrations), identifying and removing isolated noise points that significantly deviate from normal patterns, possibly due to momentary sensor malfunctions, communication interference, or recording errors. For example, a temperature reading at a certain moment that is far above the safety limit of the sterilization equipment, or a protein content calculated from a certain spectrum that is negative, such data points will be considered invalid and safely removed to ensure the basic data quality of the learning samples.

[0082] In this embodiment, labeling new samples based on expert knowledge or re-inspection results is a crucial step in giving learning significance to cleaned abnormal batch data. Expert knowledge refers to domain experts (such as process engineers and quality control experts) labeling the batch data with one or more conclusive tags based on the combination of process parameters, component change trends, and the final market feedback problem descriptions reflected in the abnormal batch data, combined with their professional experience. For example, if an expert judges that the abnormality stems from insufficient microbial control due to a slightly low sterilization temperature, a label indicating that the lower limit of sterilization temperature needs to be increased or a microbial risk label may be added; if the expert judges that the abnormality stems from insufficient compensation for fluctuations in raw milk protein, a label indicating high sensitivity of raw protein may be added. Re-inspection results refer to sending the retained batch of products back to the laboratory for comprehensive testing, using conclusive physicochemical or microbiological data to confirm the problem, and adding more precise labels accordingly, such as lactoferrin retention rate below 85% or moisture content exceeding the standard. Samples with such labels will be stored in the historical database as counterexamples or cases requiring special attention, used to enhance the system's ability to identify and avoid specific risk patterns in subsequent case matching or model training.

[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI, characterized in that, include: The historical data and model library module is used to store historical yak milk production case data and to train a yak milk-specific near-infrared spectral calibration model based on the historical yak milk production case data. A multi-stage spectral monitoring module is used to collect near-infrared spectra of flowing materials in real time at key production stages of the yak milk powder production line. The dynamic reference set generation module is used to quantify real-time near-infrared spectra based on a yak milk-specific near-infrared spectral calibration model, generate the content values ​​of key nutrients in the current production batch, and match historical production case data to form a dynamic reference case set. The dual-loop normalization and slope decision module is used to calculate key normalization coefficients based on a dynamic reference case set, construct historical case decision slopes, and generate multi-stage collaborative control slopes for the current batch. The intelligent compensation and closed-loop control module is used to output process compensation parameters based on key normalization coefficients and multi-stage collaborative control slopes through a pre-trained sequence prediction model. The process compensation parameters are then converted into control commands and sent to the production line actuators to achieve dynamic closed-loop control. The blockchain traceability and feedback optimization module is used to realize on-chain traceability of production data, provide query services, and perform feedback learning to optimize historical data and model library modules.

2. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI as described in claim 1, characterized in that, The historical data and model library module contains historical yak milk production case data, including near-infrared spectra of milk source acceptance, pasteurization, and spray drying; protein content of milk source acceptance; lactoferrin content of pasteurization; conjugated linoleic acid content of spray drying; final product quality evaluation; dairy farm environmental data; acquisition season information; and key equipment parameters used in production. The near-infrared spectral calibration model for yak milk is a partial least squares regression model or support vector machine regression model optimized for the high protein and high fat characteristics of yak milk.

3. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI as described in claim 1, characterized in that, The multi-stage spectral monitoring module includes a milk source acceptance spectral monitoring unit, a pasteurization process spectral monitoring unit, and a spray drying process spectral monitoring unit; The milk source acceptance spectral monitoring unit corresponds to the milk source acceptance stage of the production line and is installed on the conveying pipeline of the milk collection station or the milk collection tank in the factory. The pasteurization process spectral monitoring unit corresponds to the pasteurization stage of the production line. It is integrated into the heat preservation section of the pasteurizer, and the spectral probe of the pasteurization process spectral monitoring unit directly contacts the flowing dairy products through the aseptic process connector. The spectral monitoring unit for the spray drying process corresponds to the spray drying stage of the production line and is installed near the finished product outlet of the spray drying tower.

4. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 3, characterized in that, The spectral probe of the pasteurization process spectral monitoring unit is equipped with a protective sleeve with a viewing window, and the spectral probe of the spray drying process spectral monitoring unit is also equipped with a protective sleeve with a viewing window; the protective sleeve is circulated with a cooling medium to cool and protect the spectral probe.

5. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 1, characterized in that, The dynamic reference set generation module includes: The real-time quantization unit is used to call the yak milk-specific near-infrared spectral calibration model to analyze the real-time near-infrared spectrum and output the current key nutrient content values ​​of the current production batch. The case matching and filtering unit is used to compare the current key nutrient content value with the historical key nutrient content values ​​of each historical production case data in the corresponding stage, which are pre-stored in the historical data and model library module. If the absolute difference between the average value of any dimension of the current key nutrient content value and the average value of the historical key nutrient content value of the corresponding dimension of a historical production case data is less than the preset matching threshold for the corresponding dimension, then the corresponding historical production case data is added to the dynamic reference case set.

6. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 1, characterized in that, The dual-loop normalization and slope decision module includes: The normalization coefficient calculation submodule is used to calculate the first and second normalization coefficients of the current production batch based on a dynamic reference case set. The first normalization coefficient is the ratio of the content of the first type of component in the current critical nutrient content of the current production batch at the milk source acceptance stage to the sum of the content of the first type of component in the critical nutrient content of all historical production case data at the milk source acceptance stage in the dynamic reference case set. The second normalization coefficient is the ratio of the retention rate of the second type of component in the current critical nutrient content value of the current production batch during the pasteurization process to the sum of the retention rates of the second type of component in the critical nutrient content values ​​of all historical production cases in the dynamic reference case set during the pasteurization process. The case decision slope construction unit is used to determine the first decision point for each historical production case data in the dynamic reference case set, with the first normalized coefficient of the historical production case data as the x-axis and the normalized value of the key nutrient content of the final product of the historical production case data as the y-axis. The second decision point is determined with the second normalized coefficient of the historical production case data as the x-axis and the normalized value of the key nutrient content of the final product of the historical production case data as the y-axis. The first decision point and the second decision point are connected to form a decision line segment. The slope of the line connecting the midpoint of the decision line segment and the origin of the coordinate system is calculated as the decision slope of the historical production case data. The multi-stage collaborative control slope fusion unit is used to calculate the weighted average of the decision slopes of all historical production case data in the dynamic reference case set to obtain the multi-stage collaborative control slope.

7. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 1, characterized in that, The intelligent compensation and closed-loop control module includes: The prediction model unit contains a pre-trained sequence prediction model, which is a long short-term memory neural network model. The sequence prediction model takes the first normalization coefficient, the second normalization coefficient, and the multi-stage collaborative control slope as input features, and outputs the predicted values ​​of temperature compensation parameters for the material flowing into the subsequent pasteurization stage and the predicted values ​​of time compensation parameters for the spray drying stage. The parameter verification and instruction conversion unit is used to perform range verification and smoothing on the predicted values ​​of temperature compensation parameters and time compensation parameters output by the prediction model unit, and convert the processed temperature compensation parameters and time compensation parameters into standard industrial control signals. The control command issuing unit is used to send industrial control signals to the temperature controller of the pasteurization equipment and the time program controller of the spray drying equipment through the communication interface.

8. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 7, characterized in that, The training data for the long short-term memory neural network model comes from historical data and successful production cases stored in the model library module. The input of each training sample is the first normalized coefficient, the second normalized coefficient, and the multi-stage collaborative control slope of the successful production case. The training labels are the pasteurization temperature compensation amount and spray drying time compensation amount obtained by expert system or optimization algorithm, which can make the final product quality of the successful production case reach the optimal level.

9. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 1, characterized in that, The blockchain traceability and feedback optimization module includes: The end-to-end data encapsulation and on-chain unit is used to generate a unique traceability code for each production batch. It binds the key data generated during the operation of the dynamic reference set generation module, the dual-loop normalization and slope decision module, and the intelligent compensation and closed-loop control module with the unique traceability code to generate a data packet, and stores the hash value of the data packet on the blockchain network. The query service unit provides a data query interface based on a unique traceability code, responds to external query requests and returns the corresponding visual quality control report; The feedback learning unit is used to monitor and collect product market quality feedback based on unique traceability codes. When the feedback data indicates that a batch of products is of abnormal quality, the corresponding full-chain data packet is extracted, and after data cleaning and labeling, it is fed back as a new sample to the historical data and model library module.

10. The real-time traceability and quality control system for yak milk powder nutrients based on NIR spectroscopy and AI according to claim 9, characterized in that, The data cleaning and labeling operations performed by the feedback learning unit include: removing obvious abnormal data points from the end-to-end data packets, and labeling new samples with the corresponding successful production process parameter ranges or quality problem labels based on expert knowledge or re-inspection results.

Images