Method for analyzing and optimizing quality of whole process of oryzanol production by fusing multi-source heterogeneous data

Abstract

This invention relates to the field of oryzanol production quality control, and more particularly to a method for multi-source heterogeneous data fusion analysis and quality optimization throughout the entire oryzanol production process. The method includes the following steps: deploying edge acquisition nodes to collect multi-source heterogeneous data and transmitting it to the edge nodes; performing differentiated preprocessing at the edge nodes, generating virtual data in the cloud to form an enhanced dataset; constructing a multi-scale feature matrix based on the enhanced dataset, and fusing and outputting a fused feature vector using a multi-scale attention module; combining multiple model inputs with improved evidence theory to output quality prediction results; and using an improved SHAP value analysis module to locate key control nodes, and generating optimization strategies through federated reinforcement learning. This invention improves the stability and optimization accuracy of oryzanol production quality through differentiated preprocessing, multi-model fusion, SHAP analysis, and federated reinforcement learning, while ensuring data privacy and enabling quality traceability.

Description

Technical Field

[0001] This invention relates to the field of quality control in oryzanol production, and in particular to a method for multi-source heterogeneous data fusion analysis and quality optimization throughout the entire oryzanol production process. Background Technology

[0002] As an important fat-soluble active substance, oryzanol is widely used in food, health products, and pharmaceuticals. Its production process covers four core stages: raw material pretreatment, extraction, purification, and finished product preparation. Each stage requires the collection of various types of data, such as raw material properties, process parameter timing, equipment operation, environmental perception, and quality testing. Existing technologies mostly collect data through centralized systems and combine single models or human experience for quality control. Data processing and optimization decisions rely on centralized cloud computing. In some scenarios, simple data fusion or predictive models are introduced to assist in production adjustments.

[0003] Existing quality control technologies for oryzanol production have significant shortcomings. After data collection, there is a lack of systematic and differentiated preprocessing mechanisms, and the characteristics of different data types are not effectively extracted, resulting in insufficient accuracy in the fusion of multi-source heterogeneous data. In small-sample scenarios, data support is insufficient, making it difficult to adapt to special operating conditions such as sudden changes in raw material properties and abnormal equipment parameters. The fusion of multi-model prediction results lacks an efficient conflict resolution mechanism, and the prediction accuracy and confidence level are insufficient to meet industrial needs. Quality optimization is mostly limited to local adjustments in a single factory, lacking a multi-factory collaborative optimization mechanism and posing a risk of data privacy leakage. Quality traceability relies on traditional database storage, which is prone to data tampering and cannot reproduce the production process, making it difficult to accurately pinpoint the root cause of quality problems. Furthermore, the optimization and adjustment of key process parameters lacks specificity, failing to achieve precise quality control throughout the entire process. Summary of the Invention

[0004] To address the technical deficiencies in the background technology, this invention proposes a method for multi-source heterogeneous data fusion analysis and quality optimization throughout the entire oryzanol production process. This method solves the aforementioned technical problems and meets practical needs. The specific technical solution is as follows: A method for multi-source heterogeneous data fusion analysis and quality optimization throughout the entire oryzanol production process includes the following steps: Based on the entire production process of oryzanol, edge acquisition nodes are deployed to collect multi-source heterogeneous data and transmit it to the edge nodes. Based on multi-source heterogeneous data, edge nodes perform differentiated preprocessing on different types of data, synchronize them to the cloud distributed database, and generate virtual data in the cloud to supplement small sample scenarios, forming an enhanced dataset; Various data features are extracted from the augmented dataset to construct a multi-scale feature matrix. After being fused by an adaptive multi-scale attention module, the fused feature vector is output. The fused feature vectors are input into two data-driven models, while the pre-processed multi-source heterogeneous data is input into the preset digital twin and mechanism model. The quality prediction results and confidence are output by improving the evidence theory. The fused feature vector input is used to improve the SHAP value analysis module to locate key control nodes, and the optimization strategy is generated by federated reinforcement learning in combination with the quality prediction results.

[0005] Furthermore, the multi-source heterogeneous data includes raw material attributes, process parameter time series, equipment operation, environmental perception, and quality inspection data. The specific steps for the edge nodes to perform differentiated preprocessing on different types of data are as follows: The system receives raw material attributes, process parameter timing, equipment operation, environmental perception, and quality inspection data transmitted from edge acquisition nodes, carrying links, timestamps, equipment number, and 3D tags. These data are categorized into three types based on data type: structured data, time-series data, and spectral data. For structured data, dynamic standardization is adopted to update the data mean and standard deviation in real time to adapt to the fluctuation of raw material batches and eliminate the difference in data units. For time-series data, adaptive wavelet threshold denoising is first used. The threshold size is dynamically adjusted according to the data noise intensity to reduce the proportion of high-frequency interference signals. Then, the window size is dynamically adjusted according to the data frequency for feature extraction. High-frequency equipment operation data uses a short window for feature extraction, while low-frequency environmental perception data uses a long window for feature extraction. The window size is set according to the data acquisition frequency and process response characteristics. MSC correction, adaptive smoothing, and sparse dimensionality reduction are performed sequentially on the spectral data. MSC correction is used to eliminate baseline drift and scattering interference. The polynomial order of the adaptive smoothing is dynamically selected according to the complexity of the spectral signal. Sparse dimensionality reduction significantly compresses the amount of data while retaining the core information. A dual screening mechanism is used to remove abnormal data for all types of data. First, the 3σ criterion is used for preliminary screening. Then, the process rules are verified according to the reasonable range set in the oryzanol production technical specifications. Data that exceeds the range and whose duration meets the abnormal judgment criteria is judged as abnormal. For datasets with missing data after screening, the process correlation interpolation method is used to complete them, and the completion value is calculated based on the process correlation between the missing data stage and the preceding and following processes.

[0006] Furthermore, the specific steps for generating virtual data in the cloud to supplement the small sample scenario are as follows: Receive preprocessed data synchronized from edge nodes to the cloud distributed database, classify the preprocessed data according to production scenarios, identify small sample scenarios such as sudden changes in raw material properties, abnormal equipment parameters, and adjustments to process conditions, and determine the core data characteristics and distribution patterns of small sample scenarios; Based on the core data features of small sample scenarios, a generative adversarial network model is constructed. Real preprocessed data is used as training samples to train the generator and discriminator of the generative adversarial network. The generator learns the distribution characteristics of real data, and the discriminator distinguishes and verifies the generated data from the real data. Virtual data is generated by training a mature generative adversarial network. The generated virtual data must be consistent with the real data in the corresponding small sample scenario in terms of feature distribution and process logic. The generated virtual data is validated for process rationality. Abnormal virtual data that does not conform to the logic of oryzanol production process is removed, and valid virtual data that conforms to actual production is retained. By integrating effective virtual data with real preprocessed data in a cloud-distributed database, an enhanced dataset covering all scenarios is formed.

[0007] Furthermore, the specific steps for constructing the multi-scale feature matrix are as follows: Three types of raw data—structured data, time-series data, and spectral data—are extracted from the augmented dataset generated in the cloud, while retaining the production process and timestamp labels carried by the data. Based on three types of raw data, a 1DCNN network is used to perform convolution operations on short-timescale data to extract local temporal features and instantaneous change features. The short-timescale data includes real-time parameters of equipment operation and instantaneous data of the process. Based on three types of raw data, a bidirectional LSTM network is used to perform sequence learning on long-term data to extract long-range dependencies and trend features in the time dimension. The long-term data includes raw material batch attribute data, finished product quality batch detection data, and process parameter trend data. Based on three types of raw data, a CNN network is used to perform multi-layer convolution and pooling operations on the spectral data to extract feature peaks, feature valleys and texture distribution features. Short-scale local temporal features, long-scale long-range dependency features, and spectral texture features are vectorized and normalized respectively. Based on the correspondence between production links and timestamps, a multi-scale feature matrix is ​​constructed.

[0008] Furthermore, the specific steps of the adaptive multi-scale attention module fusion are as follows: Based on the multi-scale feature matrix, the total number of feature categories in the matrix and the types of quality indicators to be predicted are determined. The correlation strength between each feature and each quality indicator is calculated using a gating unit, as shown in the following formula: , in, This represents the quantified value of the correlation strength between the i-th feature and the j-th quality indicator. This represents the weight matrix of the gated unit. Represents the feature vector of the i-th class. This represents the j-th quality index parameter. This represents a vector concatenation operation. This indicates the bias term of the gating unit; The Sigmoid activation function is used to process the quantified values ​​of correlation strength, and the process correlation weights of various features with each quality index are calculated. The calculation formula is as follows: , in, This represents the process relevance weight between the i-th feature and the j-th quality indicator. represents the Sigmoid activation function, and n represents the total number of feature categories involved in the weight calculation; Represents all features corresponding to k from 1 to n Summation operation on the values; The weight matrix and bias terms of the gated unit are iteratively optimized by using the gradient descent algorithm. The various eigenvectors in the multi-scale feature matrix are weighted with the corresponding process-related weights to obtain the weighted eigenvectors. All weighted feature vectors are input into the Transformer encoder, and deep feature fusion is performed through a multi-head attention mechanism to output a fused feature vector.

[0009] Furthermore, the specific steps for constructing the digital twin are as follows: Collect physical entity information of the entire production process of oryzanol, including the physical structure parameters, installation layout, and operating principle of raw material pretreatment equipment, extraction equipment, purification equipment, and finished product preparation equipment, as well as the process flow diagram, process parameter range, and material transfer path of each link; A digital twin is built based on a 3D simulation engine and mapped 1:1 to the actual production scene, replicating the physical structure, operating parameters and process logic of various production equipment according to the physical entity information; Establish a two-way mapping mechanism between physical entities and digital twins. Through edge nodes, synchronize the real-time operation data and process parameter data of the physical entity to the digital twin, and transmit the simulation results of the digital twin back to the control terminal of the physical entity. The extraction kinetics mechanism model and the crystallization thermodynamics mechanism model are embedded in the digital twin, and the process parameter simulation and deduction function is configured for the digital twin. It can receive the pre-processed process parameters, simulate the quality change trend under different working conditions, and continuously compare and calibrate with historical actual production data.

[0010] Furthermore, the specific steps for improving the evidence theory fusion output quality prediction results and confidence levels are as follows: It receives data-driven prediction results from two data-driven models: the first is an improved gradient boosting tree model, and the second is a feature dependency capture model. It also receives simulation prediction results from the digital twin and mechanistic prediction results from the mechanistic model. Normalize the data-driven prediction results, simulation prediction results and mechanism prediction results respectively, map the numerical range to the interval of 0 to 1, calculate the basic probability allocation value of each quality index corresponding to each type of prediction result, and calculate the conflict coefficient between different prediction results. A preset conflict coefficient threshold is set. When the calculated conflict coefficient is lower than the threshold, the basic probability allocation value is fused using conventional evidence theory fusion rules. When the calculated conflict coefficient is higher than or equal to the threshold, the basic probability allocation value of the high conflict prediction result is corrected by the weighted correction mechanism, and then the evidence theory fusion rule is used for fusion calculation. The final quality prediction result is determined based on the fusion calculation results, and the confidence level of the prediction result is calculated.

[0011] Furthermore, the specific steps for the improved SHAP value analysis module to locate key control nodes are as follows: Based on the fusion feature vector, the characteristics of various process parameters contained in the fusion feature vector are clarified. Process constraint weights are introduced into the SHAP value analysis model. The process constraint weights of each link in the production of oryzanol are set according to the process importance, parameter adjustability and sensitivity of the impact on quality indicators. The process constraint weights of the extraction and purification links are higher than those of the raw material pretreatment and finished product preparation links. An improved SHAP value analysis model is developed by incorporating fused feature vector inputs into process constraint weights to calculate the SHAP value for each process parameter. All process parameters are sorted in descending order based on the absolute value of their SHAP values. The process parameters with the highest impact are selected as key control nodes. These key control nodes include extraction temperature, feed-to-liquid ratio, crystallization cooling rate, drying vacuum degree, solvent ratio, extraction time, stirring speed, alkali addition amount, crystallization settling time, and filtration accuracy. For each identified key control node, the parameter sensitivity range is determined by combining its corresponding process constraint weight and SHAP value sign. When the actual value of the parameter exceeds the sensitivity range, the system automatically marks it as a high-risk parameter. The analysis results of key control node information, SHAP values ​​of each node, and parameter sensitivity ranges are pushed to the federated reinforcement learning optimization module in real time.

[0012] Furthermore, the specific steps for generating the optimization strategy using the federated reinforcement learning method are as follows: It receives information on key control nodes, parameter sensitivity ranges, and quality prediction results, uses the quality prediction results as the state input for reinforcement learning, and uses the parameter adjustment amount of key control nodes as the action space. When in the single-factory optimization stage, a reinforcement learning agent is constructed based on key control node information, parameter sensitivity range and quality prediction results. A three-dimensional reward function is set to achieve quality target, lowest energy consumption and highest production efficiency. Quality target is determined by the fact that the purity of finished product, sterol residue and solvent residue all meet the preset product quality standards. The reinforcement learning agent and the digital twin are trained interactively. The agent outputs parameter adjustment actions based on the current quality prediction results and the status of key control nodes. The digital twin simulates the production process corresponding to the adjustment action and returns the quality change results and reward value. The agent updates the strategy based on the reward value. After multiple rounds of iterative training until the model converges, the optimal real-time optimization strategy for a single factory is obtained. When in the multi-factory collaboration stage, each factory acts as a local client, independently training reinforcement learning agents based on local production data, and only uploading the trained model parameters to the cloud federated server through encrypted transmission. The cloud-based federated server uses a federated averaging algorithm to aggregate the model parameters uploaded by each client. It dynamically allocates weights based on the production scale, data quality, and model training accuracy of each factory, calculates the weighted average of all client model parameters, and generates a globally optimized model. The global optimization model is encrypted and then distributed to each client. Each client updates its local agent parameters based on the global optimization model and outputs the optimal process parameter adjustment strategy for the current production scenario.

[0013] Furthermore, after generating the optimization strategy, it is distributed to the edge node control device, linking the blockchain and digital twin to complete quality traceability and compliance execution.

[0014] Compared with existing technologies, the method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process provided by this invention has the following beneficial effects: This invention performs differentiated preprocessing on multiple types of data through edge nodes, generates supplementary small sample scenarios by combining cloud-based virtual data, constructs a multi-scale feature matrix, and deeply integrates it through an adaptive multi-scale attention module. It enhances the accuracy of quality prediction by relying on digital twins and improved evidence theory, accurately locates key control nodes through improved SHAP value analysis, and achieves unified real-time optimization of single factories and collaborative optimization of multiple factories through federated reinforcement learning. Finally, it completes full-process quality control through edge node device control and blockchain digital twin fusion traceability, forming a complete data flow and logical closed loop. This effectively improves the stability and consistency of oryzanol production quality, enhances the targeting and accuracy of quality optimization, ensures data privacy and security during multi-factory collaboration, enables rapid location and traceability of quality problems, meets the compliance and regulatory requirements of the food and pharmaceutical industries, and has significant industrial application value. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method for multi-source heterogeneous data fusion analysis and quality optimization in the entire production process of oryzanol in this invention. Detailed Implementation

[0016] In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "middle," and "inner," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, it should be noted that unless otherwise explicitly specified and limited, the terms "installed," "connected," and "joined" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention through specific circumstances.

[0017] The embodiments of the present invention will be described below with reference to the accompanying drawings and related examples. The embodiments of the present invention are not limited to the following examples, and the present invention relates to the relevant necessary components in this technical field, which should be regarded as well-known technology in this technical field and can be known and mastered by those skilled in this technical field.

[0018] See Figure 1 This invention provides a method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process, comprising the following steps: Step S100: Based on the entire production process of oryzanol, deploy edge acquisition nodes to collect multi-source heterogeneous data and transmit it to the edge nodes; Edge acquisition nodes are distributed data acquisition terminals deployed on the production site. They integrate sensor interfaces, data caching, and preliminary transmission functions, enabling them to acquire various types of data at close range and reduce transmission latency, unlike traditional centralized acquisition architectures. Multi-source heterogeneous data refers to data sets originating from different acquisition objects and possessing different data types. In this invention, this specifically includes five categories of data: raw material attributes, process parameter time series, equipment operation, environmental sensing, and quality inspection. Data types encompass structured data, time series data, and spectral data. Edge nodes are computing units deployed on the production site, possessing local data processing and control execution capabilities. They are used to perform differentiated preprocessing of multi-source heterogeneous data, receive and execute optimization strategies to regulate on-site equipment, and reduce dependence on the cloud.

[0019] In the four core stages of oryzanol production—raw material pretreatment, extraction, purification, and finished product preparation—edge acquisition nodes are deployed. Each node integrates a sensor interface, a 5G industrial module, and a data caching unit, connecting to the corresponding detection equipment (such as near-infrared spectrometers, temperature sensors, and HPLC) and production equipment (such as extraction kettles and centrifuges) at the nearest point in the process. Data is collected including raw material attribute data (obtained via near-infrared spectrometers and moisture analyzers), process parameter time-series data (collected in real-time via temperature, pressure, and flow sensors), equipment operation data (obtained via vibration and current sensors), environmental sensing data (obtained via temperature, humidity, and dust sensors), and quality testing data (obtained via online HPLC, GC, and offline detection equipment). All collected data carries a three-dimensional tag with the stage, timestamp, and equipment number. The stage tag identifies the data source (e.g., extraction stage), the timestamp is accurate to milliseconds, and the equipment number uniquely identifies the data acquisition device, ensuring data traceability. The collected data is transmitted to the edge nodes in real-time via 5G industrial modules. The edge nodes are configured with local caching units to prevent data loss due to network fluctuations, and cached data is automatically retransmitted after network recovery.

[0020] Step S200: Based on multi-source heterogeneous data, edge nodes perform differentiated preprocessing on different types of data, synchronize them to the cloud distributed database, and generate virtual data in the cloud to supplement small sample scenarios, forming an enhanced dataset; Differential preprocessing is a process of applying customized cleaning, standardization, and transformation methods to different types of data. It aims to improve the usability and consistency of different data types and reduce noise interference. The cloud-based distributed database is a data storage system built on a distributed architecture, supporting high-concurrency writes and cross-factory data sharing. It is used to centrally store multi-source heterogeneous data after edge preprocessing, providing a data foundation for virtual data generation and model training. Small sample scenarios refer to situations where the amount of available real data is extremely limited, such as sudden changes in raw materials or the initial stage of equipment failure. It is used to identify special operating conditions requiring additional data augmentation. Virtual data is simulated data generated by algorithms to expand datasets with insufficient real samples, alleviating the problem of insufficient model training data in small sample scenarios. Augmented datasets are comprehensive training datasets composed of real preprocessed data and generated virtual data, serving as the input data source for subsequent feature extraction and model training.

[0021] Step S300: Extract various data features from the augmented dataset to construct a multi-scale feature matrix. After fusion by the adaptive multi-scale attention module, output the fused feature vector. A multi-scale feature matrix is ​​a two-dimensional array containing feature sets across multiple temporal and spatial scales, ranging from local details to global trends. It is used to characterize the dynamic evolution of oryzanol across scales during its production. An adaptive multi-scale attention module is a neural network component that dynamically adjusts the importance weights of features at different scales. It is used to perform nonlinear weighted fusion of the multi-scale feature matrix, enhancing the expressive power of features from key process stages. The fused feature vector is a low-dimensional, highly expressive feature representation formed after deep fusion processing, serving as a unified input for downstream tasks such as quality prediction and key node identification.

[0022] Step S400: Input the fused feature vector into the two data-driven models, and simultaneously input the preprocessed multi-source heterogeneous data into the preset digital twin and mechanism model. By improving the evidence theory, the quality prediction results and confidence levels are fused and output. Data-driven models are predictive models built using statistical learning or machine learning methods. They rely on historical data for training and are used to uncover hidden patterns from historical data to predict product quality indicators. Digital twins are virtual mappings of physical production systems, integrating geometric, behavioral, and state information. They are used to simulate the response behavior of the entire oryzanol production process under current operating conditions, outputting predicted values ​​of process variables for quality assessment. Mechanism models are mathematical models built based on chemical engineering and physical laws, describing raw material conversion and mass and heat transfer processes. They provide process mechanism explanations independent of statistical laws, supplementing the predictive blind spots of data-driven models under abnormal operating conditions. Improved evidence theory is an extension of DS evidence theory, used to handle fusion reasoning of uncertainties and conflicting information. It integrates the outputs of data-driven models and mechanistic models to generate quality prediction results with confidence levels.

[0023] Step S500: Input the fused feature vector into the improved SHAP value analysis module to locate key control nodes, and generate an optimization strategy by combining the quality prediction results with federated reinforcement learning.

[0024] The improved SHAP value analysis module is an optimized version of the SHAP (Shapley Additive Explanations) method, enhancing the accuracy and efficiency of attribution explanations. It is used to analyze the impact of each dimension in the fused feature vector on quality prediction and identify key control nodes. Key control nodes are process steps or equipment parameter points that significantly impact product quality. These nodes are identified through attribution analysis of the fused feature vector by the improved SHAP value analysis module and are used to guide optimization strategies to focus on the most effective intervention points. Federated reinforcement learning is a distributed learning framework combining federated learning and reinforcement learning. It achieves collaborative strategy optimization while protecting local data privacy, generating unified optimization strategies that balance real-time response in a single factory with experience sharing across multiple factories. The optimization strategy is a parameter adjustment plan designed to improve product quality. It is generated by federated reinforcement learning based on quality prediction results and key control nodes and is then distributed to edge nodes to execute specific control actions.

[0025] Taking the collaborative response of multiple factories to raw material batch mutations as an example, the multi-source heterogeneous data fusion analysis and quality optimization method for the entire process of oryzanol production in this invention can be as follows: when the initial content of oryzanol in the rice bran oil raw material entering a factory experiences abnormal fluctuations, the edge acquisition node captures the change in real time and completes attribute data preprocessing through the edge node; after receiving the data in the cloud, it identifies it as a small sample scenario, starts a virtual data generation program to expand similar working condition samples, and forms an enhanced dataset for updating the model; after multi-scale feature fusion, it is input into the dual-track prediction system, improves the evidence theory, integrates data-driven and mechanism model outputs, and gives a warning of the current batch quality deviation; the improved SHAP analysis points out that the extraction temperature range is a key control node; federated reinforcement learning calls on the historical optimization experience of other factories under similar raw material conditions, generates a joint heating compensation strategy, and distributes it to the local edge node; while executing the control, all operations and digital twin state changes are written to the blockchain for subsequent audit traceability.

[0026] This invention breaks through the bottlenecks of traditional oryzanol production quality control, such as scattered data, delayed prediction, blind optimization, and lack of privacy protection. Through the innovative integration of technologies such as edge computing, attention fusion, digital twins, and federated learning, it achieves real-time, precise, collaborative, and compliant quality control, fully meeting the stringent requirements of the food and pharmaceutical industries for product quality stability and traceability. It has significant industrial application value and promotion prospects.

[0027] In one embodiment of the present invention, the multi-source heterogeneous data includes raw material properties, process parameter time series, equipment operation, environmental perception, and quality inspection data. The specific steps for the edge node to perform differentiated preprocessing on different types of data are as follows: Step S201: Receive raw material attributes, process parameter timing, equipment operation, environmental perception and quality inspection data transmitted by the edge acquisition node, including the carrying link, timestamp, equipment number, and three-dimensional tag. The data is divided into three categories according to data type: structured data, time-series data and spectral data. Edge nodes are equipped with industrial-grade data receiving modules, receiving full production data transmitted from edge acquisition nodes via 5G industrial modules or industrial Ethernet. The data transmission protocol adopts the MQTT / OPCUA industrial standard protocol to ensure real-time and reliable data transmission. The integrity of the 3D tags for the received data's stage, timestamp, and device number is verified, rejecting data with missing tags or incorrect formats, ensuring that each piece of data is traceable to a specific production scenario. Based on the data's structural characteristics and acquisition method, a rule engine automatically classifies the data, classifying numerical, fixed-format discrete data (such as raw material weight and solvent addition) as structured data; classifying continuously changing sequence data (such as extraction temperature and stirring speed) as time-series data; and classifying continuous spectral data acquired by near-infrared or Raman spectrometers as spectral data.

[0028] Step S202: Dynamic standardization is applied to structured data, and the mean and standard deviation of the data are updated in real time to adapt to the fluctuations of raw material batches and eliminate differences in data dimensions. A sliding window algorithm is used, with the structured data of the three most recent production batches as the statistical window. The mean μ and standard deviation σ of each data dimension are calculated in real time. The window is automatically updated as new batches of data are generated. The structured data of the current batch is transformed according to the formula x'=(x-μ) / σ, where x is the original data and x' is the standardized data. The reasonable range of the standardized data is set to [-3,3]. Data outside this range is temporarily stored in the abnormal cache area for further judgment by the subsequent dual screening mechanism.

[0029] Step S203: First, adaptive wavelet threshold denoising is applied to the time series data. The threshold size is dynamically adjusted according to the data noise intensity to reduce the proportion of high-frequency interference signals. Then, the window size is dynamically adjusted according to the data frequency for feature extraction. Short window is used to extract features for high-frequency equipment operation data, and long window is used to extract features for low-frequency environmental perception data. The window size is set according to the data acquisition frequency and process response characteristics. The db4 wavelet is selected as the wavelet basis, and the time series data is decomposed into three levels of wavelet decomposition to obtain the low-frequency signal of the approximate coefficients and the high-frequency noise of the detail coefficients. An improved heuristic threshold calculation method is adopted to dynamically generate the threshold λ=σn√(2lnN) based on the noise intensity of the detail coefficients, where σn is the noise standard deviation and N is the data length. The detail coefficients are subjected to soft thresholding, with coefficients less than the threshold set to zero and coefficients greater than the threshold contracted. The data is then reconstructed through inverse wavelet transform to obtain the denoised data. The window type is automatically divided based on the data acquisition frequency. A short window of 5s is set for high-frequency data with a acquisition frequency ≥5Hz, such as equipment vibration and real-time temperature, and a long window of 30s is set for low-frequency data with a acquisition frequency <5Hz, such as workshop humidity. Statistical features such as mean, variance, peak value, valley value, trend, and slope are extracted from the time series data in each window to form a time series feature vector with fixed dimensions.

[0030] Step S204: Perform MSC correction, adaptive smoothing and sparse dimensionality reduction on the spectral data in sequence. MSC correction is used to eliminate baseline drift and scattering interference. The polynomial order of adaptive smoothing is dynamically selected according to the complexity of the spectral signal. Sparse dimensionality reduction significantly compresses the amount of data while retaining the core information. The average spectrum of all spectral data is calculated as the reference spectrum. Linear regression is performed on each original spectrum and the reference spectrum to obtain the regression coefficient slope and intercept. The baseline drift and scattering interference of the original spectrum are corrected using the regression coefficients, with the formula Acorr = (Araw - b) / a, where Araw is the absorbance of the original spectrum, a is the slope, b is the intercept, and Acorr is the corrected spectrum. The Savitzky-Golay smoothing algorithm is used, and a polynomial fit is performed on the corrected spectrum through an 11-point sliding window. Based on the complexity of the spectral signal... The polynomial order is automatically selected based on the number of characteristic peaks and the steepness of the peak shape. A second-order polynomial is selected when there are ≤5 characteristic peaks, and a third- or fourth-order polynomial is selected when there are >5 characteristic peaks. The smoothed spectral data is standardized and preprocessed to ensure that the mean and variance of each wavelength point data are 0 and 1. A PCA model is constructed to calculate the variance contribution rate of each principal component. The principal components with a cumulative variance contribution rate ≥95% are selected as the eigenvectors after dimensionality reduction. The principal component coefficients are sparsified by L1 regularization to remove wavelength point features with extremely low contribution, further compressing the data volume.

[0031] Step S205: Use a dual screening mechanism to remove abnormal data for all types of data. First, use the 3σ criterion for preliminary screening, and then verify the process rules according to the reasonable range set in the oryzanol production technical specifications. Data that exceeds the range and whose duration meets the abnormal judgment criteria is judged as abnormal. Calculate the mean μ and standard deviation σ of various preprocessed data, and determine the statistically reasonable interval as [μ-3σ, μ+3σ]. Traverse all data and mark data exceeding this interval as suspected outliers, including them in the secondary verification scope. Based on the oryzanol production technical specifications, establish the reasonable process range for each data dimension, such as the extraction temperature process range of 55-65℃ and the raw material moisture content process range of 0.1%-0.5%. For suspected outliers, determine the reasonableness of the process. If the data exceeds the process range and the duration is ≥10s to avoid misjudgment due to instantaneous fluctuations, it is finally determined as outlier and removed. If it only exceeds the statistical range but is within the process range, the data is retained and the statistical parameters are updated.

[0032] Step S206: Complete the missing datasets after screening using the process correlation interpolation method, and calculate the completion value based on the process correlation between the missing data stage and the preceding and following processes.

[0033] A mask matrix is ​​used to mark the missing locations in the dataset, clarifying the data type and corresponding timestamp of the production process to which the missing data belongs. Based on the oryzanol production process flow chart, a correlation network of data in each process is constructed. The correlation weight between the missing data and related data in the preceding and following processes is calculated. Directly related data, such as adjacent timestamp data of the same equipment, is weighted at 0.6. Indirectly related data, such as related parameters of upstream and downstream equipment, is weighted at 0.3. Unrelated data is weighted at 0.1. A weighted average interpolation method is used to calculate the missing data by weighting the effective data in the neighborhood of the missing data according to the correlation weight, and the fill value is obtained by formulating xfill=Σ(wixi), where wi is the correlation weight of the i-th neighborhood data, xi is the effective value of the i-th neighborhood data, k is the number of neighborhood data, and Σwi=1.

[0034] It should be noted that the specific steps for generating virtual data in the cloud to supplement the small sample scenario are as follows: Step S207: Receive preprocessed data synchronized from edge nodes to the cloud distributed database, classify the preprocessed data according to production scenarios, identify small sample scenarios such as sudden changes in raw material properties, abnormal equipment parameters, and adjustments to process conditions, and determine the core data characteristics and distribution patterns of small sample scenarios. The system receives preprocessed data synchronized from edge nodes to a cloud-based distributed database via an industrial communication network. This preprocessed data has undergone dynamic standardization of structured data, denoising and feature extraction of time-series data, spectral data correction, smoothing and dimensionality reduction, outlier removal, and missing data completion. An unsupervised clustering algorithm is used to automatically classify the received preprocessed data according to production scenarios. During clustering, raw material attribute parameters, process parameter ranges, equipment operating status, and environmental conditions are used as clustering features, with reasonable cluster numbers and distance thresholds set. By comparing the proportion of data volume in each cluster to the overall data volume, and combining this with data distribution characteristics, three small-sample scenarios—raw material attribute mutations, equipment parameter anomalies, and process condition adjustments—are identified. Raw material attribute mutation scenarios correspond to datasets where the content of key raw material components deviates from the normal range; equipment parameter anomaly scenarios correspond to datasets where equipment operating parameters exceed set thresholds; and process condition adjustment scenarios correspond to datasets after proactive changes to key production process parameters. For the identified small-sample scenarios, statistical analysis methods are used to calculate the mean, variance, covariance, and other statistics within the scenario to determine the core data characteristics and distribution patterns of each small-sample scenario.

[0035] Step S208: Based on the core data features of small sample scenarios, construct a generative adversarial network model. Use real preprocessed data as training samples to train the generator and discriminator of the generative adversarial network. The generator learns the distribution features of real data, and the discriminator distinguishes and verifies the generated data from the real data. Based on the core data features of small-sample scenarios, a generative adversarial network (GAN) model is constructed, consisting of a generator and a discriminator. The generator employs a deep convolutional neural network (CNN) structure. Its input is a random noise vector, and its output is virtual data with the same dimensionality as the real preprocessed data. The network has 5 to 8 layers, with the number of kernels in each layer dynamically adjusted according to the data dimensionality. The ReLU activation function is used, and the output layer uses a Sigmoid function to map the output value to a numerical range consistent with the real data. The discriminator uses a combination of a CNN and a fully connected network structure. Its input is either real preprocessed data or virtual data output by the generator, and its output is a probability judgment of the data's authenticity. The discriminator has 3 to 5 layers, uses the LeakyReLU activation function, and the output layer uses a Sigmoid function to output probability values. Using the real preprocessed data received in step one as training samples, an alternating training strategy is adopted to train the generator and discriminator. That is, the generator is fixed first to train the discriminator so that the discriminator can accurately distinguish between real data and virtual data. Then, the discriminator is fixed to train the generator so that the virtual data generated by the generator can fool the discriminator. During the training process, an adaptive moment estimation optimizer is used to adjust the network parameters and set an appropriate learning rate and number of iterations until the model converges, that is, the discriminator's recognition accuracy of real data and virtual data stabilizes at about 50%.

[0036] Step S209: Generate virtual data by training a mature generative adversarial network. The generated virtual data must be consistent with the real data of the corresponding small sample scenario in terms of feature distribution and process logic. Random noise vectors meeting preset dimensional requirements are input into a well-trained generator. The generator performs convolution, pooling, and activation operations on the random noise vectors through forward propagation, outputting virtual data with the same dimensionality as the real preprocessed data. During generation, random noise vectors corresponding to different types of small sample scenarios are input to ensure that the generated virtual data matches the core data features of the corresponding small sample scenario. The generated virtual data undergoes preliminary screening, removing virtual data whose values ​​exceed the reasonable range of physical meaning, such as indicators in raw material attribute data where the content cannot be negative, and parameter values ​​in process parameter data that exceed the operating limits of the equipment.

[0037] Step S210: Verify the rationality of the generated virtual data, remove abnormal virtual data that does not conform to the logic of the oryzanol production process, and retain valid virtual data that conforms to the actual production. A rule library for the production process of oryzanol was constructed. This library encompasses various aspects, including raw material processing constraints, extraction process constraints, purification process constraints, and finished product preparation process constraints. Specifically, it includes constraints related to the law of conservation of mass, the law of conservation of energy, equipment operating parameters, limit constraints, and logical association constraints of process parameters. A rule matching algorithm was used to verify each virtual data generated in step three. The various indicators in the virtual data were compared with the corresponding rules in the rule library to determine whether the virtual data conformed to the process logic of oryzanol production. For example, it verified whether the ratio of solvent addition to raw material input in the extraction process met the process requirements, and whether the logical relationship between cooling rate and crystal purity in the crystallization process was reasonable. Virtual data that did not conform to the process constraint rules was marked as abnormal and removed, while valid virtual data that conformed to the process logic was retained.

[0038] Step S211: Integrate the effective virtual data with the real preprocessed data in the cloud-distributed database to form an enhanced dataset covering the entire scenario.

[0039] Data stitching technology is used to integrate effective virtual data with real preprocessed data. During the integration process, the structured format of the data is maintained consistently to ensure accurate correspondence between tags such as timestamps and equipment numbers in the production process. The integrated data undergoes data normalization to map all data values ​​to the same numerical range, eliminating dimensional differences between different data types. The commonly used normalization method is min-max normalization. The normalized dataset is then subjected to integrity verification to ensure that there are no missing, duplicate, or outlier values, ultimately forming an augmented dataset covering both routine production scenarios and various small-sample scenarios.

[0040] In one embodiment of the present invention, the specific steps for constructing the multi-scale feature matrix are as follows: Step S301: Extract three types of raw data—structured data, time-series data, and spectral data—from the augmented dataset generated in the cloud, while retaining the production process and timestamp labels carried by the data; A distributed data extraction tool was used to filter and extract three types of raw data—structured data, time-series data, and spectral data—from an augmented dataset generated in the cloud. During the extraction process, a data label matching algorithm was used to retain the production process identifier and timestamp label carried by each data point, ensuring that the spatiotemporal correspondence between the data and the production process was not lost.

[0041] Step S302: Based on the three types of raw data, a 1DCNN network is used to perform convolution operations on the short-time scale data to extract local temporal features and instantaneous change features. The short-time scale data includes real-time parameters of equipment operation and instantaneous data of the process. Based on three types of raw data, a data classifier is used to filter out short-timescale data, which is then input into a 1DCNN network for convolution operations. By setting an appropriate convolution kernel size and stride, local feature capture and feature mapping are performed on the data, outputting the local temporal features and instantaneous change features of the short-timescale data. The short-timescale data refers to data with a short time span and high frequency of change, specifically including real-time parameters of equipment operation and instantaneous data of the process. Real-time parameters of equipment operation include instantaneous operating status parameters of equipment such as centrifuge speed and dryer vacuum degree, while instantaneous data of the process includes dynamic parameters during the process execution, such as instantaneous fluctuations in temperature inside the extraction vessel and instantaneous changes in solvent mixing flow rate. 1DCNN, or one-dimensional convolutional neural network, is a deep learning model specifically designed to process one-dimensional sequential data. Through the sliding convolution operation between the convolutional kernel and the input sequence, it can effectively capture local correlation features in sequential data and has parameter sharing characteristics, which can reduce model complexity. Local temporal features refer to the change patterns of short-term data within a local time window, while instantaneous change features refer to the sudden change trend and magnitude of data at a certain moment.

[0042] Step S303: Based on the three types of raw data, a bidirectional LSTM network is used to perform sequence learning on the long-term data to extract long-range dependencies and trend features in the time dimension. The long-term data includes raw material batch attribute data, finished product quality batch detection data, and process parameter trend data. Based on three types of raw data, a time span filtering algorithm was used to separate long-term data, which was then input into a bidirectional LSTM network for sequence learning. Forward and backward LSTM units were used to learn features from the forward and backward time series of the data, respectively. A feature concatenation layer then fused the outputs of the two types of units to extract the long-range dependencies and trend features of the long-term data in the time dimension. Long-term data refers to data with a long time span and low frequency of change, specifically including raw material batch attribute data, finished product quality batch inspection data, and process parameter trend data. Raw material batch attribute data covers batches of raw materials, such as average oryzanol content and average moisture content. Secondary statistical data includes batch quality testing data for finished products, covering the purity and sterol residue of each batch; process parameter trend data covers the overall trend of extraction temperature and the stable change pattern of stirring speed within a certain production stage; bidirectional LSTM networks are an improved form of long short-term memory networks. By introducing forward and backward recurrent structures, they can simultaneously capture the dependencies of sequence data in both forward and backward time dimensions, overcoming the limitation of traditional LSTM networks that can only learn sequence features in one direction; long-range dependencies refer to the mutual influence of data over a long time span, and change trend characteristics refer to the overall change pattern of data in the time dimension, such as rise, fall, and stability.

[0043] Step S304: Based on the three types of original data, a CNN network is used to perform multi-layer convolution and pooling operations on the spectral data to extract feature peaks, feature valleys and texture distribution features; Based on three types of raw data, spectral data is selected and input into a CNN network. The data undergoes alternating operations through multiple layers: input layer, convolutional layer, activation layer, and pooling layer. The convolutional kernels extract local texture features from the spectral data, the activation layers introduce nonlinear transformations to enhance feature representation, and the pooling layers downsample and compress the convolutional features. The final output is the characteristic peaks, valleys, and texture distribution features of the spectral data. A CNN, or Convolutional Neural Network, is a deep learning model with locally connected weight-sharing characteristics, exhibiting significant advantages in feature extraction from two-dimensional or high-dimensional data such as image and spectral data. Characteristic peaks refer to the peaks in the spectral curve with higher intensity than the surrounding area, corresponding to the characteristic absorption peaks of specific components of a substance. Characteristic valleys refer to the valleys in the spectral curve with lower intensity than the surrounding area, also corresponding to specific component information of a substance. Texture distribution features refer to the spatial distribution characteristics of the peaks and valleys in the spectral curve, such as their arrangement and density. Step S305: Perform vector normalization on the short-scale local temporal features, long-scale long-range dependency features, and spectral texture features respectively, and construct a multi-scale feature matrix according to the correspondence between production links and timestamps.

[0044] Vector standardization algorithms are used to standardize the extracted short-scale local temporal features, extracted long-scale long-range dependency features, and extracted spectral texture features respectively. By calculating the mean and standard deviation of each feature vector, the values ​​of all feature vectors are mapped to a unified numerical range. A feature matching mechanism is established based on production link identifiers and timestamp labels. The three types of standardized feature vectors are associated one by one according to the correspondence of production links and the synchronization relationship of timestamps, and combined to form a multi-scale feature matrix with unified dimensions and regular structure. Vector standardization is the process of converting feature vectors with different dimensions and numerical ranges into standardized vectors with unified dimensions and numerical ranges. Commonly used methods include ZScore standardization. Its core purpose is to eliminate the differences in dimensions and numerical ranges between different features and to prevent a certain feature from dominating the subsequent fusion process due to its excessively large numerical magnitude. In one embodiment of the present invention, the specific steps of the adaptive multi-scale attention module fusion are as follows: Step S306: Based on the multi-scale feature matrix, determine the total number of feature categories contained in the matrix and the types of quality indicators to be predicted, and calculate the correlation strength between each type of feature and each quality indicator through the gating unit; After receiving the constructed multi-scale feature matrix, two core operations are completed through matrix dimension parsing and feature label recognition. First, the column dimensions of the multi-scale feature matrix are traversed to count the category identifiers of different feature vectors, clarifying the total number of feature categories, such as short-scale local time-series features, long-scale long-range dependency features, and spectral texture features, contained in the matrix. Second, in conjunction with the quality control requirements of oryzanol production, the types of quality indicators to be predicted are clarified, specifically including three core quality indicators: finished product purity, sterol residue, and solvent residue, ensuring that subsequent correlation strength calculations directly correspond to the quality indicators.

[0045] Based on a multi-scale feature matrix, a fully connected layer is used to construct a gated unit, which includes an input layer, a hidden layer, and an output layer. The input layer receives the feature vector of the i-th class. With the j-th quality index parameter The hidden layer strengthens the association between features and quality indicators through nonlinear transformation (using the ReLU activation function), and the output layer outputs a quantified value of the association strength. In specific computation, first... and Perform vector concatenation, then combine it with the weight matrix of the gated unit. Perform matrix multiplication and finally add the bias terms. This yields the quantified value of the correlation strength.

[0046] The calculation formula is as follows: , in, The value representing the correlation strength between the i-th feature and the j-th quality indicator is the core quantitative result reflecting the degree of correlation between the two, and its value range is the real number domain. The weight matrix of the gated unit is a two-dimensional matrix with the number of rows matching the dimension of the concatenated vectors and the number of columns representing the number of neurons in the hidden layer. It is used to perform linear transformations on the concatenated vectors. This represents the i-th type of feature vector, which is derived from the multi-scale feature matrix and contains quantized information of local temporal features, long-range dependent features, or spectral texture features, depending on the feature category. The j-th quality indicator parameter represents the quantitative standard or target value of the purity, sterol residue, or solvent residue of the finished product. This represents a vector concatenation operation, which combines feature vectors. With quality index parameters Combine them column-wise into a new high-dimensional vector. The bias term representing the gated unit is a one-dimensional vector with the same dimension as the number of neurons in the hidden layer. It is used to compensate for the output offset after linear transformation.

[0047] Step S307: Use the Sigmoid activation function to process the correlation strength quantification value and calculate the process correlation weight between various features and each quality indicator; First, quantize the correlation strength value output by the gating unit. Input the Sigmoid activation function and map it to the interval between 0 and 1 to obtain the normalized association strength value; then count the total number of feature categories n involved in the weight calculation, and for all features with k ranging from 1 to n, the corresponding... The values ​​are summed to obtain the weighted normalized denominator; finally, the values ​​corresponding to the i-th type of feature are used. Dividing the value by the summation result yields the process relevance weight between the i-th feature and the j-th quality indicator. .

[0048] The calculation formula is as follows: , in, The value represents the process relevance weight between the i-th feature and the j-th quality indicator, ranging from 0 to 1. The sum of the weights of all features corresponding to the same quality indicator is 1. This represents the Sigmoid activation function, used to quantize the correlation strength. Normalized to the interval 0 to 1. n represents the total number of feature categories involved in the weight calculation, i.e., the total number of feature types contained in the multi-scale feature matrix. Represents all features corresponding to k from 1 to n Summation operation on the values; This represents the quantified value of the correlation strength between the k-th feature and the j-th quality indicator, and the calculation logic is the same as... Consistent.

[0049] Step S308: Iteratively optimize the weight matrix and bias terms of the gated unit using the gradient descent algorithm, and perform weighted operations on the various eigenvectors in the multi-scale feature matrix with the corresponding process-related weights to obtain the weighted eigenvectors. With the accuracy of process-related weights as the optimization objective, a loss function (using cross-entropy loss) is constructed. The loss function value reflects the deviation between the importance of features and the calculated weights in the actual process scenario. A gradient descent algorithm is used to iteratively update the weight matrix and bias terms of the gated unit with a preset learning rate (dynamically adjusted based on data features, with an initial learning rate of 0.001). During each iteration, the partial derivatives of the loss function with respect to the weight matrix and bias terms are calculated, and the parameter values ​​are adjusted according to the direction of the partial derivatives until the loss function value converges to a preset threshold (less than 0.0001), at which point parameter updates cease. Each class of feature vectors in the multi-scale feature matrix is ​​traversed, and element-wise weighting operations are performed on the corresponding feature vectors according to the process-related weights (each element of the feature vector is multiplied by the weight value). For the same quality indicator, the weighted results of all classes of feature vectors are summarized to obtain the weighted feature vector corresponding to that quality indicator; this process is repeated until the weighted feature vectors corresponding to all quality indicators are obtained.

[0050] Gradient descent is a commonly used parameter optimization algorithm. It calculates the gradient direction of the loss function and gradually adjusts the parameters along the negative gradient to minimize the loss function value. It features fast convergence and stable optimization results. The loss function measures the deviation between the model's predictions and the actual results; here, it quantifies the difference between the weights related to process features and the importance of actual features, and is a core criterion for evaluating parameter optimization. The learning rate controls the step size of each parameter update. Too large a step size can lead to parameter oscillations and non-convergence, while too small a step size can result in slow convergence. A dynamic learning rate can balance convergence speed and optimization accuracy.

[0051] Step S309: Input all weighted feature vectors into the Transformer encoder, perform deep feature fusion through multi-head attention mechanism, and output fused feature vector.

[0052] The weighted feature vectors corresponding to all quality metrics are input into the Transformer encoder, which contains six encoding layers. Each layer consists of a multi-head attention mechanism and a feedforward neural network. First, the multi-head attention mechanism (with eight attention heads) captures the interaction relationships between cross-modal features in parallel, with each attention head focusing on a different feature interaction dimension. Then, the output of the multi-head attention mechanism is input into the feedforward neural network (using the ReLU activation function) for non-linear feature transformation and dimensionality unification. Finally, through layer normalization, a fused feature vector with unified dimensions and rich information is output, with the dimension set to 512.

[0053] In one embodiment of the present invention, the specific steps for constructing the digital twin are as follows: Step S401: Collect physical entity information of the entire production process of oryzanol, including the physical structure parameters, installation layout, and operating principle of raw material pretreatment equipment, extraction equipment, purification equipment, and finished product preparation equipment, as well as the process flow diagram, process parameter range, and material transfer path of each link. A multi-dimensional data acquisition scheme was adopted for raw material pretreatment equipment, extraction equipment, purification equipment, and finished product preparation equipment. A laser 3D scanner was used to collect physical structural parameters such as the equipment's external dimensions, mounting hole positions, and component layout, with an accuracy controlled within 0.1 mm. By combining equipment manual analysis with on-site measurements, operational information such as the equipment's operating principles, power transmission paths, and control logic was recorded. A process flow diagram drawing tool was used to clarify the process logic of each production stage, including the material input / output sequence, reaction conditions, and control nodes. Simultaneously, sensor measurements and process document verification were used to determine the range of process parameters for each stage, including the allowable value ranges for parameters such as temperature, pressure, time, and flow rate. A material tracking method was used to record the material transmission path from raw material input to finished product output, clarifying the material connection relationships between each piece of equipment.

[0054] Step S402: Based on the 3D simulation engine, build a digital twin that maps to the actual production scene in a 1:1 ratio, and replicate the physical structure, operating parameters and process logic of various production equipment according to the physical entity information; A 3D simulation engine supporting industrial-grade simulation was selected, and the physical structure parameters of the equipment were imported. A 3D model was created at a 1:1 scale to reproduce details such as the appearance, structure, internal component layout, and connection methods of the equipment. Based on the equipment's operating principles and process logic, the virtual operation logic of the equipment was implemented through simulation scripts, ensuring that the start-stop status, operating parameters, adjustment, and response of the virtual equipment were consistent with those of the physical equipment. The virtual equipment in each stage was arranged according to the installation layout and material transfer path determined in step one, constructing a complete virtual scenario of the entire production process. At the same time, a parameter mapping relationship between the virtual equipment and the physical equipment was established to ensure a one-to-one correspondence between the parameters of the physical equipment and the parameters of the virtual equipment.

[0055] Step S403: Establish a two-way mapping mechanism between the physical entity and the digital twin. Through edge nodes, synchronize the real-time operation data and process parameter data of the physical entity to the digital twin, and transmit the simulation results of the digital twin back to the control terminal of the physical entity. An industrial-grade communication link is established between the edge node and the digital twin system. Industrial Ethernet or 5G industrial modules are used to achieve real-time data transmission. The real-time operating data and process parameters of the physical entity are synchronized to the virtual twin at a frequency of once per second. The transmitted data is converted into a parameter format that the virtual twin can recognize through a data parsing interface, driving the corresponding equipment in the virtual twin to update its operating status synchronously. A simulation result output interface is set in the virtual twin system to transmit the simulation data of the virtual twin, such as predicted quality indicators and equipment operating trends, back to the control terminal of the physical entity at a preset frequency, with the transmission delay controlled within 500 milliseconds.

[0056] Step S404: Embed the extraction kinetics mechanism model and the crystallization thermodynamics mechanism model into the digital twin, configure the digital twin with the process parameter simulation and deduction function, so that it can receive the pre-processed process parameters, simulate the quality change trend under different working conditions, and continuously compare and calibrate with historical actual production data.

[0057] The extraction kinetics mechanism model and the crystallization thermodynamics mechanism model are integrated into the digital twin system through a programming interface. The extraction kinetics mechanism model is constructed based on reaction kinetic equations, and its input parameters include process parameters such as extraction temperature, feed-to-liquid ratio, and stirring speed. The output parameter is the oryzanol extraction efficiency. The crystallization thermodynamics mechanism model is constructed based on phase equilibrium theory and crystal growth kinetics, and its input parameters include process parameters such as crystallization temperature and cooling rate. The output parameters are crystal purity and particle size distribution. The correlation between the mechanism model parameters and the parameters of the digital twin virtual device is established, so that changes in the virtual device parameters can drive the mechanism model to perform calculations in real time.

[0058] A process parameter input interface is set up in the digital twin system to support the reception of various pre-processed process parameters, including temperature, pressure, time, and flow rate. Based on the embedded mechanism model and virtual equipment operation logic, a working condition simulation algorithm is developed. When process parameters are input, the algorithm drives the virtual equipment to operate according to those parameters. At the same time, the mechanism model is used to calculate the quality change trend and generate simulation data. A historical actual production database is established, and the simulation results are compared with the historical actual production data batch by batch to calculate the simulation error. When the error exceeds the preset range, the mechanism model parameters and simulation algorithm coefficients are automatically adjusted to achieve continuous calibration of simulation accuracy. The calibration frequency is once every 10 batches of production.

[0059] In one embodiment of the present invention, the specific steps for the improved evidence theory fusion output quality prediction result and confidence level are as follows: Step S405: Receive the data-driven prediction results output by two data-driven models. The first data-driven model is the improved gradient boosting tree model, and the second data-driven model is the feature dependency capture model. Simultaneously, receive the simulation prediction results output by the digital twin and the mechanism prediction results output by the mechanism model. The system receives data-driven prediction results from two data-driven models. The first model is an improved gradient boosting tree model, based on a decision tree ensemble learning framework. It introduces an adaptive learning rate to adjust the training step size, iterates through the fused feature vector, and outputs a predicted value for the quality index. The second model is a feature dependency capture model (i.e., the Transformer model). This model captures long-distance dependencies between features through a multi-head attention mechanism, performs sequence modeling on the fused feature vector, and outputs the corresponding quality prediction result. The system also receives simulation prediction results from the digital twin, obtained through simulation calculations based on the input preprocessing parameters and using built-in extraction kinetics and crystallization thermodynamics mechanisms. Finally, the system receives mechanistic prediction results from the mechanistic model, derived from theoretical formulas based on the physicochemical laws of oryzanol production. All prediction results are transmitted to the fusion processing module in numerical form, using an industry-standard data format.

[0060] The improved gradient boosting tree model is an optimized version of the gradient boosting tree model. The gradient boosting tree model iteratively constructs multiple weak decision trees and weights their predictions to reduce prediction error. The improved version enhances the model's adaptability to data fluctuations through mechanisms such as adaptive learning rates. The feature dependency capture model is a deep learning model based on self-attention mechanisms. It can effectively capture the dependencies between different features in the input data, and is particularly adept at handling long sequence data.

[0061] Step S406: Normalize the data-driven prediction results, simulation prediction results and mechanism prediction results respectively, map the numerical range to the interval of 0 to 1, calculate the basic probability allocation value of each quality index corresponding to each type of prediction result, and calculate the conflict coefficient between different prediction results. The data-driven prediction results, simulation prediction results, and mechanism prediction results were normalized using the min-max normalization method, and the calculation formula is as follows: , in, The value is the normalized value. These are the original predicted values. This is the minimum value of the quality index among all prediction results. This represents the maximum value of the quality index among all prediction results. After calculation, the numerical range of all prediction results is uniformly mapped to the interval between 0 and 1.

[0062] Based on the normalized prediction results, the basic probability allocation value for each quality index corresponding to each type of prediction result is calculated using an allocation method based on prediction error. The first step is to calculate the prediction error of each model in historical data. The second step is to use the reciprocal of the error as a weight to weight the normalized prediction values, obtaining the degree of support for each quality index belonging to different value ranges; this degree of support is the basic probability allocation value.

[0063] The conflict coefficient between different prediction results is calculated using the following formula: , in, and These are the basic probability assignment functions for the two types of prediction results. and For different propositions, when and When there is no overlap, the sum of their products is calculated to obtain the conflict coefficient k, which is used to quantify the degree of inconsistency between the two types of prediction results.

[0064] Step S407: Preset conflict coefficient threshold. When the calculated conflict coefficient is lower than the threshold, use conventional evidence theory fusion rules to perform fusion calculation on the basic probability allocation value. The conflict coefficient threshold is determined by combining statistical methods with expert experience. The first step is to collect various prediction results from historical production data, calculate the conflict coefficient under different scenarios, and statistically analyze the distribution characteristics of the conflict coefficient. The second step is to determine the median or 90th quantile of the conflict coefficient as the initial threshold. The third step is to invite experts in industrial control and quality management to adjust the initial threshold based on the process characteristics of oryzanol production. The final determined conflict coefficient threshold ranges from 0.2 to 0.4. The threshold is stored in the configuration file of the fusion processing module and can be dynamically modified according to changes in the actual production scenario.

[0065] When the calculated conflict coefficient is below the threshold, the basic probability assignment values ​​are fused using the conventional evidence theory fusion rule (i.e., Dempster's combination rule), and the calculation formula is as follows: , Where k is the conflict coefficient, and A is any proposition. and Two types of evidence support the proposition. and The basic probability allocation value.

[0066] When multiple pieces of evidence need to be fused, a pairwise iterative fusion method is adopted. The first step is to fuse the basic probability assignment values ​​of the first two pieces of evidence to obtain an intermediate fusion result. The second step is to fuse the intermediate fusion result with the basic probability assignment value of the third piece of evidence. The third step is to iterate in sequence until the fusion of all evidence is completed and the final joint basic probability assignment value is obtained.

[0067] Step S408: When the calculated conflict coefficient is higher than or equal to the threshold, the basic probability allocation value of the high conflict prediction result is corrected by the weighted correction mechanism, and then the evidence theory fusion rule is used for fusion calculation. The first step is to calculate the historical prediction accuracy of each prediction model. The formula for calculating the historical prediction accuracy is as follows: , The second step is to allocate weights based on accuracy. The higher the accuracy, the greater the weight. The weight values ​​range from 0 to 1, and the sum of the weights of all models is 1.

[0068] The third step is to multiply the basic probability assignment value of each model by its corresponding weight to obtain the corrected basic probability assignment value. The correction formula is as follows: , in, The corrected basic probability assignment value. For model weights, These are the original basic probability assignment values.

[0069] For all the corrected basic probability assignments, the Dempster combination rule is used for fusion calculation. The fusion process is consistent with the pairwise iterative fusion method to obtain the final joint basic probability assignments. Step S409: Determine the final quality prediction result based on the fusion calculation result, and calculate the confidence level of the prediction result.

[0070] Based on the joint basic probability assignment values ​​obtained from the fusion operation, the quality index value corresponding to the proposition with the largest joint basic probability assignment value is selected as the final quality prediction result. If multiple propositions have the same joint basic probability assignment value, and all of them are the maximum value, then the average of the quality index values ​​corresponding to these propositions is taken as the final prediction result. The confidence score is equal to the joint basic probability assignment value of the proposition corresponding to the final prediction result, and the calculation formula is as follows: , Where A is the proposition corresponding to the final prediction result, the confidence level is rounded to two decimal places, and it is output together with the quality prediction result to the subsequent optimization stage.

[0071] In one embodiment of the present invention, the specific steps of the improved SHAP value analysis module in locating key control nodes are as follows: Step S501: Based on the fusion feature vector, clarify the characteristics of various process parameters contained in the fusion feature vector, introduce process constraint weights in the SHAP value analysis model, and set the process constraint weights of each link according to the process importance, parameter adjustability and sensitivity of the impact on quality indicators of each link in the production of oryzanol. The process constraint weights of the extraction link and the purification link are higher than those of the raw material pretreatment link and the finished product preparation link. Based on the fusion feature vector output by the adaptive multi-scale attention module, the feature parsing algorithm is used to decompose the fusion feature vector into dimensions, identify the process parameter feature categories and corresponding data dimensions contained in the vector one by one, and clarify the source production link (raw material pretreatment, extraction, purification, finished product preparation) and data type (structured, time-series, spectral derived features) of each process parameter feature. A process constraint weight configuration unit is added to the input layer of the SHAP value analysis model to construct a weight allocation matrix. The weight allocation process adopts the analytic hierarchy process. First, the entire production process of oryzanol is divided into four primary evaluation indicators: raw material pretreatment, extraction, purification, and finished product preparation. Then, for each primary indicator, three secondary evaluation indicators are used for quantitative scoring: process importance (the core degree of influence on product quality), parameter adjustability (the feasibility and operational difficulty of parameter adjustment in actual production), and quality impact sensitivity (the amplitude of quality index fluctuation caused by small parameter changes). After consistency verification, the weight coefficient of each primary indicator is calculated to ensure that the sum of the weight coefficients of the extraction and purification stages is not less than the sum of the weight coefficients of the raw material pretreatment and finished product preparation stages. The final process constraint weight matrix is ​​embedded in the loss function of the SHAP value analysis model and participates in the model's forward calculation and backward propagation process.

[0072] Step S502: Introduce the fused feature vector input into the improved SHAP value analysis model of process constraint weights, and calculate the SHAP value corresponding to each process parameter; The parsed fused feature vectors are converted into input data acceptable to the SHAP value analysis model according to a preset data format. For different types of process parameter features (structured derived features, time-series derived features, and spectral derived features) in the fused feature vectors, corresponding SHAP value calculation methods are adopted: For structured and time-series derived features, the Tree SHAP algorithm based on a tree model is used to calculate the marginal contribution value by traversing the feature decision path; for spectral derived features, the Deep SHAP algorithm is used to approximate the feature contribution value using the gradient information of the deep learning model. During the calculation process, the background dataset of the model uses a standardized feature matrix obtained after feature extraction from the augmented dataset to ensure that the distribution of the background data is consistent with the distribution of the input data. After the calculation is completed, the SHAP value matrix corresponding to each process parameter feature is output, where the row vectors in the matrix correspond to a single sample, and the column vectors correspond to the SHAP value of a single process parameter.

[0073] Step S503: Sort all process parameters in descending order according to the absolute value of SHAP value, and select the process parameters with the highest impact as key control nodes. The key control nodes include extraction temperature, material-liquid ratio, crystallization cooling rate, drying vacuum degree, solvent ratio, extraction time, stirring speed, alkali addition amount, crystallization settling time, and filtration accuracy. The output SHAP value matrix is ​​row-normalized to obtain the relative proportion of each process parameter's SHAP value in each sample. Then, the global average SHAP absolute value of each process parameter is obtained by calculating the mean of the absolute values ​​of all sample SHAP values. This is used as a quantitative evaluation standard for the degree of influence of process parameters on quality indicators. A quicksort algorithm is used to sort all process parameters from largest to smallest based on their global average SHAP absolute value. Based on the process complexity and optimization requirements of oryzanol production, a screening threshold is set, and the top 10 process parameters with the highest global average SHAP absolute value are selected as key control nodes. Among the selected key control nodes, extraction temperature, material-to-liquid ratio, extraction time, and stirring speed correspond to the core parameters of the extraction stage; crystallization cooling rate and crystallization settling time correspond to the core parameters of the purification stage; and drying vacuum degree, solvent ratio, alkali addition amount, and filtration accuracy correspond to the key parameters of the purification and finished product preparation stages, respectively, ensuring that the key control nodes cover the core process stages of the entire production process.

[0074] Step S504: For each key control node identified, the parameter sensitivity range is determined by combining its corresponding process constraint weight and SHAP value sign. When the actual value of the parameter exceeds the sensitivity range, the system automatically marks it as a high-risk parameter. For each key control node, the sign of its corresponding SHAP value is extracted (a positive sign indicates a positive correlation between the parameter and the quality indicator, and a negative sign indicates a negative correlation). Combined with the process constraint weighting coefficients set in step one, a parameter sensitivity interval calculation model is constructed. The model is based on the historical normal operating parameter range of the key control nodes, introducing a SHAP value sign correction factor and a process constraint weighting correction factor. The SHAP value sign correction factor adjusts the bias of the interval based on the sign (positively correlated parameters expand the upper limit interval, negatively correlated parameters shrink the upper limit interval), and the process constraint weighting correction factor adjusts the width of the interval based on the weight (the larger the weight, the narrower the interval, and the stricter the parameter control). By statistically analyzing the parameter distribution of key control nodes corresponding to historical qualified products, and adjusting the initial interval using the correction factors, the parameter sensitivity interval for each key control node is finally determined. Parameters whose actual values ​​exceed this interval are marked as high-risk, generating a high-risk parameter warning list.

[0075] Step S505: Push the analysis results of key control node information, SHAP values ​​of each node and parameter sensitivity range to the federated reinforcement learning optimization module in real time.

[0076] A data transmission interface is constructed, employing a message queue mechanism to facilitate data interaction between the improved SHAP value analysis module and the federated reinforcement learning optimization module. Data transmission uses JSON format and includes core data fields such as the identifier of key control nodes, the global average absolute SHAP value of each node, the SHAP value sign, the parameter sensitivity range, the current actual value, and risk status markers. A data transmission trigger mechanism is set up so that the data push process is automatically triggered after the analysis module completes a round of analysis of key control nodes. The message queue caches and verifies the transmitted data to ensure data integrity and transmission stability. After receiving the data, the federated reinforcement learning optimization module extracts core information through the data parsing unit and stores it in the optimization module's parameter database, providing real-time data support for policy generation of the reinforcement learning agent.

[0077] In one embodiment of the present invention, the specific steps of generating the optimization strategy through federated reinforcement learning are as follows: Step S506: Receive information on key control nodes, parameter sensitivity ranges, and quality prediction results. Use the quality prediction results as the state input for reinforcement learning and the parameter adjustment amount of key control nodes as the action space. The system receives key control node information, parameter sensitivity ranges, and quality prediction results fused from the improved SHAP value analysis module. Data transmission is completed using the industrial Ethernet protocol, with cyclic redundancy check (CRC) verifying data integrity and eliminating abnormal data caused by transmission errors. The quality prediction results are directly mapped to reinforcement learning state input vectors, with the state vector dimension consistent with the types of quality indicators and the number of key control nodes. Simultaneously, the parameter adjustment amounts of the key control nodes are defined as the reinforcement learning action space, with boundary values ​​set based on the parameter sensitivity range to ensure that the action outputs are always within the allowable range of the process.

[0078] Step S507: When in the single-factory optimization stage, based on the key control node information, parameter sensitivity range and quality prediction results, construct a reinforcement learning agent and set a three-dimensional reward function that achieves quality standards, minimizes energy consumption and maximizes production efficiency. Quality standards are judged based on the fact that the purity of the finished product, the amount of sterol residue and the amount of solvent residue all meet the preset product quality standards. When in the single-factory optimization stage, a reinforcement learning agent based on a deep policy gradient architecture is constructed based on the received key control node information, parameter sensitivity range, and quality prediction results. The agent includes an input layer, a hidden layer, and an output layer. The dimension of the input layer matches the dimension of the state vector. The hidden layer uses the ReLU activation function to improve feature fitting ability, and the output layer uses the Softmax activation function to output the action probability distribution. A three-dimensional reward function is set to achieve quality compliance, lowest energy consumption, and highest production efficiency. The weight coefficients of the three-dimensional objectives are determined by the analytic hierarchy process (AHP). The weight coefficient for achieving quality compliance is the highest. The weight coefficients for lowest energy consumption and highest production efficiency are dynamically adjusted according to the factory's production priority. Quality compliance is judged by the fact that the purity of the finished product, the amount of sterol residue, and the amount of solvent residue all meet the relevant national food and pharmaceutical industry product quality standards. The specific values ​​of the standards are calibrated by statistically analyzing the factory's historical qualified product data.

[0079] Step S508: The reinforcement learning agent and the digital twin interact and train. The agent outputs parameter adjustment actions based on the current quality prediction results and the status of key control nodes. The digital twin simulates the production process corresponding to the adjustment action and returns the quality change results and reward value. The agent updates the strategy based on the reward value. After multiple rounds of iterative training until the model converges, the optimal real-time optimization strategy for a single factory is obtained. A real-time interactive interface is established between the reinforcement learning agent and the digital twin. The interaction frequency is consistent with the response cycle of the production process. The agent adjusts its actions by outputting parameters through the output layer based on the current quality prediction results and the status of key control nodes. After receiving the adjustment actions, the digital twin simulates the production process corresponding to the adjustment action based on the built-in extraction kinetics mechanism model, crystallization thermodynamics mechanism model, and the entire process logic. It outputs the quality change results and the corresponding reward value, which is calculated using a three-dimensional reward function. The agent uses the quality change results as a feedback signal and updates the policy network parameters using a temporal difference learning algorithm. During iterative training, the average reward value of each iteration is recorded. When the fluctuation range of the average reward value is less than 0.01 for 50 consecutive iterations, the model is considered to have converged, training is stopped, and the current policy network parameters are saved, thus obtaining the optimal real-time optimization strategy for a single factory.

[0080] Step S509: When in the multi-factory collaboration stage, each factory acts as a local client, independently training reinforcement learning agents based on local production data, and only uploading the trained model parameters to the cloud federated server through encrypted transmission. When in the multi-factory collaboration stage, each factory acts as an independent local client, replicating the reinforcement learning agent training process of the single-factory optimization stage based on its own production data. During local training, it only uses data such as raw material attributes, process parameters, and quality inspection collected by its own factory, and does not transmit raw production data to any external nodes. After training, the model parameters are encrypted using an asymmetric encryption algorithm to generate an encrypted parameter data package. The data package contains key information such as the model weight matrix, bias vector, and number of training iterations. The encrypted parameter data package is then uploaded to the cloud federated server through a virtual private network.

[0081] Step S510: The cloud federated server uses the federated averaging algorithm to aggregate the model parameters uploaded by each client, dynamically allocates weights according to the production scale, data quality and model training accuracy of each factory, calculates the weighted average of the model parameters of all clients, and generates a global optimized model. After receiving encrypted parameter data packets uploaded by each local client, the cloud-based federated server decrypts them using the corresponding private key to obtain the model parameters of each client. Weight coefficients are dynamically allocated based on the production scale, data quality, and model training accuracy of each factory. Production scale is quantified by annual oryzanol production, data quality is evaluated based on data integrity and accuracy, and model training accuracy is measured by the average reward value of the single-factory optimization strategy. The weight coefficients are normalized to ensure a sum of 1. A federated averaging algorithm is used to aggregate the model parameters of each client. First, the product of each client's model parameter and its corresponding weight coefficient is calculated, then all product results are summed to obtain the global model parameters. Based on these parameters, a global optimization model adapted to multi-factory scenarios is constructed.

[0082] Step S511: After encrypting the global optimization model, it is sent to each client. Each client updates its local agent parameters based on the global optimization model and outputs the optimal process parameter adjustment strategy for the current production scenario.

[0083] The parameters of the global optimization model are re-encrypted using an asymmetric encryption algorithm to generate an encrypted global model data packet, which is then distributed to each local client via a virtual private network. Upon receiving the packet, each client decrypts it using its private key to obtain the global model parameters. A model parameter fusion strategy is then used to weight and fuse the global model parameters with the local model parameters. The fusion weights are set based on the similarity of the distribution between the local factory production data and the global data. After fusion, the policy network parameters of the local reinforcement learning agent are updated. Through an iterative process of uploading local training parameters, aggregating them in the cloud, distributing the global model, and updating them locally, the performance of the local agent is continuously optimized. Finally, the optimal process parameter adjustment strategy for the current production scenario is output, ensuring that the generalization accuracy of the global model in each factory meets the requirements of industrial applications.

[0084] It should be noted that step S600, after generating the optimization strategy, is also included, and then distributed to the edge node control device to link the blockchain and digital twin to complete quality traceability and compliance execution. The specific steps are as follows: Step S601: Receive the optimal process parameter adjustment strategy, and send the strategy to the corresponding edge node through the industrial communication network. The edge node adjusts the equipment operating parameters of each production link according to the strategy instructions. Step S602: During the production process and parameter adjustment, collect preprocessed data, fused feature vectors, quality prediction results, optimization parameters, and risk event records in real time. After classifying and organizing the data by production batch, transmit it to the consortium blockchain. Step S603: The consortium blockchain stores the data of the entire process, and configures a unique batch number, precise timestamp, data hash value and node signature for each batch of data. The chain storage structure is used to associate each new data block with the previous data block through the hash value. Step S604: When a quality problem occurs, retrieve the full process data of the corresponding batch from the blockchain through the batch number, input the data into the digital twin, and drive the digital twin to reproduce the production process and parameter change trajectory of the batch. Step S605: Pre-set quality compliance rules that conform to the food and pharmaceutical industry standards in the blockchain smart contract, obtain quality testing data and quality prediction results in real time, and compare them with the pre-set compliance rules; Step S606: When the comparison result meets all the requirements of the compliance rules, the smart contract automatically executes the product warehousing operation, generates an warehousing voucher containing batch number, product quantity, quality indicator test value, and warehousing time, and synchronizes it to the blockchain. Step S607: When the comparison result fails to meet any of the compliance rules, the smart contract automatically executes the product isolation operation, generates an isolation notification containing the isolation area, isolation period and subsequent handling suggestions, and suspends the shipment of the batch of products. Step S608: After the smart contract executes the storage or isolation operation, a quality report is automatically generated. The report includes a summary of process parameters for the entire production process, quality inspection results for each stage, comparative analysis of predicted and actual inspection results, and compliance judgment conclusions. After encryption, the report is stored on the blockchain for authorized nodes to retrieve and verify.

[0085] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process, characterized in that, Includes the following steps: Based on the entire production process of oryzanol, edge acquisition nodes are deployed to collect multi-source heterogeneous data and transmit it to the edge nodes. Based on multi-source heterogeneous data, edge nodes perform differentiated preprocessing on different types of data, synchronize them to the cloud distributed database, and generate virtual data in the cloud to supplement small sample scenarios, forming an enhanced dataset; Various data features are extracted from the augmented dataset to construct a multi-scale feature matrix. After being fused by an adaptive multi-scale attention module, the fused feature vector is output. The fused feature vectors are input into two data-driven models, while the pre-processed multi-source heterogeneous data is input into the preset digital twin and mechanism model. The quality prediction results and confidence are output by improving the evidence theory. The fused feature vector input is used to improve the SHAP value analysis module to locate key control nodes, and the optimization strategy is generated by federated reinforcement learning in combination with the quality prediction results.

2. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The multi-source heterogeneous data includes raw material properties, process parameter time series, equipment operation, environmental perception, and quality inspection data. The specific steps for edge nodes to perform differentiated preprocessing on different types of data are as follows: The system receives raw material attributes, process parameter timing, equipment operation, environmental perception, and quality inspection data transmitted from edge acquisition nodes, carrying links, timestamps, equipment number, and 3D tags. These data are categorized into three types based on data type: structured data, time-series data, and spectral data. For structured data, dynamic standardization is adopted to update the data mean and standard deviation in real time to adapt to the fluctuation of raw material batches and eliminate the difference in data units. For time-series data, adaptive wavelet threshold denoising is first used. The threshold size is dynamically adjusted according to the data noise intensity to reduce the proportion of high-frequency interference signals. Then, the window size is dynamically adjusted according to the data frequency for feature extraction. High-frequency equipment operation data uses a short window for feature extraction, while low-frequency environmental perception data uses a long window for feature extraction. The window size is set according to the data acquisition frequency and process response characteristics. MSC correction, adaptive smoothing, and sparse dimensionality reduction are performed sequentially on the spectral data. MSC correction is used to eliminate baseline drift and scattering interference. The polynomial order of the adaptive smoothing is dynamically selected according to the complexity of the spectral signal. Sparse dimensionality reduction significantly compresses the amount of data while retaining the core information. A dual screening mechanism is used to remove abnormal data for all types of data. First, the 3σ criterion is used for preliminary screening. Then, the process rules are verified according to the reasonable range set in the oryzanol production technical specifications. Data that exceeds the range and whose duration meets the abnormal judgment criteria is judged as abnormal. For datasets with missing data after screening, the process correlation interpolation method is used to complete them, and the completion value is calculated based on the process correlation between the missing data stage and the preceding and following processes.

3. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The specific steps for generating virtual data in the cloud to supplement the small sample scenario are as follows: Receive preprocessed data synchronized from edge nodes to the cloud distributed database, classify the preprocessed data according to production scenarios, identify small sample scenarios such as sudden changes in raw material properties, abnormal equipment parameters, and adjustments to process conditions, and determine the core data characteristics and distribution patterns of small sample scenarios; Based on the core data features of small sample scenarios, a generative adversarial network model is constructed. Real preprocessed data is used as training samples to train the generator and discriminator of the generative adversarial network. The generator learns the distribution characteristics of real data, and the discriminator distinguishes and verifies the generated data from the real data. Virtual data is generated by training a mature generative adversarial network. The generated virtual data must be consistent with the real data in the corresponding small sample scenario in terms of feature distribution and process logic. The generated virtual data is validated for process rationality. Abnormal virtual data that does not conform to the logic of oryzanol production process is removed, and valid virtual data that conforms to actual production is retained. By integrating effective virtual data with real preprocessed data in a cloud-distributed database, an enhanced dataset covering all scenarios is formed.

4. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The specific steps for constructing the multi-scale feature matrix are as follows: Three types of raw data—structured data, time-series data, and spectral data—are extracted from the augmented dataset generated in the cloud, while retaining the production process and timestamp labels carried by the data. Based on three types of raw data, a 1DCNN network is used to perform convolution operations on short-timescale data to extract local temporal features and instantaneous change features. The short-timescale data includes real-time parameters of equipment operation and instantaneous data of the process. Based on three types of raw data, a bidirectional LSTM network is used to perform sequence learning on long-term data to extract long-range dependencies and trend features in the time dimension. The long-term data includes raw material batch attribute data, finished product quality batch detection data, and process parameter trend data. Based on three types of raw data, a CNN network is used to perform multi-layer convolution and pooling operations on the spectral data to extract feature peaks, feature valleys and texture distribution features. Short-scale local temporal features, long-scale long-range dependency features, and spectral texture features are vectorized and normalized respectively. Based on the correspondence between production links and timestamps, a multi-scale feature matrix is ​​constructed.

5. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The specific steps of the adaptive multi-scale attention module fusion are as follows: Based on the multi-scale feature matrix, the total number of feature categories in the matrix and the types of quality indicators to be predicted are determined. The correlation strength between each feature and each quality indicator is calculated using a gating unit, as shown in the following formula: ， in, This represents the quantified value of the correlation strength between the i-th feature and the j-th quality indicator. This represents the weight matrix of the gated unit. Represents the feature vector of the i-th class. This represents the j-th quality index parameter. This represents a vector concatenation operation. Indicates the bias term of the gating unit; The Sigmoid activation function is used to process the quantified values ​​of correlation strength, and the process correlation weights of various features with each quality index are calculated. The calculation formula is as follows: ， in, This represents the process relevance weight between the i-th feature and the j-th quality indicator. represents the Sigmoid activation function, and n represents the total number of feature categories involved in the weight calculation; Represents all features corresponding to k from 1 to n Summation operation on the values; The weight matrix and bias terms of the gated unit are iteratively optimized by using the gradient descent algorithm. The various eigenvectors in the multi-scale feature matrix are weighted with the corresponding process-related weights to obtain the weighted eigenvectors. All weighted feature vectors are input into the Transformer encoder, and deep feature fusion is performed through a multi-head attention mechanism to output a fused feature vector.

6. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The specific steps for constructing the digital twin are as follows: Collect physical entity information of the entire production process of oryzanol, including the physical structure parameters, installation layout, and operating principle of raw material pretreatment equipment, extraction equipment, purification equipment, and finished product preparation equipment, as well as the process flow diagram, process parameter range, and material transfer path of each link; A digital twin is built based on a 3D simulation engine and mapped 1:1 to the actual production scene, replicating the physical structure, operating parameters and process logic of various production equipment according to the physical entity information; Establish a two-way mapping mechanism between physical entities and digital twins. Through edge nodes, synchronize the real-time operation data and process parameter data of the physical entity to the digital twin, and transmit the simulation results of the digital twin back to the control terminal of the physical entity. The extraction kinetics mechanism model and the crystallization thermodynamics mechanism model are embedded in the digital twin, and the process parameter simulation and deduction function is configured for the digital twin. It can receive the pre-processed process parameters, simulate the quality change trend under different working conditions, and continuously compare and calibrate with historical actual production data.

7. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The specific steps for fusing the improved evidence theory to output quality prediction results and confidence levels are as follows: It receives data-driven prediction results from two data-driven models: the first is an improved gradient boosting tree model, and the second is a feature dependency capture model. It also receives simulation prediction results from the digital twin and mechanistic prediction results from the mechanistic model. Normalize the data-driven prediction results, simulation prediction results and mechanism prediction results respectively, map the numerical range to the interval of 0 to 1, calculate the basic probability allocation value of each quality index corresponding to each type of prediction result, and calculate the conflict coefficient between different prediction results. A preset conflict coefficient threshold is set. When the calculated conflict coefficient is lower than the threshold, the basic probability allocation value is fused using conventional evidence theory fusion rules. When the calculated conflict coefficient is higher than or equal to the threshold, the basic probability allocation value of the high conflict prediction result is corrected by the weighted correction mechanism, and then the evidence theory fusion rule is used for fusion calculation. The final quality prediction result is determined based on the fusion calculation results, and the confidence level of the prediction result is calculated.

8. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 1, characterized in that, The specific steps for locating key control nodes using the improved SHAP value analysis module are as follows: Based on the fusion feature vector, the characteristics of various process parameters contained in the fusion feature vector are clarified. Process constraint weights are introduced into the SHAP value analysis model. The process constraint weights of each link in the production of oryzanol are set according to the process importance, parameter adjustability and sensitivity of the impact on quality indicators. The process constraint weights of the extraction and purification links are higher than those of the raw material pretreatment and finished product preparation links. An improved SHAP value analysis model is developed by incorporating fused feature vector inputs into process constraint weights to calculate the SHAP value for each process parameter. All process parameters are sorted in descending order based on the absolute value of their SHAP values. The process parameters with the highest impact are selected as key control nodes. These key control nodes include extraction temperature, feed-to-liquid ratio, crystallization cooling rate, drying vacuum degree, solvent ratio, extraction time, stirring speed, alkali addition amount, crystallization settling time, and filtration accuracy. For each identified key control node, the parameter sensitivity range is determined by combining its corresponding process constraint weight and SHAP value sign. When the actual value of the parameter exceeds the sensitivity range, the system automatically marks it as a high-risk parameter. The analysis results of key control node information, SHAP values ​​of each node, and parameter sensitivity ranges are pushed to the federated reinforcement learning optimization module in real time.

9. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 8, characterized in that, The specific steps for generating optimization strategies using federated reinforcement learning are as follows: It receives information on key control nodes, parameter sensitivity ranges, and quality prediction results, uses the quality prediction results as the state input for reinforcement learning, and uses the parameter adjustment amount of key control nodes as the action space. When in the single-factory optimization stage, a reinforcement learning agent is constructed based on key control node information, parameter sensitivity range and quality prediction results. A three-dimensional reward function is set to achieve quality target, lowest energy consumption and highest production efficiency. Quality target is determined by the fact that the purity of finished product, sterol residue and solvent residue all meet the preset product quality standards. The reinforcement learning agent and the digital twin are trained interactively. The agent outputs parameter adjustment actions based on the current quality prediction results and the status of key control nodes. The digital twin simulates the production process corresponding to the adjustment action and returns the quality change results and reward value. The agent updates the strategy based on the reward value. After multiple rounds of iterative training until the model converges, the optimal real-time optimization strategy for a single factory is obtained. When in the multi-factory collaboration stage, each factory acts as a local client, independently training reinforcement learning agents based on local production data, and only uploading the trained model parameters to the cloud federated server through encrypted transmission. The cloud-based federated server uses a federated averaging algorithm to aggregate the model parameters uploaded by each client. It dynamically allocates weights based on the production scale, data quality, and model training accuracy of each factory, calculates the weighted average of all client model parameters, and generates a globally optimized model. The global optimization model is encrypted and then distributed to each client. Each client updates its local agent parameters based on the global optimization model and outputs the optimal process parameter adjustment strategy for the current production scenario.

10. The method for multi-source heterogeneous data fusion analysis and quality optimization of the entire oryzanol production process according to claim 9, characterized in that, After generating the optimization strategy, it is distributed to the edge node control device, linking the blockchain and digital twin to complete quality traceability and compliance execution.

Images