Intelligent regulation and control method and system for whole process of low-temperature extraction and separation of scallion oil

Abstract

The invention discloses an intelligent regulation and control method and system for the whole process of low-temperature extraction and separation of scallion oil, and relates to the technical field of food processing.In the production process of extraction and separation of scallion oil, quantitative evaluation is conducted on raw material characteristics, and technological parameters are optimized.The method comprises the steps that near infrared spectrum data of scallion raw materials are collected and preprocessed; performing prediction and similar case retrieval on raw material characteristic parameters; extracting, collecting time sequence data in real time, predicting a product quality index, and evaluating and comparing with an expected effect; detecting a quality deviation risk, calculating an adjustment scheme, performing security constraint check, and then performing parameter adjustment; carrying out separation and purification and real-time monitoring, and removing impurities by adopting a three-stage separation strategy; quality detection is carried out, a quality detection report is generated, and production data arrangement and quality prediction model updating are carried out. According to the method, through adaptive optimization of technological parameters, product quality prediction control and whole-process collaborative optimization, whole-process intelligent management, including extraction and separation processes, of scallion oil production is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology in food processing, and more specifically, to an intelligent control method and system for the entire process of low-temperature extraction and separation of scallion oil. Background Technology

[0002] Scallion oil is a functional oil product extracted from scallions through an extraction process. It is rich in bioactive components such as sulfides and polyphenols, and possesses a unique flavor and health benefits. The sulfides in scallion oil have antioxidant, antibacterial, and lipid-lowering physiological activities, while the polyphenols have the effect of scavenging free radicals and delaying aging. It has wide applications in condiments, health foods, and other fields.

[0003] Traditional scallion oil production relies heavily on manual experience to control process parameters. Operators judge the quality of raw materials based on sensory characteristics such as appearance and odor, and set parameters like extraction temperature and time based on experience. This method has several problems: the judgment of raw material characteristics is highly subjective and difficult to quantify; there are large batch-to-batch quality fluctuations, with significant differences in raw material composition between different batches, making it difficult to ensure stable quality using fixed process parameters; raw material utilization is low, as process parameters are not optimized according to raw material characteristics, leading to insufficient extraction or excessive decomposition of functional components; and production efficiency is unstable, requiring extended extraction times or increased temperatures to correct product quality deviations, increasing energy consumption and costs.

[0004] With the intelligent development of the food industry, the application of sensor technology, machine learning algorithms, and automatic control technology in food processing has become an important trend. Near-infrared spectroscopy can quickly and non-destructively detect the components of raw materials, machine learning algorithms can extract the relationship between raw material characteristics and process parameters and product quality from historical data, and model predictive control algorithms can achieve dynamic optimization of process parameters under constraints.

[0005] However, scallion oil production involves multiple stages, including raw material characteristic identification, extraction process control, and separation and purification, with complex coupling relationships between these stages. For example, in actual production, the following problem is frequently encountered: when using a fixed extraction temperature and time, the sulfide content of the final product fluctuates significantly due to differences in the initial sulfide content, moisture content, and freshness of different batches of raw materials. Some batches fall below the target lower limit and fail to meet quality standards; others suffer from excessive sulfide decomposition due to excessively high temperatures or prolonged extraction times. By the time quality deviations are detected, the extraction process is nearing completion, leaving limited room for adjustment and making effective correction difficult. The root cause of this problem lies in the lack of quantitative assessment methods for raw material characteristics, the inability to pre-optimize process parameters based on raw material characteristics, and the lack of quality prediction and dynamic adjustment mechanisms during the process.

[0006] How to achieve rapid and accurate identification of raw material characteristics, adaptive optimization of process parameters, real-time prediction of product quality, and collaborative control of multiple links, and how to build a scalable intelligent control system and achieve continuous learning and performance improvement are the current technical challenges facing the intelligent production of scallion oil. Summary of the Invention

[0007] This invention provides an intelligent control method and system for the entire process of low-temperature extraction and separation of scallion oil, which solves the technical problem in related technologies of lacking quantitative evaluation means of raw material characteristics and being unable to pre-optimize process parameters based on raw material characteristics.

[0008] This invention provides an intelligent control method for the entire process of low-temperature extraction and separation of scallion oil, including: Near-infrared spectral data of scallion raw materials were collected and preprocessed. The characteristic parameters of the raw materials were predicted and similar cases were searched to obtain recommended values ​​of process parameters and expected effect evaluation results. Extraction is performed based on recommended values ​​of process parameters, and process parameters are collected in real time. Product quality indicators are predicted and compared with the expected effect evaluation results to obtain product quality prediction results. Based on the product quality prediction results, the risk of quality deviation is detected, the process parameter adjustment plan is calculated, the parameter adjustment is executed after safety constraint check, and the final state information of extraction is obtained. Based on the final state information of the extraction, separation and purification were carried out and monitored in real time. A three-stage separation strategy was adopted to remove impurities and obtain purified scallion oil product. The purified scallion oil product undergoes quality testing, a quality testing report is generated, production data is organized and updated, and an updated knowledge base is obtained. Traceability information is generated based on the quality inspection report, and data is classified and protected by digital signature to obtain traceability data packages and QR code identifiers.

[0009] In a preferred embodiment, the similar case retrieval employs a case-based reasoning method, specifically including: Construct the feature vector of the current batch of cases. The feature vector includes moisture content, soluble solids content, initial sulfide content, variety code, origin code, freshness index and uniformity index. Normalize each feature in the case feature vector, mapping the numerical features to the interval between 0 and 1; The similarity between the current case and each historical case in the historical database is calculated using weighted Euclidean distance. The weight coefficients of the weighted Euclidean distance are determined according to the degree of influence of each feature on the extraction effect, with the initial sulfide content having the highest weight, followed by the variety code and freshness index. Sort the cases by similarity from high to low, and select the top K most similar historical cases. The K value is set to 5 to 20. The process parameters of the selected K historical cases were weighted and averaged, with the weight of each historical case being the value after similarity normalization, to obtain the recommended extraction temperature, extraction time, oil-to-material ratio and enzyme addition amount. It also outputs an assessment of expected results, including expected sulfide content, expected polyphenol content, and expected color parameters.

[0010] In a preferred embodiment, the predicted product quality index employs a long short-term memory network, specifically including: The input features for constructing a long short-term memory network include temporal features and static features; The time-series features include: temperature time-series data, pressure time-series data, pH time-series data, and cumulative enzyme addition time-series data, with a sampling frequency of 1 to 10 times per minute; Static characteristics include: raw material moisture content, raw material soluble solids content, initial raw material sulfide content, raw material freshness index, raw material uniformity index, and recommended oil-to-seed ratio; A sliding time window is used to process the temporal features. The time window length is set to 10 to 60 time steps to form a three-dimensional tensor input. The tensor dimension is batch size × time step length × number of features. A three-dimensional tensor is input into a long short-term memory network. The network consists of an input layer, two LSTM hidden layers, and a fully connected output layer. The number of units in the first LSTM layer ranges from 64 to 128, and the number of units in the second LSTM layer ranges from 32 to 64. The Long Short-Term Memory Network outputs predicted values ​​for sulfide content, polyphenol content, volatile flavor compound content, and color parameters. The predicted values ​​are compared with the target quality indicators to calculate the quality deviation and deviation rate.

[0011] In a preferred embodiment, the calculation of the optimal process parameter adjustment scheme employs a model predictive control algorithm, specifically including: When a quality deviation risk is detected, the model predictive control algorithm is activated. An optimization problem is constructed based on a pre-established prediction model of the extraction process. The prediction model adopts a linear state space representation. The state variables include temperature, pH value, extraction time, and cumulative enzyme addition. The control variables include temperature adjustment, time adjustment, and enzyme addition adjustment. The prediction time domain is set to 10 to 30 time steps, and the control time domain is set to 3 to 10 time steps; Construct an optimization objective function that includes a quality tracking error term and a control change penalty term, with the weight of the quality tracking error term being higher than that of the control change penalty term; The following hard constraints are set: temperature hard constraint is 40℃ to 80℃, and temperature change rate constraint is no more than 2℃ per minute; extraction time hard constraint is 30 minutes to 180 minutes; enzyme addition hard constraint is 0.1% to 2% of the raw material mass, and the single addition amount shall not exceed 30% of the total amount; pressure hard constraint is -0.05MPa to 0.1MPa; pH value hard constraint is 4.5 to 7.5. Set soft constraints, including: energy consumption per batch does not exceed the preset energy consumption limit, and the total extraction time is preferably controlled within the preset time range; The optimization problem is solved using a quadratic programming method, the optimal control sequence is calculated, and the adjustment scheme of process parameters is output.

[0012] In a preferred embodiment, obtaining the purified onion oil product specifically includes: The system receives the final state information of the extraction process and transmits the parameters and state information of the extraction process to the separation control module. The separation control module adjusts the separation process parameters according to the linkage control model. Estimate the viscosity of the mixture based on the final extraction temperature and adjust the centrifuge speed accordingly; determine whether to adjust the pH value of the separation stage based on the final extraction pH value. The first-stage coarse separation removes large solid particles through gravity sedimentation; when the sludge layer reaches a preset thickness threshold, the sludge is discharged; the upper clarified oil phase flows into the next stage centrifugal separation equipment. The second stage of centrifugal separation removes fine particles and some moisture through centrifugal force; The third-stage precision filtration removes minute impurities through multiple layers of filter membranes; when the differential pressure exceeds the preset differential pressure threshold, an alarm is triggered to prompt the replacement of the filter element; During the three-stage separation process, the oil phase quality is monitored according to the second preset acquisition frequency; when the monitored value exceeds the target value, the parameters are adjusted.

[0013] In a preferred embodiment, the updated knowledge base specifically includes: Quality assessment is conducted using a combination of online and offline testing; a quality inspection report is generated; and complete data for this batch is stored in a historical database. The actual quality test results are compared with the expected effect assessment, and the prediction deviation is calculated. When the accumulated number of historical batches reaches the third preset batch number, the model update mechanism is triggered; Model retraining employs an incremental learning approach; if the new model outperforms the old model, the new model is downloaded to the edge device to replace the old model. Identify the success factors and failure reasons for this batch; extract the identified experiences into rules and add them to the intelligent knowledge base.

[0014] In a preferred embodiment, the data hierarchical processing adopts a three-level hierarchical rule, specifically including: The traceability information in this batch is classified into three levels according to its level of openness: Level 1 Public Data, Level 2 Semi-Public Data, and Level 3 Confidential Data. Level 1 publicly available data includes: product batch number, production date, manufacturer name, manufacturer address, product specifications, shelf life, quality grade, and applicable standards; Secondary semi-public data includes: raw material origin, raw material variety, raw material freshness grade, extraction temperature curve, extraction time, oil-to-material ratio, separation process type, and sulfide content and polyphenol content in quality test results; Level 3 confidential data includes: detailed raw material characteristic parameters, complete process parameter time series data, enzyme addition amount, pH adjustment records, quality prediction model parameters, model prediction control algorithm parameters, cost data, supplier information, and energy consumption data; The primary public data and the secondary semi-public data are serialized, the digital fingerprint is calculated using the SHA-256 hash algorithm, and the hash value is digitally signed using the enterprise's private key. Level 3 confidential data is encrypted using the AES-256 symmetric encryption algorithm and stored encrypted on the enterprise's local server. Generate a QR code traceability identifier, which contains the product batch number and a traceability query link.

[0015] In a preferred embodiment, the prediction of raw material characteristic parameters adopts a neural network spectral analysis model, which includes an input layer, a hidden layer and an output layer. The hidden layer uses the ReLU activation function and a Dropout layer is set after the hidden layer.

[0016] In a preferred embodiment, the extraction process prediction model adopts a linear state-space representation, and the state variables include temperature, pH value, extraction time, and cumulative enzyme addition, while the control variables include temperature adjustment, time adjustment, and enzyme addition adjustment.

[0017] This invention provides an intelligent control system for the entire process of low-temperature extraction and separation of scallion oil, used to execute the aforementioned intelligent control method for the entire process of low-temperature extraction and separation of scallion oil, comprising: The raw material characteristic prediction module is used to collect near-infrared spectral data of onion raw materials and perform preprocessing, predict raw material characteristic parameters and search for similar cases, and obtain recommended values ​​of process parameters and expected effect evaluation results. The extraction quality prediction module performs extraction based on recommended values ​​of process parameters and collects process parameters in real time. It predicts product quality indicators and compares them with the expected effect evaluation results to obtain product quality prediction results. The parameter optimization and adjustment module detects the quality deviation risk based on the product quality prediction result, calculates the process parameter adjustment plan, performs safety constraint checks, and then executes parameter adjustment to obtain the final state information of leaching; The separation and purification control module performs separation and purification based on the final state information of leaching and monitors it in real time. It adopts a three-stage separation strategy to remove impurities and obtains purified scallion oil products; The model update and feedback module is used to conduct quality inspections on the purified scallion oil products, generate quality inspection reports, organize and update production data, and obtain an updated knowledge base; The traceability system construction module generates traceability information based on the quality inspection report, conducts data classification processing and digital signature protection, and obtains traceability data packets and two-dimensional code identifiers.

[0018] The beneficial effects of the present invention are as follows: By constructing a hierarchical and scalable intelligent control system, the intelligent management of the entire scallion oil production process is realized. By using near-infrared spectroscopy technology combined with a neural network model, the characteristics of raw materials can be quickly and accurately identified, providing a reliable basis for process parameter optimization. The decision-making method based on case-based reasoning makes full use of historical production data, improving the accuracy and interpretability of process parameter recommendations.

[0019] Using a long short-term memory network for dynamic prediction of product quality can detect quality deviation risks in advance during the mid-term of the leaching process,争取时间窗口为及时调整工艺参数争取时间窗口。采用模型预测控制算法进行参数优化,在满足安全约束和经济约束的前提下,实现了质量指标的精确控制。通过多环节联动控制策略,将浸提环节的状态信息传递给分离环节,实现了工艺参数的自适应调整和全流程协同优化。

[0020] A closed-loop feedback mechanism is established. By continuously accumulating production data and updating the prediction model and knowledge base, the continuous learning and performance improvement of the system are realized. The product traceability system constructed using digital signature technology ensures the authenticity and integrity of traceability data, enhancing the market trust of the product. The present invention improves the stability of scallion oil product quality, production efficiency, and raw material utilization rate, reduces production costs, and has good application value and promotion prospects. Brief Description of the Drawings

[0021] Figure 1 is the main flow chart of the present invention; Figure 2 is the detailed flow chart of the present invention; Figure 3 is the module diagram of the full-process intelligent control system of the present invention. Detailed Embodiments

[0022] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0023] At least one embodiment of the present invention discloses an intelligent control method for the entire process of low-temperature extraction and separation of scallion oil, such as... Figure 1 - Figure 2 As shown, it includes: Step 1: Collect near-infrared spectral data of scallion raw materials and perform preprocessing; predict raw material characteristic parameters and search similar cases to obtain recommended values ​​of process parameters and expected effect evaluation results. Specifically, the following steps are included: Step 1.1, Raw material sample collection and spectral data acquisition; When a new batch of scallion raw materials arrives, the operator randomly selects a sample from the raw materials, preferably 300 to 800 grams. After simple crushing, the sample is placed in a sample cup and placed in the detection window of a portable near-infrared spectrometer. The spectrometer scans within a preset wavelength range, preferably 800 to 2500 nanometers, to obtain the near-infrared spectral data of the raw materials.

[0024] After obtaining the raw spectral data, preprocessing is performed to improve data quality and model prediction accuracy. Preprocessing steps include: baseline correction to eliminate spectral drift and background interference; standard normal variable transformation (SNV) to eliminate the influence of sample particle size and optical path variation; and normalization to scale the spectral data to the range of 0 to 1, facilitating processing by the neural network model. The preprocessed spectral data forms spectral vectors, which serve as input to the subsequent neural network model.

[0025] Step 1.2, Prediction of raw material characteristic parameters based on neural networks; A pre-trained neural network spectral analysis model is used to quickly identify the characteristics of raw materials. This model employs a multi-layer fully connected neural network structure, including an input layer, hidden layers, and an output layer. The input layer receives pre-processed spectral vectors; the dimension of the spectral vectors depends on the number of wavelength sampling points, preferably between 200 and 500. The hidden layers use the ReLU activation function to extract nonlinear features from the spectral data and integrate the feature information. The number of nodes in the hidden layers is preferably between 100 and 300, and 1 to 3 hidden layers can be set. To prevent overfitting, a Dropout layer is added after the hidden layers, with a preferred Dropout ratio of 0.1 to 0.3.

[0026] The output layer outputs predicted values ​​for multiple raw material characteristic parameters. Specifically, the output layer contains multiple output nodes, each corresponding to a raw material characteristic parameter. The number of output nodes equals the number of raw material characteristic parameters to be predicted. In this embodiment, five output nodes are set, corresponding to moisture content, soluble solids content, initial sulfide content, variety code, and origin code, respectively. Each output node outputs the predicted value of the corresponding parameter through a linear activation function. For continuous parameters such as moisture content, soluble solids content, and initial sulfide content, the output node directly outputs the numerical value; for categorical parameters such as variety code and origin code, the output node outputs the numerical code of the category label.

[0027] The spectral analysis model is pre-trained on the company's historical production data, with an optimal training data volume of 200 to 500 batches. The model's prediction accuracy meets preset requirements, with an optimal prediction correlation coefficient greater than 0.85, and the average absolute error of each raw material characteristic parameter is controlled within 5%. The preprocessed spectral vector is input into the model, which quickly outputs the predicted values ​​of the raw material characteristic parameters, with a preferred prediction time of 3 to 8 seconds.

[0028] The output of the spectral analysis model includes basic raw material characteristic parameters such as moisture content, soluble solids content, initial sulfide content, variety code, and origin code.

[0029] Step 1.3, Calculation of raw material freshness index and uniformity index; After obtaining the basic raw material characteristic parameters, the raw material freshness index and raw material uniformity index are further calculated.

[0030] The raw material freshness index is a comprehensive indicator calculated by analyzing characteristic bands in the near-infrared spectrum related to chlorophyll degradation. The specific calculation method is as follows: extract data from the 680-750 nm and 1400-1500 nm bands from the spectral data; analyze the absorption intensity and peak shape characteristics of these bands; calculate the average reflectance value of each band; multiply the average reflectance of the 680-750 nm band by a weight of 0.6, and multiply the average reflectance of the 1400-1500 nm band by a weight of 0.4; add the two values ​​together and divide by the maximum possible value of the reflectance of the two bands to obtain the freshness index, which ranges from 0 to 100. A higher value indicates fresher raw materials.

[0031] The raw material uniformity index is calculated by analyzing the stability of spectral data. Multiple repeated measurements are performed on the same batch of raw materials, preferably 3 to 8 times. The coefficient of variation (COP) of each raw material characteristic parameter is calculated, and the COP is obtained by dividing the standard deviation by the mean. The uniformity index is calculated by subtracting the weighted average of the COP percentages from 100. The value ranges from 0 to 100; a higher value indicates a more uniform distribution of raw material components.

[0032] This step yields the raw material freshness index and raw material uniformity index. Combined with the basic raw material characteristic parameters obtained in step 1.2, a complete raw material characteristic parameter vector is formed, including seven parameters: moisture content, soluble solids content, initial sulfide content, variety code, origin code, freshness index, and uniformity index.

[0033] Step 1.4, similar case retrieval based on historical data; Based on the company's historical production data, historical batches with similar characteristics to the current raw materials are retrieved. Historical production data includes the characteristic parameters of different batches of raw materials, the process parameters used, and the final product quality results. This data allows for the establishment of a mapping relationship between raw material characteristics and optimal process parameters.

[0034] A case-based reasoning decision-making method is adopted to retrieve similar cases from the historical database. Similarity calculation uses weighted Euclidean distance, assigning weights to each characteristic parameter. In this embodiment, the weight of initial sulfide content is set to 0.30, the weight of raw material freshness index is set to 0.25, the weight of soluble solids content is set to 0.20, the weight of moisture content is set to 0.15, and the weights of variety code and origin code are set to 0.05 and 0.05 respectively, with the sum of all weights being 1. The specific calculation process is as follows: the raw material characteristic parameter vector of the current batch is compared with the raw material characteristic parameter vector of each batch in the historical database, and the weighted Euclidean distance is calculated. The formula is: the distance is equal to the square root of the sum of the squares of the differences between each parameter and the products of their corresponding weights; the smaller the distance, the higher the similarity.

[0035] Based on the similarity scores, the most similar historical batches were retrieved, with an optimal retrieval number of 5 to 15 batches. Statistical analysis was then performed on the process parameters of these similar historical batches to extract common patterns. Specifically, the analysis method involved: cluster analysis of the process parameters of the retrieved similar batches to identify frequently occurring parameter combination patterns; calculation of the mean and standard deviation of each process parameter to determine reasonable value ranges; and combining this with the quality results of the historical batches to screen out process parameter combinations that consistently yield qualified products, identifying the key parameters with the most significant impact on quality.

[0036] This step obtains several historical batches with process parameter statistical characteristics that are most similar to the current raw material characteristics, providing a data basis for subsequent process parameter recommendations.

[0037] Step 1.5, Recommendation of process parameters and decision output; Based on the similar historical batches retrieved in step 1.4 and their statistical analysis results, the optimal process parameters for this batch are recommended. The specific recommendation method is as follows: a weighted average of the process parameters of several similar historical batches is calculated. The weight of each historical batch is determined based on its similarity; batches with higher similarity have greater weights. Specifically, the weight of each batch is equal to its similarity divided by the sum of the similarities of all retrieved batches, and normalization is used to ensure that the sum of all weights is 1. The recommended process parameter values ​​are obtained by multiplying the corresponding parameter values ​​of each historical batch by their respective weights and then summing the results. The recommended process parameters include extraction temperature, extraction time, enzyme dosage, ultrasonic power density, pH value, and oil-to-material ratio.

[0038] Simultaneously, interpretable information for decision-making is provided. By analyzing similar cases, key factors affecting the quality of this batch of products are identified. Based on the levels of raw material freshness and uniformity indices, combined with key parameters such as initial sulfide content, appropriate extraction temperatures and times are recommended. This ensures sufficient extraction of functional components such as sulfides and polyphenols while avoiding excessive decomposition of sulfides and volatilization of flavor compounds due to excessively high temperatures or prolonged extraction times. This interpretable information helps operators understand the basis of their decisions, enhancing their trust and acceptance of intelligent control methods.

[0039] The output also includes an expected outcome assessment. Based on statistical analysis of quality results from similar historical batches, the expected range of product quality indicators that this batch may achieve after adopting the recommended process parameters is predicted, including the expected ranges for sulfide content, polyphenol content, volatile flavor compound content, and color parameters. These expected outcome assessments provide operators with a reference for quality expectations and also serve as a benchmark for subsequent quality predictions and comparisons.

[0040] This step yields recommended optimal process parameters based on historical data matching, as well as interpretability information for the decision and an assessment of expected effects, providing scientifically reliable control setpoints for starting the extraction process.

[0041] Step 2: Extraction is performed based on the recommended values ​​of process parameters and process parameters are collected in real time. Product quality indicators are predicted and compared with the expected effect evaluation results to obtain product quality prediction results. Specifically, the following steps are included: Step 2.1: Initiation of the extraction reaction and precise temperature control; Add the raw materials and vegetable oil to the extraction tank according to the oil-to-material ratio recommended in step 1.5, preferably between 2:1 and 5:1, and start the automatic control device. The temperature control device achieves precise temperature control through an electric heater and cooling water circulation, with the control accuracy meeting the process requirements, preferably ±0.5 to 2 degrees Celsius.

[0042] Precise temperature control is achieved using a Proportional Integral Derivative (PID) control algorithm. Its working principle involves calculating heating or cooling power based on the deviation between the current temperature and the target temperature, the integral of the deviation, and the rate of change of the deviation. The output of the PID controller equals the proportional term plus the integral term plus the derivative term. The proportional term is used for rapid response to deviations, the integral term is used to eliminate steady-state errors, and the derivative term is used to suppress temperature fluctuations. These three terms work together to quickly stabilize the temperature near the target value. Temperature data is collected at a first preset sampling frequency, the PID output is calculated, and the heater power is adjusted to form a closed-loop control.

[0043] This step initiates the extraction reaction and enables precise temperature control, providing a stable temperature environment for the extraction reaction.

[0044] Step 2.2: Real-time acquisition and dynamic adjustment of process parameters; During the extraction process, process parameters such as temperature, pressure, and pH are collected every second, forming a time-series data stream. This data is stored in real time on edge computing devices, providing a data foundation for subsequent quality prediction.

[0045] Simultaneously, process parameters are dynamically adjusted. Enzyme preparation is added every 30 minutes according to a preset time interval, with the amount added dynamically adjusted based on the current extraction progress and temperature. In the initial extraction stage, the enzyme addition is set at 35% to 45% of the total amount to disrupt cell wall structure; in the middle stage, the enzyme addition is set at 30% to 40% of the total amount to accelerate component dissolution; and in the later stage, the enzyme addition is set at 20% to 30% of the total amount to extract poorly soluble components. This staged addition strategy improves enzyme utilization efficiency and maintains enzyme activity at a consistently high level.

[0046] pH control employs an intermittent adjustment strategy. pH values ​​are measured at preset time intervals, preferably 3 to 8 minutes. When the actual pH value deviates from the target value by more than a preset deviation threshold, an acid or alkali pump is activated for adjustment. After pumping is complete, sufficient mixing is allowed before re-measurement. This intermittent control strategy avoids pH fluctuations caused by continuous adjustment, thus improving control stability.

[0047] This step established a complete time-series data stream of process parameters and enabled precise control of enzyme addition and pH value.

[0048] Step 2.3, Dynamic prediction of product quality based on LSTM network; The system receives the process parameter timing data stream collected in step 2.2. When the extraction reaches a preset time point, preferably between 40% and 60% of the total extraction time, the quality prediction function is activated. A pre-trained Long Short-Term Memory (LSTM) network is used to predict the final product quality.

[0049] This LSTM quality prediction model was pre-trained on historical production data from the enterprise. The training dataset contains complete time-series data of process parameters for 200 to 500 batches and corresponding quality inspection results, divided into training, validation, and test sets in a 7:2:1 ratio. The model employs a 2-3 layer LSTM structure and 1-2 layers of fully connected layers. The preferred number of neurons in the LSTM layers is 64-128 in the first layer and 32-64 in the second layer. The LSTM layers can capture long-term dependencies in the time-series data of process parameters and extract short-term patterns and long-term dependency features. The preferred number of neurons in the fully connected layers is 16-32, mapping the extracted features to predicted values ​​of quality indicators. The model's prediction accuracy meets the preset requirements, with a preferred prediction accuracy greater than 85%, and the average absolute error of each quality indicator controlled within 8%.

[0050] The time-series data of process parameters collected in step 2.2 for the preceding period are used as input, including multiple feature dimensions such as temperature curves, pH changes, enzyme addition records, pressure changes, and flow rate changes. This data is normalized to form a time-series matrix, which is then input into the LSTM model. The model quickly outputs predicted values ​​for four quality indicators: sulfide content, polyphenol content, volatile flavor compound content, and color parameters. The prediction time is preferably 5 to 15 seconds.

[0051] The predicted values ​​are compared with the target quality indicators to determine if there is a risk of quality non-compliance, and the quality deviation and deviation rate are calculated. Simultaneously, the predicted values ​​are compared with the expected effect evaluation output in step 1.5 to verify the accuracy of the initial process parameter recommendations. Based on the current process status and historical evolution patterns, a quality evolution trend prediction is output, forecasting the change curve of quality indicators in the future period, providing a time window and adjustment direction for process parameter optimization.

[0052] This step yields product quality prediction results, including predicted values, quality deviations, deviation rates, and evolution trends for four quality indicators: sulfide content, polyphenol content, volatile flavor compound content, and color parameters. This provides a reliable basis for intelligent optimization of process parameters.

[0053] Step 3: Based on the product quality prediction results, detect the risk of quality deviation, calculate the process parameter adjustment plan, perform safety constraint checks, and then execute the parameter adjustment to obtain the final extraction state information; Specifically, the following steps are included: Step 3.1, Quality Risk Assessment and Optimization Initiation; Receive the product quality prediction results output from step 2.3, including predicted values, quality deviations, deviation rates, and evolution trend prediction curves for four quality indicators: sulfide content, polyphenol content, volatile flavor compound content, and color parameters. Perform a risk assessment on each predicted value of the quality indicators.

[0054] The risk assessment criteria are as follows: For each quality indicator, determine whether the predicted value is within the target range. The target range is determined according to the product quality standard. If the predicted value exceeds the target range, the indicator is considered to be at risk. Calculate the deviation rate, which is equal to the difference between the predicted value and the target value divided by the target value. When the absolute value of the deviation rate exceeds a preset threshold, the indicator is also considered to be at risk. The preferred range for the deviation rate threshold is 5% to 15%. If any quality indicator is considered to be at risk, the batch is considered to have a risk of failing to meet quality standards, and process parameter optimization needs to be initiated.

[0055] Upon detecting a risk of substandard quality, the process parameter optimization module is activated. This module employs Model Predictive Control (MPC) algorithm, which, while meeting constraints, optimizes the control sequence over a future period to make the predicted quality indicators as close as possible to the target values, thereby achieving precise control of the output.

[0056] This step completes the quality risk assessment and activates the optimization module, preparing for process parameter optimization.

[0057] Step 3.2, Problem construction for prediction model establishment and optimization; The MPC algorithm employs a pre-established prediction model for the extraction process. Based on historical production data, this model uses a linear state-space representation. State variables include current temperature, pH value, extraction time, and cumulative enzyme addition. Control variables include temperature adjustment, time adjustment, and enzyme addition adjustment. Output variables are predicted quality indicators such as sulfide content and polyphenol content. The model's prediction accuracy meets preset requirements, with an optimal prediction accuracy greater than 85%, accurately describing the dynamic characteristics of the extraction process. The model uses discrete-time processing, with a sampling period consistent with the monitoring interval, enabling real-time status updates and continuous optimization.

[0058] Based on the prediction model, the MPC algorithm constructs an optimization problem. Within the prediction time domain, it seeks the optimal control sequence that makes the predicted quality index as close as possible to the target value, while ensuring smooth changes in the control input and avoiding drastic adjustments. The optimization objective function includes a quality tracking error term and a control change penalty term. The quality tracking error term measures the deviation between the predicted quality index and the target value, while the control change penalty term measures the magnitude of changes in the control input between adjacent time points. By adjusting the weight coefficients of these two terms, a balance is struck between quality control accuracy and control smoothness. The preferred prediction time domain is 30 to 90 minutes, and the preferred control time domain is 10 to 30 minutes.

[0059] The optimization problem must meet certain constraints. Hard constraints include: temperature range of 35 to 70 degrees Celsius, time range of 120 to 300 minutes, enzyme addition range of 0.03% to 0.5% of raw material mass, and pH range of 5.5 to 7.5. Exceeding these ranges may lead to product quality problems or safety risks. Soft constraints include: energy consumption increase limit of no more than 15%, time extension limit of no more than 40 minutes, and cost increase limit of no more than 8%, etc., which are economic and efficiency constraints.

[0060] This step establishes a prediction model and constructs an optimization problem, providing a mathematical foundation for solving the optimal control sequence.

[0061] Step 3.3, Solving for the optimal control sequence and evaluating its effectiveness; The optimization problem constructed in step 3.2 is received and solved using quadratic programming. Since the objective function is quadratic and the constraints are linear inequalities, this problem is a convex optimization problem with a global optimum. Numerical solutions are performed using the interior-point method or the active set method. These methods approximate the optimal solution iteratively, updating the control variables in each iteration until the convergence condition is met. The computation time is prioritized to be within 10 seconds to meet real-time control requirements.

[0062] The system receives current status information, including extraction time, current temperature, current pH value, and predicted quality indicators, and retrieves raw material characteristic information. Based on multi-dimensional status information, the MPC algorithm calculates the optimal control sequence. The optimization results display the adjustment scheme for process parameters, including specific values ​​for temperature adjustment, time adjustment, and enzyme addition adjustment. The control sequence adopts a rolling optimization strategy, that is, only the first control action is executed at a time, and the status is remeasured and a new optimization problem is solved at the next sampling time, achieving feedback correction.

[0063] Simultaneously, the system outputs a prediction of the adjustment effect of the optimization scheme. Based on the prediction model, the extraction process after adopting the optimization scheme is simulated, and the change trajectory of each quality indicator is predicted. It is determined whether the predicted values ​​of each quality indicator after adjustment fall within the target range, and whether the increase in energy consumption, time, and cost is within an acceptable range. The adjustment effect prediction includes the predicted values ​​of each quality indicator after adjustment, the increase in energy consumption, the extension of time, and the increase in cost, providing decision-making reference for operators. If the adjustment effect prediction is not ideal, the optimization parameters can be adjusted and the solution recalculated.

[0064] This step yields the optimal process parameter adjustment scheme and its adjustment effect prediction, providing a scientific basis for safe execution.

[0065] Step 3.4, Multi-layer safety constraint check; Before adjusting the parameters, conduct multi-layered safety constraint checks to ensure that the adjustment plan will not lead to safety risks or quality problems.

[0066] Check the hard constraints, which are the safety restrictions that cannot be violated. Temperature hard constraint check: Determine if the adjusted temperature is within the equipment's allowable range and meets food safety standards; Time hard constraint check: Determine if the adjusted total extraction time exceeds the equipment's continuous operating time limit; Enzyme dosage hard constraint check: Determine if the adjusted total enzyme dosage is within the safe operating range and meets food additive usage standards; Pressure hard constraint check: Determine if the current extraction tank pressure is within the safe range. If all hard constraint checks pass, proceed to the soft constraint check.

[0067] Checking soft constraints refers to economic and efficiency constraints that can be violated to a reasonable extent. Energy consumption soft constraint check: Determine if the increase in energy consumption due to temperature adjustments is within an acceptable economic range; if so, record the increase and assess its acceptability. Efficiency soft constraint check: Determine the impact of time adjustments on production efficiency and assess whether the efficiency loss incurred to ensure quality standards is acceptable; if the impact is significant, record the efficiency decrease. Raw material consumption soft constraint check: Determine if parameter adjustments will increase raw material consumption.

[0068] This step confirms whether the adjustment plan meets all safety constraints and assesses the acceptability of safety risks and economic and efficiency impacts.

[0069] Step 3.5: Parameter adjustment execution and effect monitoring; After all constraint checks pass, parameter adjustment execution instructions are generated, including adjusting the target temperature of the temperature controller, adjusting the remaining time of the timer, and replenishing enzyme preparations for the enzyme addition controller. These instructions are then sent to each controller in the execution layer.

[0070] Upon receiving the new target temperature, the temperature controller smoothly adjusts the heating power using a PID algorithm. To avoid the adverse effects of sudden temperature changes on product quality, a ramp-up heating strategy is adopted, where the temperature gradually increases at a preset rate, preferably 0.5 to 2 degrees Celsius per minute. During the heating process, the PID controller continuously monitors temperature changes and dynamically adjusts the heating power to ensure a smooth and stable heating curve.

[0071] The timer updates the remaining extraction time, and the control interface displays the estimated completion time, allowing operators to plan subsequent work accordingly. The enzyme addition controller automatically replenishes enzyme preparation at preset time points, calculates the total amount of raw materials in the tank, calculates the total amount of enzyme preparation that needs to be added, and controls the metering pump to accurately pump the corresponding volume of enzyme solution.

[0072] After parameter adjustments, continuously monitor the process, periodically rerun the quality prediction model, and dynamically evaluate the adjustment effect, with an optimal monitoring interval of 10 to 30 minutes. When the model predicts that the final quality indicators have entered the target range, it is determined that no further adjustments are needed, and the current parameters are maintained until the extraction is complete. If the quality indicators still do not meet the standards, restart the optimization process for a second adjustment.

[0073] After extraction is complete, the final extraction status information is output, including the adjusted temperature, pH value, actual extraction time, total amount of enzyme added, predicted quality index values, and parameter adjustment execution records.

[0074] This step ensures the safe optimization and execution of process parameters, guaranteeing product quality while ensuring production safety.

[0075] Step 4: Based on the final state information of the extraction, separation and purification are carried out and monitored in real time. A three-stage separation strategy is adopted to remove impurities and obtain purified scallion oil product. Specifically, the following steps are included: Step 4.1: Adaptively adjust the separation parameters based on upstream information; After extraction, the mixture enters the separation and purification stage. The final extraction status information output from step 3.5 is received, including adjusted temperature, pH value, actual extraction time, total enzyme addition, predicted quality indicators, and parameter adjustment execution records. A multi-stage linkage control strategy is adopted, transmitting key parameters and status information of the extraction stage to the separation control module through a standardized data interface. The standardized data interface encapsulates data in JSON format and transmits it via the industrial Ethernet protocol, with transmission latency preferably controlled within 50 to 200 milliseconds to ensure data real-time performance and integrity.

[0076] After receiving information from upstream, the separation control module automatically calculates and adjusts the separation process parameters based on a pre-established linkage control model. This linkage control model is built upon historical production data and employs multiple linear regression or neural network methods. The inputs are the key parameters of the extraction process, and the outputs are the optimal parameter settings for the separation process. The model's prediction accuracy meets preset requirements, with a preferred correlation coefficient greater than 0.80.

[0077] The current viscosity of the mixture is estimated based on the final extraction temperature. Viscosity estimation uses an empirical formula, where viscosity has an exponential relationship with temperature. Viscosity equals the reference viscosity multiplied by an exponential term of the natural constant, where the exponential term is the temperature coefficient multiplied by the temperature difference. In this embodiment, when the extraction temperature is 58 degrees Celsius, the estimated viscosity of the mixture is approximately 45 to 55 mPa·s. Based on the viscosity estimation result, the separation module automatically adjusts the centrifuge speed. For higher viscosity, the centrifuge speed is increased to enhance the separation effect; a preferred speed adjustment range is 2800 to 3500 rpm. For lower viscosity, the centrifuge speed is decreased to save energy and reduce equipment wear; a preferred speed adjustment range is 2000 to 2800 rpm. Based on the viscosity estimation value, the centrifuge speed is set to 3200 rpm.

[0078] The effect of the mixture's acidity or alkalinity on oil-water separation is determined by the final pH value of the extraction. When the pH value is in the range of 6.0 to 7.0, the interfacial tension between oil and water is moderate, which is conducive to oil-water separation, and there is no need to adjust the pH value of the separation stage. When the pH value deviates from this range, pH adjustment is required before separation, with the preferred adjustment target being 6.5 ± 0.3. When the final pH value of the extraction is 6.5, it is within the optimal range, and it is confirmed that there is no need to adjust the pH value of the separation stage, and the original setting should be maintained.

[0079] The extraction effect is judged based on the predicted quality index values ​​output in step 3.5. When the predicted sulfide content reaches the target range, it indicates sufficient extraction, and the separation process can proceed according to the standard procedure. When the predicted sulfide content is low, it indicates insufficient extraction, and special measures need to be taken in the separation process, such as lowering the separation temperature to reduce sulfide volatilization loss or extending the settling time to improve the recovery rate. When the predicted sulfide content is 1.72 grams per kilogram, it has entered the target range, and no special measures are needed in the separation process; it can proceed according to the standard procedure.

[0080] Linked application of extraction time information: The actual extraction time is used to determine the degree of component release from the raw material. A longer extraction time indicates sufficient component release and a higher solid particle content in the mixture, requiring extended settling time or increased centrifugation intensity. A shorter extraction time indicates insufficient component release and a lower solid particle content in the mixture, allowing for a shorter separation time to improve production efficiency. When the actual extraction time is 195 minutes, which is within the normal slightly longer range, the separation module automatically adjusts the settling time from the standard 30 minutes to 35 minutes and the centrifugation time from the standard 15 minutes to 18 minutes to ensure effective separation.

[0081] This step enables adaptive adjustment of separation parameters, achieving synergistic optimization of the extraction and separation processes, and providing optimized parameter settings for the subsequent three-stage separation process.

[0082] Step 4.2, precise control of the three-stage separation process; The separation process employs a three-stage separation strategy to progressively remove solid particles and moisture, yielding a pure oil phase product. The three-stage separation includes coarse separation, centrifugal separation, and precision filtration, with each stage targeting impurities of different particle sizes to achieve efficient separation.

[0083] First-stage coarse separation: Gravity sedimentation removes large solid particles. This specifically includes the following steps: The mixture is pumped from the extraction tank to the settling tank via pipeline. The pumping flow rate is controlled within a preset range, preferably 0.5 to 2 cubic meters per hour, to avoid emulsification of the mixture due to excessive flow rate. After entering the settling tank, the mixture is allowed to settle by gravity. Based on the linkage adjustment results in step 4.1, the settling time is set to 35 minutes.

[0084] The settling temperature is precisely controlled via a jacketed heating or cooling system. The temperature control device employs a PID control algorithm, with a preferred control accuracy of ±1 to 3 degrees Celsius. The settling temperature is set at 50 degrees Celsius, at which temperature the oil viscosity is approximately 30 to 40 mPa·s, ensuring a stable settling rate while preventing excessive temperature from causing loss of volatile flavor compounds. Too low a temperature leads to increased viscosity and reduced settling rate; too high a temperature causes flavor compounds to volatilize, resulting in a decline in product quality.

[0085] The settling process follows Stokes' Law. According to this law, the settling velocity of solid particles is equal to 2 times the acceleration due to gravity, the square of the particle radius, the density difference between the particle and the fluid, divided by 9 times the fluid dynamic viscosity. In this embodiment, the density of the onion residue particles is approximately 1.2 to 1.4 g / cm³, the density of the oil phase is approximately 0.91 to 0.93 g / cm³, and the density difference is approximately 0.27 to 0.49 g / cm³. Large solid residue particles preferably have a diameter greater than 100 micrometers, which result in a faster settling velocity under gravity, preferably 0.5 to 2 mm / s.

[0086] Solid particles gradually settle to the bottom of the tank under gravity, forming a sludge layer. The thickness of the sludge layer is monitored by a liquid level sensor. When the sludge layer thickness reaches a preset threshold, preferably 10% to 20% of the tank height, the sludge discharge program is automatically initiated. Sludge discharge is carried out through a bottom sludge discharge valve, which is pneumatically controlled and preferably opens for 30 to 90 seconds to ensure complete discharge of the sludge. The discharged sludge can be used as a feed additive or organic fertilizer raw material, achieving comprehensive resource utilization.

[0087] The upper clarified oil phase flows automatically into the next stage centrifugal separator through the overflow port at the top of the tank. The overflow port height is designed to ensure that only the upper clarified oil phase flows out, preventing bottom sediment from being carried out. After coarse separation, the solid particle content in the oil phase is reduced from the initial 5% to 8% to 0.5% to 1.5%, with a removal efficiency of 80% to 90%.

[0088] Second-stage centrifugation: Centrifugal force removes fine particles and some moisture. This specifically includes the following steps: The clarified oil phase enters a horizontal spiral centrifuge for fine separation. The centrifuge model is LW450 horizontal spiral sedimentation centrifuge, with a drum diameter of 450 mm, a length-to-diameter ratio of 3 to 4, and a processing capacity of 0.3 to 1.0 cubic meters per hour.

[0089] The centrifuge speed is set to 3200 revolutions per minute based on the linkage adjustment result in step 4.1. At this speed, the centrifugal acceleration generated on the inner wall of the drum is approximately 35 times the acceleration due to gravity. The calculation formula is: centrifugal acceleration equals the square of the rotational speed multiplied by the drum radius divided by 900. Under this centrifugal force, denser solid particles and water are thrown towards the outer wall of the centrifuge, while the less dense oil phase remains in the inner layer.

[0090] The centrifuge is equipped with a screw propeller, and there is a speed difference between the screw propeller and the rotating drum, preferably 5 to 15 revolutions per minute. The screw propeller pushes the solid particles and water deposited on the outer wall toward the discharge port, while the clarified oil phase flows out from the central outlet. This continuous discharge method ensures that the centrifuge can operate stably for a long time without frequent shutdowns for cleaning.

[0091] The efficiency of centrifugal separation depends on centrifugal force, residence time, and density difference. The oil phase density is approximately 0.91 to 0.93 g / cm³, the aqueous phase density is approximately 1.00 to 1.02 g / cm³, and the solid particle density is approximately 1.2 to 1.4 g / cm³. This significant density difference is beneficial for separation. The residence time of the material in the centrifuge is controlled by the feed flow rate. Based on the linkage adjustment results in step 4.1, the centrifugation time is set to 18 minutes to ensure thorough separation.

[0092] After centrifugation, the solid particle content in the oil phase decreased from 0.5% to 1.5% to 0.05% to 0.15%, and the moisture content decreased from the initial 3% to 5% to 0.5% to 1.0%, achieving a removal efficiency of 85% to 95%. The separated aqueous phase and solid mixture are discharged through the slag discharge port and can be further processed to recover residual oil.

[0093] Third-stage precision filtration: Multi-layer filter membranes remove minute impurities. Specifically, it includes the following steps: The oil phase after centrifugation enters a precision filtration system for final purification. The filtration system employs a three-stage cascade filtration process to remove fine particles step by step. The first stage, coarse filtration, uses a 316L stainless steel filter screen with a pore size of 10 microns and a filtration area of ​​2 to 5 square meters; the second stage, medium filtration, uses a polypropylene filter element with a pore size of 5 microns and a filtration area of ​​1 to 3 square meters; the third stage, fine filtration, uses a polyethersulfone or polyvinylidene fluoride microporous filter membrane with a pore size of 1 micron and a pleated design.

[0094] The filtration process takes place in a closed system to prevent oxidation and contamination. The system is protected by nitrogen gas, with a nitrogen purity preferably greater than 99.9% and an oxygen content controlled below 0.1%. The filtration pressure is controlled by a pressure regulating valve, preferably between 0.15 and 0.25 MPa. The filtration rate follows Darcy's law, meaning the filtration rate is directly proportional to the pressure difference and inversely proportional to the membrane resistance and fluid viscosity. In this embodiment, at a pressure of 0.2 MPa, the filtration rate is approximately 50 to 100 liters per square meter per hour.

[0095] The system is equipped with a differential pressure monitoring device to monitor the pressure difference across the filter membrane in real time. When the pressure difference exceeds a preset threshold (preferably 0.15 to 0.25 MPa), an automatic alarm will sound to prompt for filter replacement. The filter replacement cycle is determined based on the throughput, with a preferred replacement cycle of 50 to 200 batches. After filtration, the solid particle content in the oil phase is reduced to below 0.01%, and the product transmittance is preferably greater than 95%, achieving a clear and transparent standard.

[0096] During the three-stage separation process, multiple online sensors are deployed to monitor the oil phase quality in real time. The sensor acquisition frequency is set to a second preset acquisition frequency, with an optimal value of 0.1 to 1 Hz.

[0097] Online turbidity meters are installed at the centrifuge outlet and the precision filter outlet, using the 90-degree scattering light principle to measure the concentration of suspended particles. The target turbidity value is preferably less than 3 NTU, and the measurement accuracy is preferably ±0.5 NTU. An online moisture meter is installed at the centrifuge outlet, using the capacitance method to measure the moisture content. The target moisture content is preferably less than 0.2%, and the measurement accuracy is preferably ±0.05%. Temperature sensors and flow meters are installed at the inlet and outlet of each stage of the separation equipment to monitor temperature and flow rate changes in real time.

[0098] When the monitored value exceeds the target value, the parameter adjustment program is automatically triggered. The adjustment strategy uses a fuzzy control algorithm to determine the adjustment range based on the magnitude and trend of the deviation. When turbidity exceeds the standard, the centrifuge speed is increased first, by 100 to 200 revolutions per minute each time, with a maximum of 4000 revolutions per minute; after the speed reaches the upper limit, the feed flow rate is reduced by 10% to 20%. When moisture exceeds the standard, the centrifuge disc angle is adjusted (30 to 45 degrees) or the speed difference is adjusted (5 to 15 revolutions per minute). When the temperature exceeds 60 degrees Celsius, the cooling device is activated, with the cooling rate controlled at 2 to 5 degrees Celsius per minute, and cooling is stopped when the temperature drops to 50 to 55 degrees Celsius.

[0099] This step achieves precise control and real-time monitoring of the three-stage separation process, resulting in a clear, transparent, purified scallion oil product with low moisture and impurity content. The product has a solid particle content of less than 0.01%, a moisture content of less than 0.2%, a turbidity of less than 3 NTU, and a light transmittance of greater than 95%, meeting the quality standards for high-quality scallion oil.

[0100] Step 5: Conduct quality testing on the purified scallion oil product, generate a quality testing report, organize and update production data, and obtain an updated knowledge base; Specifically, the following steps are included: Step 5.1, quality assessment combining online and offline detection; The purified onion oil product output from step 4.2 is received and proceeds to the quality inspection stage. A combination of online and offline testing methods is used to comprehensively evaluate the product quality.

[0101] Online inspection is conducted in real time using sensors installed on the production line. Color parameters are measured using a colorimeter, which employs the tristimulus principle: the sample is illuminated by three standard light sources—red, green, and blue—and the intensity of the reflected light is measured and converted into L, a, and b values ​​in the CIE Lab color space. The L value represents lightness, ranging from 0 (black) to 100 (white); the a value represents red-green hue, with positive values ​​for red and negative values ​​for green; and the b value represents yellow-blue hue, with positive values ​​for yellow and negative values ​​for blue. Color is a crucial sensory indicator for scallion oil; a satisfactory color indicates that the product has not been excessively oxidized or charred. The measurement results are compared with target ranges: L value 44 to 47, a value -3 to -2, and b value 27 to 30, to determine if the color is acceptable.

[0102] Viscosity is measured using an online viscometer based on the rotational method, where a rotor rotates within the sample, and the resistance torque experienced by the rotor is measured and converted into a viscosity value. Viscosity reflects the oil's fluidity and taste; excessively high viscosity indicates the presence of large molecules or polymers, while excessively low viscosity indicates the presence of solvents or light components. The measurement results are compared with a target viscosity range, preferably 25 to 45 mPa·s (measured at 25 degrees Celsius), to determine if the viscosity is within acceptable limits.

[0103] The refractive index is measured using an online refractometer. The refractometer employs the critical angle method, meaning that when light enters the glass prism from the oil phase, it refracts at the interface, and the angle of refraction is proportional to the refractive index. The refractive index reflects the purity and composition of the oil; an abnormal refractive index may indicate the presence of other substances in the oil. The measurement results are compared with a target range, preferably 1.465 to 1.475 (measured at 20 degrees Celsius), to determine if the refractive index is within acceptable limits.

[0104] Offline testing, conducted through laboratory analysis, provides a more accurate and comprehensive quality assessment. Samples are sent to the laboratory, where gas chromatography (GC) is used to determine the sulfide content. The principle of GC is that the sample is vaporized and enters the chromatographic column. Different components have different retention times in the column and are detected sequentially by detectors. Qualitative analysis is based on retention time, and quantification is based on peak area. A flame photometric detector (FPD) is used, which has high sensitivity and selectivity for sulfur-containing compounds. The measured results are compared with predicted values ​​to verify the accuracy of the quality prediction model. Sulfides are the main functional components of scallion oil; meeting the content standards indicates excellent product quality.

[0105] The polyphenol content was determined using spectrophotometry. The principle of spectrophotometry is that polyphenols react with Folin-Ciocalteu reagent to form a blue compound, which has maximum absorption at a specific wavelength; the absorbance is directly proportional to the polyphenol content. The measurement results were compared with the target range to determine whether the polyphenol content was within acceptable limits. Polyphenols are important antioxidant components of scallion oil; an appropriate content indicates that the product has good health benefits.

[0106] Volatile flavor compounds were determined using headspace solid-phase microextraction combined with gas chromatography-mass spectrometry (GC-MS). The principle of this method is as follows: the sample is heated in a sealed container, releasing volatile components into the headspace, where they are adsorbed by a solid-phase microextraction fiber. The fiber is then inserted into the GC inlet, where the volatile components undergo thermal desorption and separation into the chromatographic column. The mass spectrometer detector performs qualitative and quantitative analysis based on the mass spectrum. The results are compared with the target range to determine whether the flavor compound content is within acceptable limits. Volatile flavor compounds are the source of the aroma of scallion oil; an appropriate content indicates that the product has a rich scallion flavor.

[0107] Based on the combined results of online and offline testing, all quality indicators of this batch of products meet the target requirements and are therefore deemed qualified. A quality inspection report is automatically generated, including the batch number, testing date, testing items, measured values, target range, qualification criteria, and testing personnel. The report is stored electronically in the database and can be exported as a PDF for customer review or regulatory approval.

[0108] Step 5.2: Automatic organization of production data and model update; Receive the quality inspection report output in step 5.1. After the quality inspection is completed, start the closed-loop feedback mechanism and use the complete data of this batch for model optimization and updating.

[0109] Complete data for this batch is automatically organized and stored in the historical database. Regardless of whether the product quality is up to standard, all batch data is saved. Qualified batches are used to optimize model performance, while unqualified batches are used to analyze the causes of failure and improve control strategies.

[0110] The stored data includes the following: raw material characteristic parameters, including moisture content, soluble solids content, initial sulfide content, variety code, and origin code; time-series data of process parameters, including temperature curves, pH curves, pressure curves, flow rate curves, and enzyme addition records; quality prediction results; actual quality test results, including sulfide content, polyphenol content, volatile flavor compound content, and color parameters; parameter adjustment records; and abnormal event records.

[0111] The actual quality inspection results in step 5.1 are compared with the expected effect evaluation in step 1.5 to calculate the prediction deviation and verify the accuracy of the raw material characteristic detection and process parameter recommendation models. Simultaneously, the quality prediction results in step 2.3 are compared with the actual inspection results in step 5.1 to calculate the prediction error and verify the accuracy of the LSTM quality prediction model.

[0112] These data are automatically labeled with information including batch type, quality grade, raw material characteristics, process features, quality results, and optimization effects. The labeling uses a multi-tag classification method, allowing each batch to have multiple tags for easy subsequent retrieval and analysis. For example, it can retrieve all batches that underwent effective mid-term temperature adjustments to analyze the optimal timing and magnitude of these adjustments; it can also retrieve non-conforming batches to analyze the causes and patterns of quality control failures.

[0113] The system counts the current cumulative number of historical batches. When the number reaches the third preset batch count, the model update mechanism is triggered. The third preset batch count is a set model update threshold, with an optimal value of 15 to 30 batches. When the cumulative data reaches this threshold, it indicates that there is enough new data available for model optimization, and the model retraining process is automatically initiated.

[0114] Model retraining employs an incremental learning approach. The advantage of incremental learning is that it retains the knowledge already learned by the original model while only learning new patterns in the new data, avoiding the high computational cost and knowledge forgetting issues caused by training from scratch. Specifically, the network structure and parameters of some layers of the original LSTM model are retained; these parameters have already learned the general feature extraction capabilities of time-series data. Only the parameters of other layers are fine-tuned, enabling the model to learn new patterns in the latest data, such as new raw material characteristic ranges, new combinations of process parameters, and new quality response relationships.

[0115] The fine-tuning process uses the latest accumulated data, divided proportionally into training and validation sets, with an optimal ratio of 7-9:1-3. Training employs an optimization algorithm with a relatively small learning rate, preferably 0.05-0.2 times the initial learning rate, as the model is already close to optimal and requires fine-tuning in small steps. The batch size is preferably 8-32, and the number of training epochs is preferably 30-100. The training process is performed on a cloud server.

[0116] After training, evaluate the performance of the new model on the test set. The test set contains recent data that was not used in model training. Evaluation metrics include prediction accuracy, mean and standard deviation of prediction error, etc. Compare the performance of the new model with the old model to determine if there is a significant improvement.

[0117] If the new model outperforms the old model, the evaluation results will be recorded in the model performance log, and a performance improvement report will be generated. The report will include: update time, amount of data used, training time, performance improvement, and typical case analysis. The report will be available for technical personnel to review, monitor model performance evolution trends, and evaluate the effectiveness of continuous learning.

[0118] Once the new model passes evaluation, it is automatically downloaded to the edge device, replacing the old model. The replacement process uses a hot update method, meaning the model switch is completed without halting production. Specifically, the new model is first loaded into memory; then, between batches, the model pointer is switched from the old model to the new model, releasing the old model's memory. The entire switchover process does not affect production continuity. If the new model's performance does not meet expectations, the old model is retained and used, the reason for the update failure is recorded, and retraining is performed after sufficient data has been accumulated.

[0119] Step 5.3: Extraction and knowledge accumulation of excellent process experience; Receive the complete batch of data compiled in step 5.2, conduct in-depth analysis of process parameters and quality results, extract best practices and accumulate them in the knowledge base.

[0120] Identify the key success factors or reasons for failure in this batch. For qualified batches, analyze key factors such as the effectiveness of process parameter adjustments, the accuracy of quality prediction, and the timeliness of parameter optimization; for unqualified batches, analyze the reasons for quality out-of-control issues, the sources of prediction deviations, and the inadequacy of optimization strategies. Taking this example, the key success factors for this batch were identified as follows: In the middle stage of extraction, the quality prediction model promptly detected that the predicted value of the quality index was lower than the target lower limit, and parameter optimization was immediately initiated. The optimal adjustment scheme was calculated using the MPC algorithm, and adjustments were made to the temperature, time, and enzyme addition amount. After the adjustment, the predicted quality value entered the target range, and the final actual test results deviated very little from the predicted values, verifying the effectiveness of prediction and adjustment.

[0121] The identified experiences are extracted into rules and added to the intelligent knowledge base. The rules are expressed as follows: When the raw material characteristic parameters are within a specific range, if the predicted quality index during the mid-stage extraction is lower than the target lower limit and the deviation is within a certain range, it can be corrected by adjusting the temperature and extending the extraction time, which is expected to bring the final quality index to the target requirements. The rules also include information such as applicable conditions, adjustment range, expected effects, and confidence level, with the confidence level calculated based on the success rate of similar historical cases.

[0122] These rules will be prioritized when similar situations arise in the future. Specifically, when the current raw material characteristics and process status are detected to match the applicable conditions of a rule, the knowledge base is automatically searched to find the rule with the highest matching degree. The adjustment scheme recommended by this rule is then used as the initial or reference solution for the optimization algorithm, accelerating the optimization convergence speed, improving decision-making accuracy, and allowing operators or automated control devices to quickly adopt the suggestion without waiting for the optimization algorithm to calculate, thus improving response speed.

[0123] Simultaneously, analyze any abnormalities in this batch. Record the abnormal events, analyze the causes, and document the lessons learned as precautions. These precautions will serve as reminders for operators and automated control devices in future production, enabling them to prepare accordingly and prevent similar problems from recurring.

[0124] Through the three sub-steps in step 5, a closed-loop process of quality inspection, data feedback, and knowledge accumulation is completed, resulting in a quality inspection report, an optimized prediction model, and an updated knowledge base. The quality inspection report is used for traceability information generation in step 6, while the optimized prediction model and updated knowledge base are applied to subsequent batch production through a closed-loop feedback mechanism, enabling continuous system learning and performance improvement.

[0125] Step 6: Generate traceability information based on the quality inspection report, perform data classification processing and digital signature protection, and obtain traceability data package and QR code identifier; Specifically, the following steps are included: Step 6.1: Structured organization and hierarchical processing of traceability data; After receiving the quality inspection report output in step 5.1 and confirming that the product quality has passed inspection, it proceeds to the packaging and traceability labeling stage. An enhanced traceability scheme, namely Scheme B, is adopted, which uses digital signature technology to achieve quality traceability.

[0126] The key traceability information for this batch was compiled into structured data and classified according to the degree of public disclosure.

[0127] The data classification rules are as follows: Public data (Level 0): stored and transmitted in plaintext, directly displayed on product packaging QR codes or traceability platforms; Semi-public data (Level 1): stored in plaintext, but can only be viewed after authorization verification, such as by enterprise customers or regulatory authorities; Confidential data (Level 2): ​​stored using AES-256 encryption, authorized for access only by internal systems, and not released to the public.

[0128] Publicly available data includes: batch number, which is unique and consists of the date, production line number, and batch sequence number; production date; origin of raw materials; type of raw materials; quality grade, determined based on a comprehensive assessment of sulfide content, polyphenol content, etc.; shelf life, determined based on accelerated aging tests; manufacturer's name; production license number; and product standard number. This publicly available data will be completely transparently presented to consumers, helping them understand the basic information about the product.

[0129] Semi-public data includes: extraction temperature range (only the range is disclosed, without revealing the specific temperature curves, thus showcasing the technological level while protecting process details); extraction time range; pH range; sulfide content, accurate to two decimal places, demonstrating product quality; polyphenol content; volatile flavor compound content; color parameters; raw material utilization rate, demonstrating production efficiency and environmental protection level; energy consumption, demonstrating energy-saving level, etc. This semi-public data will be selectively presented to consumers or specific clients to meet different levels of information needs.

[0130] The confidential data includes: detailed process parameter time-series data, including all data points such as temperature, pressure, and pH value collected at the first preset sampling frequency; enzyme formulations, including enzyme type, source, and activity unit; specific dosage and addition time; optimization algorithm parameters and models; supplier information; cost data, and other trade secrets. This confidential data is stored and used only internally within the company and will not be disclosed to the public to protect the company's trade secrets and competitive advantage.

[0131] This step completes the structured organization and hierarchical processing of traceability data, forming a hierarchical traceability dataset, which prepares for subsequent data signing and release.

[0132] Step 6.2, Data integrity protection based on digital signatures; To ensure the authenticity and integrity of traceability data, both publicly available and semi-publicly available data are digitally signed.

[0133] The publicly available and semi-publicly available data in this batch are serialized into JSON format strings. The SHA-256 hash algorithm is used to calculate the digital fingerprint of the data. This algorithm maps data of arbitrary length to a fixed-length 256-bit hash value, represented as 64 characters in hexadecimal. The SHA-256 algorithm has properties such as one-wayness, collision resistance, and avalanche effect, ensuring that any tampering with the original data will result in a completely different hash value, thus being detected.

[0134] The hash value is digitally signed using the company's private key. The digital signature employs the RSA algorithm, with a preferred key length of 2048 bits or more. The signing process involves encrypting the hash value using the company's private key to obtain the digital signature. Digital signatures serve two crucial purposes: proving the authenticity of the data source (only the company holding the private key can generate a valid signature); and proving the integrity of the data (any tampering with the data will cause signature verification to fail).

[0135] After signing is completed, the public data, semi-public data, hash value, and digital signature are packaged, stored in the enterprise database, and uploaded to the cloud server. The cloud server provides a public query interface, allowing consumers to query product information by batch number and verify the authenticity of the data using the enterprise's public key.

[0136] Different protection strategies are employed for confidential data. Confidential data includes detailed process parameter timing data, enzyme formulations, optimization algorithm parameters, supplier information, and cost data—all commercially sensitive information. This data is stored encrypted only on the company's local server and is not publicly disclosed. The encryption algorithm uses AES-256 symmetric encryption, employing a 256-bit key to encrypt the data. The key is kept by the company's security administrator and stored in a hardware security module (HSM). This device provides secure key generation, storage, and use, preventing key leakage. Keys are replaced periodically, preferably every 1 to 6 months. Old keys are used to decrypt historical data, and new keys are used to encrypt new data.

[0137] This step completes the digital signature and hierarchical storage of traceability data, forming a traceability data package with tamper-proof capabilities, ensuring the authenticity and integrity of publicly available data, while protecting the confidentiality of confidential data.

[0138] Step 6.3, Traceability Identification Generation and Consumer Inquiry Service; Based on the data signing and storage completed in step 6.2, a QR code traceability identifier is generated and a consumer query service is deployed.

[0139] The QR code uses the QR code format, with an error correction level prioritized at H, tolerating 30% damage to the code area to ensure readability during transportation and use. The QR code size is preferably 2 to 5 centimeters, containing information such as the batch number and a query link. The QR code is printed on the product packaging label using a waterproof and ink-resistant material to ensure clarity and durability. The label uses self-adhesive material and is affixed to a prominent position on the packaging for easy scanning by consumers. Each bottle of product has a unique QR code to prevent mass counterfeiting.

[0140] The query link leads to the company's traceability query webpage, which is deployed on a cloud server, adopts a responsive design, and supports access from various devices such as mobile phones, tablets, and computers. The webpage can be developed using Node.js, Java, or other technology stacks, providing a user-friendly interface and a smooth interactive experience.

[0141] After consumers scan the QR code on the product packaging with their mobile phones, the phones automatically recognize the query link and open a browser to redirect to a traceability query webpage. The webpage uses a layered display strategy to meet the information needs of different consumers.

[0142] The webpage displays basic product information, including publicly available data such as product images, batch number, production date, raw material origin, raw material type, quality grade, shelf life, and manufacturer. The information is presented in card format, combining text and images for easy reading.

[0143] After a consumer clicks the "View Details" button, the webpage retrieves semi-public data from the server via an asynchronous request, including information such as process parameter ranges, quality inspection results, raw material utilization, and energy consumption. This information is displayed in tables and charts; for example, sulfide content is shown in a bar chart and compared with the target range, intuitively displaying the product quality level; process parameter ranges are displayed in interval format, showcasing the advancement of the process while protecting specific process details.

[0144] The website also offers a "Verify Data Authenticity" function. After a consumer clicks this button, the website retrieves the hash value and digital signature of the data from the server, calculates the hash value of the currently displayed data locally, and verifies the digital signature using the company's public key. If the verification passes, the website displays a message indicating successful data verification; if verification fails, the website displays a warning message indicating that the data may have been tampered with. This function allows consumers to independently verify the authenticity and integrity of the data, enhancing their trust in the product.

[0145] To prevent counterfeit products, the system records the number of queries and the query time for each QR code in the database. Under normal circumstances, each QR code is queried a limited number of times, preferably 1 to 5 times. If a QR code is queried a large number of times in a short period, such as exceeding a preset threshold within a preset time period, the system automatically alerts the system, indicating the potential presence of counterfeit products. The alert information is sent to the company's quality management department and relevant supervisory departments, enabling timely measures to combat counterfeit and substandard products.

[0146] This step yields a traceability data package and a QR code identifier. The traceability data package includes a hierarchical traceability dataset, a hash value, and a digital signature. The QR code identifier is used by consumers to scan and query.

[0147] A fully intelligent control system for the low-temperature extraction and separation of scallion oil is provided to execute the aforementioned intelligent control method for the low-temperature extraction and separation of scallion oil. Figure 3 As shown, it includes: The raw material characteristic prediction module is used to collect near-infrared spectral data of onion raw materials and perform preprocessing, predict raw material characteristic parameters and search for similar cases, and obtain recommended values ​​of process parameters and expected effect evaluation results. The extraction quality prediction module performs extraction based on recommended values ​​of process parameters and collects process parameters in real time. It predicts product quality indicators and compares them with the expected effect evaluation results to obtain product quality prediction results. The parameter optimization and adjustment module detects the risk of quality deviation based on the product quality prediction results, calculates the process parameter adjustment plan, performs safety constraint checks, and then executes the parameter adjustment to obtain the final extraction state information. The separation and purification control module performs separation and purification based on the final state information of the extraction and monitors it in real time. It adopts a three-stage separation strategy to remove impurities and obtain purified scallion oil product. The model update feedback module is used to perform quality testing on purified scallion oil products, generate quality testing reports, organize and update production data, and obtain an updated knowledge base. The traceability system construction module generates traceability information based on the quality inspection report and performs data classification processing and digital signature protection to obtain traceability data packages and QR code identifiers.

[0148] In one embodiment of the present invention, a specific example is provided: This embodiment underwent a 90-day field test at a medium-sized scallion oil production enterprise. During the test, a complete hierarchical intelligent control architecture was deployed, including 12 temperature sensors, 8 pressure sensors, 6 flow sensors, 4 level sensors, 1 portable near-infrared spectrometer, 2 edge computing devices, and 1 cloud data analysis platform. The test production line had a daily processing capacity of 500 kg of scallion raw materials, and a total of 30 batches of production were completed, yielding abundant experimental data.

[0149] Upon arrival of the scallion raw material batch 20241215-A-001, it was analyzed using a near-infrared spectrometer, and the spectral analysis model output basic raw material characteristic parameters. By performing five repeated measurements on the raw material, the freshness index and uniformity index were calculated. The raw material characteristic test data are shown in Table 1. Table 1: Test data of raw material characteristics for batch 20241215-A-001; Based on the aforementioned raw material characteristic parameter vector, eight most similar historical batches were retrieved from the historical database, with similarity scores of 0.92, 0.89, 0.87, 0.85, 0.83, 0.81, 0.79, and 0.77, respectively. Statistical analysis showed that these similar batches used extraction temperatures ranging from 54 to 58 degrees Celsius, extraction times ranging from 175 to 195 minutes, enzyme addition amounts ranging from 0.12% to 0.18% of the raw material mass, pH values ​​ranging from 6.3 to 6.7, and oil-to-material ratios ranging from 3:1 to 4:1.

[0150] Based on the weighted average of 8 similar historical batches, the recommended optimal process parameters are: extraction temperature 56 degrees Celsius, extraction time 185 minutes, enzyme addition of 0.15% of raw material mass, ultrasonic power density of 0.8 watts per cubic centimeter, pH value of 6.5, and oil-to-material ratio of 3.5 to 1.

[0151] Interpretability information for this batch of raw materials indicates that the freshness index is 87.5, which is relatively high; the uniformity index is 82.3, which is above average; and the initial sulfide content is 0.42%, which is within the normal range. Based on these characteristics, a moderate extraction temperature of 56 degrees Celsius and a moderate extraction time of 185 minutes are recommended.

[0152] The expected effect assessment shows that the sulfide content is expected to be 1.6 to 1.9 grams per kilogram, the polyphenol content is expected to be 400 to 450 milligrams per kilogram, the volatile flavor substance content is expected to be 2.5 to 3.2 grams per kilogram, and the color parameters are expected to be L value 44 to 47, a value -3 to -2, and b value 27 to 30.

[0153] The extraction process was started according to the recommended process parameters. At the 110-minute mark of extraction, approximately 60% of the total extraction time, the LSTM quality prediction model output the following prediction: the predicted sulfide content was 1.55 grams per kilogram, which was 1.6 grams lower than the target lower limit, with a deviation rate of 3.1%.

[0154] The MPC optimization algorithm was activated, and the following adjustment plan was calculated: the extraction temperature was increased from 56 degrees Celsius to 58 degrees Celsius, the extraction time was extended from 185 minutes to 195 minutes, and the amount of supplementary enzyme added was 0.03% of the raw material mass. After the adjustment was implemented, the predicted quality value was continuously monitored. At the 150-minute mark, the predicted sulfide content rose to 1.72 grams per kilogram, entering the target range.

[0155] After extraction, the mixture enters a three-stage separation process. Based on the extraction temperature of 58 degrees Celsius, the viscosity of the mixture is estimated, and the centrifuge speed is set to 3200 rpm. After separation, a clear and transparent purified onion oil product is obtained.

[0156] After the quality inspection is completed, the final product quality data is obtained, as shown in Table 2: Table 2: Final product quality inspection data for batch 20241215-A-001; All indicators are within the target range, the product is judged to be qualified, and the quality grade is rated as excellent.

[0157] The actual quality test results were compared with the expected results of step 1. The sulfide content of 1.75 g was within the expected range of 1.6 to 1.9 g, the polyphenol content of 435 mg was within the expected range of 400 to 450 mg, the volatile flavor substance content of 2.9 g was within the expected range of 2.5 to 3.2 g, and the color parameters were all within the expected range, which verified the accuracy of the raw material characteristic test and process parameter recommendation model.

[0158] Traceability information was generated, batch number 20241215-A-001, production date December 15, 2024, raw material origin Zhangqiu A District, Shandong, raw material variety Zhangqiu scallion, quality grade superior, shelf life 18 months. Semi-public data includes: extraction temperature range 55 to 58 degrees Celsius, extraction time range 185 to 195 minutes, pH value range 6.4 to 6.6, sulfide content 1.75 grams per kilogram, polyphenol content 435 milligrams per kilogram, raw material utilization rate 92%, and energy consumption 2.4 kilowatt-hours per kilogram of product.

[0159] This application example demonstrates the complete process of the present invention, from raw material testing, parameter recommendation, quality prediction, parameter optimization, separation and purification to quality traceability, and verifies the effectiveness and accuracy of the intelligent control method.

[0160] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments based on the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent control of the entire process of low-temperature extraction and separation of scallion oil, characterized in that, Includes the following steps: Near-infrared spectral data of scallion raw materials were collected and preprocessed. Raw material characteristic parameters were predicted and similar cases were searched to obtain recommended values ​​of process parameters and expected effect evaluation results. Extraction is performed based on recommended values ​​of process parameters, and process parameters are collected in real time. Product quality indicators are predicted and compared with the expected effect evaluation results to obtain product quality prediction results. Based on the product quality prediction results, the risk of quality deviation is detected, the process parameter adjustment plan is calculated, the parameter adjustment is executed after safety constraint check, and the final state information of extraction is obtained. Based on the final state information of the extraction, separation and purification were carried out and monitored in real time. A three-stage separation strategy was adopted to remove impurities and obtain purified scallion oil product. The purified scallion oil product undergoes quality testing, a quality testing report is generated, production data is organized and updated, and an updated knowledge base is obtained. Traceability information is generated based on the quality inspection report, and data is classified and protected by digital signature to obtain traceability data packages and QR code identifiers.

2. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 1, characterized in that, The similar case retrieval employs a case-based reasoning method, specifically including: Construct the feature vector of the current batch of cases. The feature vector includes moisture content, soluble solids content, initial sulfide content, variety code, origin code, freshness index and uniformity index. Normalize each feature in the case feature vector, mapping the numerical features to the 0 to 1 interval; The similarity between the current case and each historical case in the historical database is calculated using weighted Euclidean distance. The weight coefficients of the weighted Euclidean distance are determined according to the degree of influence of each feature on the extraction effect, with the initial sulfide content having the highest weight, followed by the variety code and freshness index. Sort the cases by similarity from high to low, and select the top K most similar historical cases. The K value is set to 5 to 20. The process parameters of the selected K historical cases were weighted and averaged, with the weight of each historical case being the value after similarity normalization, to obtain the recommended extraction temperature, extraction time, oil-to-material ratio and enzyme addition amount. It also outputs an assessment of expected results, including expected sulfide content, expected polyphenol content, and expected color parameters.

3. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 1, characterized in that, The predicted product quality indicators employ a long short-term memory network, specifically including: The input features for constructing a long short-term memory network include temporal features and static features; The time-series features include: temperature time-series data, pressure time-series data, pH time-series data, and cumulative enzyme addition time-series data, with a sampling frequency of 1 to 10 times per minute; Static characteristics include: raw material moisture content, raw material soluble solids content, initial sulfide content of raw material, raw material freshness index, raw material uniformity index, and recommended oil-to-seed ratio; A sliding time window is used to process the temporal features. The time window length is set to 10 to 60 time steps to form a three-dimensional tensor input. The tensor dimension is batch size × time step length × number of features. A three-dimensional tensor is input into a long short-term memory network. The network consists of an input layer, two LSTM hidden layers, and a fully connected output layer. The number of units in the first LSTM layer ranges from 64 to 128, and the number of units in the second LSTM layer ranges from 32 to 64. The Long Short-Term Memory Network outputs predicted values ​​for sulfide content, polyphenol content, volatile flavor compound content, and color parameters. The predicted values ​​are compared with the target quality indicators to calculate the quality deviation and deviation rate.

4. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 1, characterized in that, The calculation of the optimal process parameter adjustment scheme adopts a model predictive control algorithm, specifically including: When a quality deviation risk is detected, the model predictive control algorithm is activated. An optimization problem is constructed based on a pre-established prediction model of the extraction process. The prediction model adopts a linear state space representation. The state variables include temperature, pH value, extraction time, and cumulative enzyme addition. The control variables include temperature adjustment, time adjustment, and enzyme addition adjustment. The prediction time domain is set to 10 to 30 time steps, and the control time domain is set to 3 to 10 time steps; Construct an optimization objective function that includes a quality tracking error term and a control change penalty term, with the weight of the quality tracking error term being higher than that of the control change penalty term; The following hard constraints are set: temperature hard constraint is 40℃ to 80℃, and temperature change rate constraint is no more than 2℃ per minute; extraction time hard constraint is 30 minutes to 180 minutes; enzyme addition hard constraint is 0.1% to 2% of the raw material mass, and the single addition amount shall not exceed 30% of the total amount; pressure hard constraint is -0.05MPa to 0.1MPa; pH value hard constraint is 4.5 to 7.

5. Set soft constraints, including: energy consumption per batch does not exceed the preset energy consumption limit, and the total extraction time is preferably controlled within the preset time range; The optimization problem is solved using a quadratic programming method, the optimal control sequence is calculated, and the adjustment scheme of process parameters is output.

5. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 1, characterized in that, The purified onion oil product obtained specifically includes: The system receives the final state information of the extraction process and transmits the parameters and state information of the extraction process to the separation control module. The separation control module adjusts the separation process parameters according to the linkage control model. Estimate the viscosity of the mixture based on the final extraction temperature and adjust the centrifuge speed accordingly; determine whether to adjust the pH value of the separation stage based on the final extraction pH value. The first-stage coarse separation removes large solid particles through gravity sedimentation; when the sludge layer reaches a preset thickness threshold, the sludge is discharged; the upper clarified oil phase flows into the next stage centrifugal separation equipment. The second stage of centrifugal separation removes fine particles and some moisture through centrifugal force; The third-stage precision filtration removes minute impurities through multiple layers of filter membranes; when the differential pressure exceeds the preset differential pressure threshold, an alarm is triggered to prompt the replacement of the filter element; During the three-stage separation process, the oil phase quality is monitored according to the second preset acquisition frequency; when the monitored value exceeds the target value, the parameters are adjusted.

6. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 1, characterized in that, The updated knowledge base specifically includes: Quality assessment is conducted using a combination of online and offline testing; a quality inspection report is generated; and complete data for this batch is stored in a historical database. The actual quality test results are compared with the expected effect assessment, and the prediction deviation is calculated. When the accumulated number of historical batches reaches the third preset batch number, the model update mechanism is triggered; Model retraining employs an incremental learning approach; if the new model outperforms the old model, the new model is downloaded to the edge device to replace the old model. Identify the success factors and failure reasons for this batch; extract the identified experiences into rules and add them to the intelligent knowledge base.

7. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 1, characterized in that, The data hierarchical processing adopts a three-level hierarchical rule, specifically including: The traceability information in this batch is classified into three levels according to its level of openness: Level 1 Public Data, Level 2 Semi-Public Data, and Level 3 Confidential Data. Level 1 publicly available data includes: product batch number, production date, manufacturer name, manufacturer address, product specifications, shelf life, quality grade, and applicable standards; Secondary semi-public data includes: raw material origin, raw material variety, raw material freshness grade, extraction temperature curve, extraction time, oil-to-material ratio, separation process type, and sulfide content and polyphenol content in quality test results; Level 3 confidential data includes: detailed raw material characteristic parameters, complete process parameter time series data, enzyme addition amount, pH adjustment records, quality prediction model parameters, model prediction control algorithm parameters, cost data, supplier information, and energy consumption data; The primary public data and the secondary semi-public data are serialized, the digital fingerprint is calculated using the SHA-256 hash algorithm, and the hash value is digitally signed using the enterprise's private key. Level 3 confidential data is encrypted using the AES-256 symmetric encryption algorithm and stored encrypted on the enterprise's local server. Generate a QR code traceability identifier, which contains the product batch number and a traceability query link.

8. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 2, characterized in that, The prediction of raw material characteristic parameters adopts a neural network spectral analysis model, which includes an input layer, a hidden layer and an output layer. The hidden layer uses the ReLU activation function and a Dropout layer is set after the hidden layer.

9. The intelligent control method for the entire process of low-temperature extraction and separation of scallion oil according to claim 4, characterized in that, The extraction process prediction model uses a linear state-space representation. The state variables include temperature, pH value, extraction time, and cumulative enzyme addition. The control variables include temperature adjustment, time adjustment, and enzyme addition adjustment.

10. A fully intelligent control system for the low-temperature extraction and separation of scallion oil, characterized in that, A method for intelligent control of the entire process of low-temperature extraction and separation of scallion oil as described in any one of claims 1-9 includes: The raw material characteristic prediction module is used to collect near-infrared spectral data of scallion raw materials and perform preprocessing, predict raw material characteristic parameters and search for similar cases, and obtain recommended values ​​of process parameters and expected effect evaluation results. The extraction quality prediction module performs extraction based on recommended values ​​of process parameters and collects process parameters in real time. It predicts product quality indicators and compares them with the expected effect evaluation results to obtain product quality prediction results. The parameter optimization and adjustment module detects the risk of quality deviation based on the product quality prediction results, calculates the process parameter adjustment plan, performs safety constraint checks, and then executes the parameter adjustment to obtain the final extraction state information. The separation and purification control module performs separation and purification based on the final state information of the extraction and monitors it in real time. It adopts a three-stage separation strategy to remove impurities and obtain purified scallion oil product. The model update feedback module is used to perform quality testing on purified scallion oil products, generate quality testing reports, organize and update production data, and obtain an updated knowledge base. The traceability system construction module generates traceability information based on the quality inspection report and performs data classification processing and digital signature protection to obtain traceability data packages and QR code identifiers.

Images