Intelligent perception-based quality and quantity wine picking method and system

Abstract

The application discloses a quality and quantity liquor picking method and system based on intelligent sensing, which is applied to liquor distillation production line distillate grading liquor picking. The method pre-constructs timing-aligned base liquor calibration dataset, pre-trains spectrum-alcohol content-sensory grade two-way closed loop coupling model; through the pre-completed hardware level timing calibration integrated detection device, the multi-dimensional data of the distillate liquor is synchronously collected, and after temperature correction, the model is input, through the three-layer logic of forward hard constraint, dynamic soft weight multi-classification SVM fine classification judgment and reverse bias correction, the alcohol content prediction and grade determination are completed, and finally the automatic grading storage is realized through the intelligent shunting unit with anti-shake control. The method realizes the standardization and intelligentization of liquor picking, and improves the base liquor grading accuracy and production stability.

Description

Technical Field

[0001] This invention relates to the field of technology, specifically to a method and system for measuring and extracting wine based on intelligent sensing. Background Technology

[0002] Quality-based distillation is the core process in baijiu distillation, directly determining the quality grade of the base liquor, the yield of premium liquor, and the stability of batch production. Currently, the mainstream manual distillation method in the baijiu industry heavily relies on the individual experience of distillers and tasters, resulting in inherent flaws such as highly subjective grading standards, long training cycles for professionals, significant differences in grading results among different operators, substantial batch-to-batch quality fluctuations in base liquor, and the difficulty in scaling up traditional techniques. Existing intelligent distillation technologies generally suffer from problems such as inaccurate timing of multi-source detection data, insufficient robustness of models due to temperature fluctuations, high misjudgment rates due to the non-linear correlation between alcohol content and sensory grades, and the potential for mixed grades caused by frequent valve switching in industrial settings. These issues fail to meet the precise and stable grading requirements of continuous industrial baijiu production. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for measuring and extracting wine based on intelligent sensing, which is capable of...

[0004] In a first aspect, this application provides a method for graded distillation based on intelligent sensing, applied to the graded distillation of distillate in a baijiu distillation production line, characterized by the following steps:

[0005] S1: A pre-constructed time-aligned base wine calibration dataset, which includes one-to-one corresponding near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data anchored based on national standards. The sensory grades are divided into six levels: head, premium, first-grade, second-grade, third-grade, and tail, with each level corresponding to a preset alcohol content threshold range.

[0006] S2: Based on the calibration dataset, a two-way closed-loop coupling model of spectrum-alcohol content-sensory level is pre-trained and constructed. The two-way closed-loop coupling model integrates the quantitative alcohol content correction sub-model and the qualitative sensory level discrimination sub-model.

[0007] S3: Pre-completion of hardware-level timing calibration of the integrated detection device: Calculate the flow path delay time Δt between the spectral acquisition flow cell and the alcohol content detection module, use the same clock source to trigger spectral acquisition and alcohol content detection, and after the spectral acquisition is completed, delay Δt to latch the corresponding alcohol content and temperature data to achieve acquisition timing synchronization; during the distillation process, through the calibrated integrated detection device, the near-infrared spectral data, alcohol content data and temperature data of the distilled liquor are acquired in real time and synchronously.

[0008] S4: Based on real-time temperature data, perform temperature-adaptive baseline correction on the spectral data. Input the corrected spectral data into the pre-trained bidirectional closed-loop coupled model, and synchronously complete alcohol content prediction and grade determination according to the following hierarchical logic:

[0009] Ⅰ Forward hard constraint: The alcohol content quantitative correction sub-model outputs the real-time alcohol content prediction value of the distilled liquid. Based on the alcohol content threshold range of each level preset by S1, it defines the range of candidate sensory levels that meet the threshold requirements and excludes non-candidate levels outside the threshold range.

[0010] II. Sub-classification: Within the candidate sensory level range, the sensory level qualitative discrimination sub-model adopts a multi-class SVM algorithm with dynamic soft weights, using the corrected spectral data and real-time alcohol content prediction as joint inputs to complete the final sensory level sub-classification.

[0011] III. Reverse Bias Correction: The final sensory rating result is used as a bias correction factor to systematically correct the predicted alcohol content within the range of that rating, and the final alcohol content result is output.

[0012] S5: Based on the final sensory rating result, the intelligent diversion unit with anti-shake control logic completes the automatic grading and storage of wine of the corresponding grade.

[0013] In some embodiments, step S1 specifically includes the following sub-steps:

[0014] S11: Based on the national standard for the target aroma type of baijiu, and using the corresponding national standard baijiu sample as the anchor benchmark, establish the classification rules for 6 sensory levels, the alcohol content threshold range for each level, and standardized sensory evaluation indicators.

[0015] S12: Collect base liquor calibration samples at the receiving port of the baijiu distillation production line, divided into three stages: head, middle and tail of liquor, covering different production shifts, batches and all working conditions of the cellars. The total number of samples shall not be less than 200, and the number of single-grade samples shall not be less than 15.

[0016] S13: Complete synchronous collection for each calibration sample, obtain the near-infrared spectral data, measured alcohol content data and synchronous temperature data of the sample, and complete the sensory level calibration of the sample according to the standardized sensory evaluation indicators.

[0017] S14: The initial data alignment is completed by hardware-level time-series calibration with the pre-completed integrated detection device. Then, the cross-correlation algorithm is used to perform secondary time-series alignment of the spectral sequence and the alcohol content sequence, and finally a one-to-one corresponding calibration dataset is formed.

[0018] In some embodiments, the specific steps of hardware-level timing calibration in step S3 are as follows:

[0019] S31: Calculate the flow path volume L between the spectral acquisition flow cell and the alcohol content detection module, and determine the theoretical delay time t0 = L / v by combining the design flow rate v of the distillate;

[0020] S32: Perform actual delay calibration using the step calibration method: Inject a standard concentration ethanol solution into the flow path, and record the time t1 when the spectral acquisition module detects the concentration change and the time t2 when the alcohol content detection module detects the concentration change, respectively. Calculate the actual flow path delay time Δt = t2 - t1.

[0021] S33: Write Δt into the control system, use the same 100Hz clock source to synchronously trigger the spectral acquisition module and the alcohol content detection module, and after the spectral acquisition is completed, delay Δt time to latch the corresponding alcohol content and temperature detection values ​​to achieve hardware-level timing synchronization with a synchronization error ≤0.1s.

[0022] In some embodiments, the specific steps of temperature adaptive baseline correction in step S4 are as follows:

[0023] S41: Pre-collect near-infrared spectra of real base wine standard samples with different grades and fractions of concentration gradients of 5%-80%vol within the temperature range of 20℃-60℃, and establish a quantitative correction model for temperature and spectral baseline offset.

[0024] S42: For the raw near-infrared spectral data acquired in real time, baseline drift correction is completed through a quantitative correction model using synchronously acquired temperature data, eliminating the interference of temperature fluctuations on the spectrum;

[0025] S43: Perform preprocessing on the corrected spectral data, sequentially performing outlier removal, Savitzky-Golay convolutional smoothing and denoising, adaptive baseline correction, and wavelength normalization to obtain standard spectral feature data.

[0026] In some embodiments, step S2 specifically includes the following sub-steps:

[0027] S21: Perform temperature adaptive correction and preprocessing on the spectral data of the calibration dataset to obtain a standard spectral feature dataset;

[0028] S22: Using the Kennard-Stone algorithm, the standard spectral feature dataset, the corresponding measured alcohol content data, and the sensory grade data are divided into a modeling set and a validation set in a 9:1 ratio.

[0029] S23: Construct a quantitative alcohol content calibration sub-model, using partial least squares regression (PLS) algorithm, with standard spectral characteristic data as input and measured alcohol content as output, and determine the optimal number of principal components to be 5-8 through cross-validation, with a cumulative contribution rate ≥95%;

[0030] S24: Construct a qualitative sensory grade discrimination sub-model, using a multi-class SVM algorithm with dynamic soft weights, with standard spectral feature data and the synchronous output of the alcohol content quantitative correction sub-model as joint inputs and the sensory grade of the base wine as the output;

[0031] S25: Set up a two-way closed-loop coupling rule for model training and inference, which corresponds completely to the hierarchical logic of step S4: During the training phase, the sample classification range is defined by the measured alcohol content as a hard constraint, and then the level training is completed by SVM with soft weights. Finally, the training bias of the alcohol content sub-model is corrected in reverse by the level calibration result; During the inference phase, the logic of steps I-III of step S4 is executed.

[0032] S26: The model is trained using a modeling set, the model parameters are optimized using leave-one-out cross-validation, and the model's generalization ability is verified using a validation set to ensure that the model's accuracy meets the standards.

[0033] In some embodiments, the specific rules for the dynamic soft weights in step S4Ⅱ are as follows:

[0034] S61: The dynamic soft weighting only applies within the candidate sensory level range defined in step S4Ⅰ. The weighting transitions smoothly with the rate of change of alcohol content through the Sigmoid function, without any hard threshold abrupt changes.

[0035] S62: The weight transition rule is as follows:

[0036] When the rate of change of alcohol content v < 1.5%vol / min, the weight of alcohol content is 0.5, and the weight of spectral characteristics is 0.5.

[0037] When the rate of change of alcohol content is 1.5%vol / min≤v≤2.5%vol / min, the weights transition smoothly through the Sigmoid function, with the alcohol content weight linearly increasing from 0.5 to 0.6 and the spectral feature weight linearly decreasing from 0.5 to 0.4.

[0038] When the rate of change of alcohol content v > 2.5% vol / min, the weight of alcohol content is 0.6 and the weight of spectral characteristics is 0.4.

[0039] S63: The rate of change of alcohol content is calculated using the slope of a linear fit with a 10-second sliding window. The input alcohol content data is first processed by a 5-point sliding window Kalman filter to eliminate detection noise.

[0040] S64: The multi-class SVM algorithm uses the RBF kernel function and optimizes the penalty coefficient C and kernel function parameter gamma through grid search combined with 5-fold cross-validation. The optimization range of the penalty coefficient C is 1~10.

[0041] In some embodiments, step S5 specifically includes the following sub-steps:

[0042] S51: The intelligent control system receives the final sensory level result output by the model and initiates the anti-shake verification logic;

[0043] S52: The valve switching action of the corresponding level is triggered only when the level judgment results are completely consistent for 3 consecutive times with an interval of 0.5s. The valve operation is not executed when the judgment result changes once.

[0044] S53: Through the explosion-proof intelligent valve group, the wine of the corresponding grade is diverted to the storage tank of the corresponding grade to complete the automatic graded storage.

[0045] In some embodiments, in step S26, the model accuracy must simultaneously meet the following indicators: alcohol content derivation error ≤ ±0.5%vol, sensory grade judgment accuracy ≥95%, abnormal wine identification accuracy ≥98%, validation set prediction deviation ≤ ±2%, and model accuracy fluctuation within the range of wine sample temperature fluctuation ±3℃ ≤2%.

[0046] In some embodiments, the S7 full-process data closure and model adaptation steps are also included, specifically in the following steps:

[0047] S71: Record the entire production process data of each batch of liquor, build a unique digital passport for each batch of base liquor, with 100% data integrity and storage time ≥ 5 years;

[0048] S72: Based on production operation data and standard wine sample verification results, complete the parameter fine-tuning and optimization of the coupled model every quarter;

[0049] S73: To meet the adaptation needs across wineries and aroma types, a model adaptation method based on transfer learning is adopted to map the feature space of the pre-trained model to the feature space of the target scene. The target scene provides ≤50 sets of standard wine samples to complete the rapid model adaptation.

[0050] Secondly, this application provides a quality-based liquor extraction system based on intelligent perception, used to implement the above-mentioned quality-based liquor extraction method, applied to a liquor distillation production line, including a grading benchmark storage module, a calibration dataset management module, a model calculation module, an integrated online detection unit, an intelligent control unit, and a diversion execution unit;

[0051] The grading benchmark storage module is used to store the sensory grade classification rules of six base liquors anchored to national standards, the alcohol content threshold range corresponding to each grade, and standardized sensory evaluation indicators.

[0052] The calibration dataset management module is used to store pre-built time-aligned base wine calibration datasets, which contain one-to-one corresponding near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data.

[0053] The integrated online detection unit has completed hardware-level timing calibration in advance and has built-in near-infrared spectral acquisition module, high-precision alcohol content detection module and temperature acquisition module triggered by the same clock source, which are used to synchronously acquire near-infrared spectral data, alcohol content data and temperature data of distilled liquor in real time.

[0054] The model computation module incorporates a pre-trained bidirectional closed-loop coupled model of spectral-alcohol content-sensory level. This model integrates a quantitative alcohol content correction sub-model and a qualitative sensory level discrimination sub-model. The model computation module is used to perform temperature-adaptive baseline correction of the spectral data and completes inference operations according to the following hierarchical logic: Ⅰ Forward hard constraint: The real-time alcohol content prediction value is output through the quantitative alcohol content correction sub-model to define the candidate sensory level range; Ⅱ Fine classification determination: Within the candidate range, the sensory level is finely classified using a multi-class SVM algorithm with dynamic soft weights; Ⅲ Backward bias correction: The systematic bias of the alcohol content prediction value is corrected using the final level result as a correction factor.

[0055] The intelligent control unit has built-in anti-shake control logic, which is used to drive the diversion execution unit to complete the automatic grading and storage of wine of the corresponding grade according to the final sensory level result output by the model.

[0056] The shunt execution unit is electrically connected to the intelligent control unit.

[0057] The beneficial effects of this invention are as follows: This method achieves accurate synchronous acquisition of multi-source data (spectral density, alcohol content, and temperature) through hardware-level time-series calibration, eliminating the data timing inaccuracy problem caused by flow path delay and improving data consistency during model training and inference. Through the three-level inference logic of the bidirectional closed-loop coupled model, the problem of spectral ambiguity caused by the nonlinear correspondence between alcohol content and sensory grade is solved, significantly reducing the misjudgment rate of grade determination. The dynamic soft-weight algorithm adapts to the alcohol content change characteristics at different stages of distillation, improving the determination accuracy at grade boundaries. The anti-jitter control logic avoids the risks of frequent valve switching and mixed-grade liquor, achieving accurate graded storage of base liquor. This method transforms traditional brewing experience into a standardized intelligent process, realizing the automation and intelligence of the distillation process, effectively improving the batch stability and production controllability of base liquor. Attached Figure Description

[0058] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0060] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0061] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0062] In the description of the embodiments of the present invention, it should be noted that if terms such as "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first," "second," and "third" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0063] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0064] Furthermore, the use of terms such as "horizontal," "vertical," and "sag" does not imply that the component must be absolutely horizontal or suspended, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal relative to "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0065] Currently, the traditional method of distillation in the baijiu industry relies on "observing the foam" and "judging by taste," depending entirely on the personal experience of master distillers. This involves observing the shape, size, and duration of the foam, combined with taste and smell to determine the alcohol content and flavor level of the liquor, dividing it into heads, middle, and tail sections, and establishing the start and end points and segmentation standards for distillation. Existing intelligent distillation technologies largely rely on laboratory quantitative analysis of seven components, including total acid, total esters, and acetic acid, and then comparing this data with spectral data to create a model, attempting to define the quality grade of the liquor through limited component indicators.

[0066] The existing technology has the following problems:

[0067] (i) Skills inheritance: Traditional skills rely on personal experience, the judgment criteria are subjective and cannot be quantified, making them difficult to replicate and pass on; senior brewing masters are scarce and aging, and young practitioners have difficulty mastering core skills, highlighting the risk of skills being lost.

[0068] (ii) Talent cultivation: The training cycle for excellent wine tasters is long (from several years to several decades), lacks systematic quantitative standards, and relies entirely on personal understanding and mentorship. The industry has an insufficient supply of wine tasters, which restricts the stability of wine production quality and the implementation of intelligent technology.

[0069] (III) Quality fluctuation: Human judgment is affected by individual differences, environment and physical condition, resulting in large deviations (error rate of 15%-20%). The existing intelligent wine-making process only relies on seven component indicators for grading, ignoring the complexity of the components of sauce-flavored wine and strong-aroma wine, making it difficult to match consumers' taste needs, resulting in inaccurate wine-making segmentation, uneven batch wine quality, and low yield of high-quality wine.

[0070] (iv) Efficiency bottleneck: Existing intelligent winemaking requires laboratory component testing, has a long modeling cycle (usually 3-6 months), and has extremely low implementation efficiency.

[0071] (V) Data Black Box Level: Traditional manual distillation decisions lack quantitative records, while existing intelligent distillation relies on complex component detection, making data acquisition cumbersome and disconnected from taste, both of which make it difficult to construct a quality profile that meets market demands. In view of this, this application provides a quantitative and qualitative distillation method and system based on intelligent perception, which is applied to the graded distillation of distillate in a baijiu distillation production line.

[0072] Example 1

[0073] In view of this, refer to Figure 1 The first aspect of this application provides a method for graded distillation based on intelligent sensing, applied to the graded distillation of distillate in a baijiu distillation production line, comprising:

[0074] S1: A pre-built time-aligned base wine calibration dataset. The dataset contains one-to-one corresponding near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data anchored based on national standards. The sensory grades are divided into six levels: head, premium, first-grade, second-grade, third-grade, and tail. Each level corresponds to a preset alcohol content threshold range.

[0075] The specific execution process of pre-building the time-aligned base wine calibration dataset is as follows:

[0076] S11: Based on the national standard for the target aroma type of baijiu, and using the corresponding national standard baijiu sample as the anchor benchmark, organize process experts, quality experts, and at least 3 senior tasters to jointly establish the classification rules for 6 sensory levels, the corresponding alcohol content threshold range for each level, and standardized sensory evaluation indicators. Senior tasters participating in the development of indicators must meet the requirements of having ≥10 years of professional experience and possessing national-level baijiu tasting qualifications. After the indicators are developed, all participating tasters must sign and confirm them. The standardized sensory evaluation indicators are based on GB / T33404-2016 "Guidelines for Sensory Evaluation of Baijiu". The standardized sensory evaluation indicators include four quantitative dimensions: aroma intensity, taste purity, aftertaste length, and style fit. Each dimension has a fixed scoring standard of 0 to 10 points. The scoring standard of each dimension corresponds to a clear and reproducible sensory description. The scoring weights of the four dimensions are equal. The total score of the four dimensions is used to classify six sensory levels. The six sensory levels are: head, premium, first-grade, second-grade, third-grade, and tail. The alcohol content threshold range corresponding to each level is adapted to the internal control standard of the base liquor quality of the corresponding distillery. The upper and lower limits of the alcohol content threshold range do not exceed ±0.5% vol.

[0077] S12: At the receiving point of the baijiu distillation production line, collect base liquor calibration samples in three stages: head, middle, and tail, covering different production teams, batches, and fermentation pits under all operating conditions. The total number of samples should not be less than 200, and the number of samples for each grade should not be less than 15. The specific execution rules for the sampling process are as follows: collect 2 head samples, 5 middle samples, and 3 tail samples from each receiving point. The sampling should cover all receiving points in the same production workshop and cover all operating conditions of different production teams, different production batches, and fermentation pits of different ages. This ensures that the collected calibration samples cover all possible changes in liquor quality during the distillation process and avoids insufficient model generalization ability due to uneven sample distribution. After collection, the wine samples were sealed in brown ground glass bottles with polytetrafluoroethylene (PTFE) tape to prevent evaporation and changes in composition. The storage environment was maintained at a temperature of 20°C to 25°C and a relative humidity of 40% to 60%. The storage time was no more than 48 hours. Near-infrared spectral data acquisition, measured alcohol content data detection, and synchronous temperature data acquisition were completed within 4 hours after sample collection to avoid inaccurate calibration data caused by changes in wine sample composition over time.

[0078] S13: Simultaneous data acquisition is performed on each calibration sample to obtain near-infrared spectral data, measured alcohol content data, and synchronous temperature data. Sensory level calibration of the samples is then completed according to standardized sensory evaluation indicators. The synchronous data acquisition process for calibration samples utilizes an integrated detection device consistent with subsequent production processes, employing the same acquisition parameters and hardware-level time-series calibration parameters. This ensures consistency between the acquired data of the calibration samples and the real-time acquisition data of the subsequent production process, eliminating data distribution deviations in model training and inference caused by differences in acquisition equipment and parameters. The sensory level calibration process employs a blind tasting method with accompanying standard samples, conducted by at least three senior wine tasters with ≥10 years of experience and national-level wine tasting qualifications. During the blind tasting, all source information, production information, and alcohol content information of the wine samples are masked to avoid interference from irrelevant information for the calibration personnel. For every 10 wine samples to be tested, one national-level standard wine sample of the corresponding aroma type is inserted to calibrate the scoring benchmark of the calibration personnel, ensuring consistency in scoring results among different calibration personnel. The entire blind tasting process is recorded. The grading result of a single wine sample must meet the requirement that the grading consistency rate of at least 3 sommeliers is ≥85% in order to be considered as the final valid grading result. For wine samples that do not meet the consistency rate, a fourth senior sommelier with equivalent qualifications will review and decide, and the review result will be used as the final grading result to ensure the reproducibility and accuracy of the grading results.

[0079] S14: Initial data alignment is achieved through hardware-level timing calibration consistent with subsequent production stages. Then, a cross-correlation algorithm is used for secondary timing alignment of the spectral and alcohol content sequences, ultimately forming a one-to-one corresponding calibration dataset. The specific execution process of the cross-correlation algorithm is as follows: the acquired near-infrared spectral absorbance time series is defined as x(n), and the synchronously acquired alcohol content time series is defined as y(n). The sampling frequencies of the two sequences are consistent, and the number of sampling points is N. The cross-correlation function R of the two sequences is calculated. x The formula for calculating the cross-correlation function ᵧ(τ) is:

[0080] R x ᵧ(τ)=Σx(n)y(n+τ)

[0081] Where τ is the time offset, and the summation range is the effective value interval of n. The time offset τ corresponding to the peak value of the cross-correlation function is calculated. max , τ max This is the optimal alignment offset between the two sequences, based on the optimal alignment offset τ. max The spectral and alcohol content sequences are shifted and aligned to eliminate minor temporal deviations during acquisition, ultimately forming a calibration dataset with time-aligned, one-to-one correspondence of near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data.

[0082] S2: Based on the calibration dataset, a two-way closed-loop coupled model of spectral density, alcohol content, and sensory level is pre-trained and constructed. This model integrates a quantitative alcohol content correction sub-model and a qualitative sensory level discrimination sub-model. The specific pre-training process for constructing the two-way closed-loop coupled model of spectral density, alcohol content, and sensory level is as follows:

[0083] S21: Perform temperature adaptive correction and preprocessing on the spectral data of the calibration dataset to obtain a standard spectral feature dataset. The temperature adaptive correction process is consistent with the temperature adaptive baseline correction process in the subsequent production process. The preprocessing process sequentially performs outlier removal, Savitzky-Golay convolutional smoothing and denoising, multivariate scattering correction, and standard normal variable transformation to obtain the standard spectral feature dataset.

[0084] Outlier removal was performed using Mahalanobis distance combined with the Grubbs test. The Mahalanobis distance was calculated as follows: the spectral data matrix of the calibration dataset was defined as X, an n x m matrix, where n is the number of calibration samples and m is the number of wavelength points in the spectral data. The mean vector μ of the spectral data matrix was calculated, and the covariance matrix S of the spectral data matrix was calculated. The formula for calculating the covariance matrix S is as follows:

[0085] S=(X-μ)ᵀ(X-μ) / (n-1)

[0086] Calculate the Mahalanobis distance MDᵢ for each sample. The formula for calculating the Mahalanobis distance is:

[0087] MDᵢ=√[(Xᵢ-μ)S⁻¹(Xᵢ-μ)ᵀ]

[0088] Where Xᵢ is the row vector of the spectral data of the i-th sample. The confidence interval for the Mahalanobis distance is set to 95%, and outlier samples outside the confidence interval are removed. The significance level of the Grubbs test is set to 0.05, and a second outlier test is performed on the samples filtered by Mahalanobis distance to remove samples with outlier results, ensuring that the spectral data used for model training is free from outlier interference.

[0089] The Savitzky-Golay convolutional smoothing denoising uses a 7-point smoothing window and a 3rd-order polynomial. The convolution kernel coefficients in the smoothing process are obtained by least squares fitting. The objective function of the fitting process is to minimize the sum of squared residuals. The spectral data is smoothed through convolution operations to eliminate random noise in the spectral data while retaining the effective characteristic peaks and peak shape information of the spectral data.

[0090] The execution process of multivariate scattering correction is as follows: calculate the average spectrum of all spectral data in the calibration dataset, use the average spectrum as the reference spectrum, perform univariate linear regression on the spectral data of each sample and the reference spectrum to obtain the regression coefficient and intercept of each sample, and correct the spectral data of each sample according to the regression coefficient and intercept to eliminate the spectral baseline shift and slope change caused by the turbidity of the wine sample and the scattering of suspended particles, and obtain the corrected spectral matrix.

[0091] The standard normal variable transformation process is as follows: for the spectral data of each sample, calculate the mean and standard deviation of the absorbance of the sample across all wavelengths, subtract the mean from the absorbance value of each wavelength point of the sample and divide by the standard deviation to obtain the transformed spectral data, thus eliminating the spectral baseline shift caused by differences in optical path length and sample packing.

[0092] S22: The Kennard-Stone algorithm is used to divide the standard spectral feature dataset, corresponding measured alcohol content data, and sensory grade data into a modeling set and a validation set at a 9:1 ratio. The Kennard-Stone algorithm executes by first calculating the Euclidean distance between all samples. The formula for calculating the Euclidean distance is:

[0093] dᵢⱼ=√[Σ(Xᵢ k -Xⱼ k )²]

[0094] Among them, Xᵢ k Let Xⱼ be the k-th feature value of the i-th sample. kLet be the k-th feature value of the j-th sample, and the summation range covers all feature dimensions. Select the two samples with the largest Euclidean distance as initial samples and add them to the modeling set. Then, calculate the minimum Euclidean distance between each remaining sample and all samples in the modeling set, and select the sample with the largest minimum Euclidean distance to add to the modeling set. Repeat the above sample selection process until the number of samples in the modeling set reaches 90% of the total number of samples. The remaining 10% of samples are used as the validation set to ensure that the samples in the modeling set and validation set cover all levels and all working conditions, and that the sample distribution is uniform, avoiding insufficient model generalization ability caused by the difference in sample distribution between the modeling set and the validation set.

[0095] S23: Construct a quantitative alcohol content correction sub-model, employing the Partial Least Squares Regression (PLS) algorithm. Using standard spectral characteristic data as input and measured alcohol content as output, the optimal number of principal components is determined to be 5-8 through cross-validation, with a cumulative contribution rate ≥95%. The specific derivation and execution steps of the PLS algorithm are as follows:

[0096] The first step is data standardization. The standard spectral feature data matrix of the modeling set is defined as X, where X is an n x m matrix, n is the number of samples in the modeling set, and m is the dimension of the spectral features. The corresponding measured alcohol content data matrix of the modeling set is defined as Y, where Y is an n x 1 matrix. The X and Y matrices are then standardized to obtain the standardized matrices E0 and F0. The standardization calculation formula is as follows:

[0097] E0=(X-μ x ) / σ x

[0098] F0=(Y-μᵧ) / σᵧ

[0099] Where μ x Let σ be the mean vector of each column of matrix X. x Let X be the standard deviation vector of each column of matrix X, μᵧ be the mean of matrix Y, and σᵧ be the standard deviation of matrix Y.

[0100] The second step is principal component extraction (PCE). The maximum number of iterations for PCE is set to 300, and the initial iteration count h=1. E0 is used as the spectral data matrix for the current iteration, and F0 is used as the alcohol content data matrix for the current iteration. The weight vector w is then calculated. h The formula for calculating the weight vector is:

[0101] w h =E₍ h ₋1₎ᵀF₍ h ₋1₎ / ||E₍ h ₋1₎ᵀF₍ h ₋1₎||

[0102] Where ||·|| is the 2-norm of the vector, for the weight vector w h After normalization, we obtain the unit weight vector w. h * The formula for normalization is:

[0103] w h *=w h / ||w h ||

[0104] Calculate the principal component score vector t of the spectral data h The formula for calculating the principal component score vector is:

[0105] t h =E₍ h ₋1₎w h *

[0106] Calculate the load vector p of the spectral data h The formula for calculating the load vector is:

[0107] p h =E₍ h ₋1₎ᵀt h / t h ᵀt h

[0108] Calculate the loading vector q of alcohol content data h The formula for calculating the load vector is:

[0109] q h =F₍ h ₋1₎ᵀt h / t h ᵀt h

[0110] For matrix E₍ h ₋1₎ and F₍ h ₋1₎ Perform residual update. The formula for residual update is:

[0111] E h =E₍ h ₋1₎-t h p h ᵀ

[0112] F h =F₍ h ₋1₎-t h q h ᵀ

[0113] Increase the iteration number h by 1 and repeat the principal component extraction steps above until the preset number of principal components is reached, or the sum of squares of the residual matrix is ​​lower than the preset threshold.

[0114] The third step is to determine the optimal number of principal components. Leave-one-out cross-validation is used to determine the optimal number of principal components, with the selection range being 1 to 15. The process of leave-one-out cross-validation is as follows: One sample is removed from the modeling set each time, the remaining samples are used to build the PLS model, the removed samples are used for validation, and the predicted residuals of the model are calculated. This process is repeated until all samples have been removed and validated once. The root mean square error of cross-validation (RMSECV) is then calculated for different numbers of principal components. The formula for RMSECV is:

[0115] RMSECV=√[Σ(yᵢ-yᵢ^)² / n]

[0116] Where yᵢ is the measured alcohol content of the sample, yᵢ^ is the predicted value of the model, and n is the number of samples in the modeling set. The number of principal components corresponding to the minimum value of RMSECV is selected as the optimal number of principal components. The optimal number of principal components ranges from 5 to 8, and the cumulative contribution rate of the corresponding principal components is ≥95%.

[0117] The fourth step is to calculate the regression coefficients. Based on the number of extracted optimal principal components, the regression coefficient matrix B of the PLS model is calculated. The formula for calculating the regression coefficient matrix is ​​as follows:

[0118] B=W*(PᵀW)⁻¹Q*ᵀ

[0119] Where W is the matrix composed of unit weight vectors, P is the matrix composed of spectral loading vectors, and Q* is the matrix composed of alcohol content loading vectors. A quantitative prediction model from spectral feature data to alcohol content data is constructed based on the regression coefficient matrix. The prediction formula of the model is:

[0120] Y^=XB+Y0

[0121] Where Y^ is the predicted alcohol content matrix, X is the standardized spectral feature data matrix, and Y0 is the mean of the alcohol content data.

[0122] The fifth step is model validation. The standard spectral feature data of the validation set is input into the trained PLS model to obtain the alcohol content prediction value. The root mean square error (RMSEP) and the coefficient of determination (R²) of the model are calculated to ensure that the alcohol content derivation error of the model is ≤ ±0.5% vol and the cross-validation coefficient of determination (R²) is ≥ 0.99.

[0123] In some embodiments, the alcohol content quantification correction submodel is constructed using the Extreme Learning Machine (ELM) algorithm. The number of nodes in the input layer of the ELM is consistent with the dimension of the spectral features, the number of nodes in the hidden layer is set to 20, and the activation function of the hidden layer is the Sigmoid function. The training process of the model is to solve the output weights by solving the generalized inverse matrix, without the need for iterative optimization. The training time of the model is shortened by more than 40% compared with the partial least squares regression algorithm, making it suitable for application scenarios that require rapid model construction and updates.

[0124] S24: Construct a qualitative sensory level discrimination sub-model, employing a multi-class SVM algorithm with dynamic soft weights. The standard spectral feature data and the synchronous output of the alcohol content quantitative correction sub-model are used as joint inputs, with the sensory level of the base wine as the output. The multi-class SVM algorithm uses a one-to-one multi-class strategy, decomposing the 6 sensory level multi-class task into 15 binary sub-tasks. Each binary sub-task corresponds to the construction of classification boundaries for two different sensory levels. The specific derivation and execution steps of the multi-class SVM algorithm are as follows:

[0125] The first step is input data preprocessing. The standard spectral feature data matrix of the modeling set is defined as X, and the alcohol content data matrix output by the alcohol content quantification correction sub-model is defined as Y. al c ohol Construct the joint input matrix Z. The formula for calculating the joint input matrix is:

[0126] Z=[α·X,β·Y al c ohol ]

[0127] Where α is the spectral feature weight, β is the alcohol content weight, α+β=1, and α and β are dynamic soft weights. The weight values ​​are smoothly adjusted by the Sigmoid function as the alcohol content changes.

[0128] The second step involves dynamic soft weight calculation. The rate of change of alcohol content, v, is calculated using the slope of a linear fit within a 10-second sliding window. The alcohol content data sequence within the sliding window is y1, y2, ..., y k k is the number of sampling points within the sliding window, the sampling interval is 1 second, and the linear fitting formula is y = a·t + b, where t is the time series, a is the slope obtained from the fitting, i.e., the rate of change of alcohol content v = a, and the unit of v is %vol / min. The unit and calculation method are completely consistent with the subsequent criteria for determining the rate of change of alcohol content. The input alcohol content data first undergoes a 5-point sliding window Kalman filter to eliminate random noise in the alcohol content detection data. The specific execution steps of the Kalman filter are as follows:

[0129] 1. State prediction: x(k|k-1)=x(k-1|k-1), where x(k|k-1) is the prior state estimate at time k, and x(k-1|k-1) is the posterior state estimate at time k-1;

[0130] 2. Covariance prediction: P(k|k-1)=P(k-1|k-1)+Q, where P(k|k-1) is the prior estimated covariance, P(k-1|k-1) is the posterior estimated covariance, and Q is the process noise covariance, with Q set to 0.01;

[0131] 3. Kalman gain calculation: K(k)=P(k|k-1) / (P(k|k-1)+R), where K(k) is the Kalman gain, R is the observation noise covariance, and R is set to 0.09;

[0132] 4. State update: x(k|k)=x(k|k-1)+K(k)·(z(k)-x(k|k-1)), where x(k|k) is the posterior state estimate at time k, and z(k) is the alcohol content observation at time k;

[0133] 5. Covariance update: P(k|k)=(1-K(k))·P(k|k-1), where P(k|k) is the posterior estimated covariance at time k.

[0134] The Sigmoid transition function for dynamic soft weights is:

[0135] β(v)=0.5+0.1·(1 / (1+exp(-k·(v-2))))

[0136] Where k is the slope coefficient, the value of which is determined through parameter sensitivity analysis and grid search optimization of the modeling set. The optimization process is as follows: within the range of k=110, different k values ​​are set with a step size of 0.5, and multi-class SVM models are trained respectively. Five-fold cross-validation is used to calculate the accuracy of the model's grade determination and the misclassification rate at the boundary between the head and tail of wine under different k values. The k value with the highest cross-validation accuracy and the lowest misclassification rate at the boundary is selected as the optimal value. Experimental verification shows that within the range of k value 37, the model's grade determination accuracy fluctuates by no more than 1.2%, and a smooth weight transition and stable classification effect can be achieved. k=5 is the optimal value, corresponding to the model with the highest cross-validation accuracy and the lowest misclassification rate at the boundary.

[0137] The transition rules for dynamic soft weights are as follows:

[0138] When the rate of change of alcohol content v < 1.5%vol / min, β(v) approaches 0.5, and α = 0.5;

[0139] When the rate of change of alcohol content is 1.5%vol / min≤v≤2.5%vol / min, β(v) smoothly transitions from 0.5 to 0.6, and α smoothly transitions from 0.5 to 0.4.

[0140] When the rate of change of alcohol content v > 2.5%vol / min, β(v) approaches 0.6 and α = 0.4.

[0141] Dynamic soft weights only take effect within the defined range of candidate sensory levels. The weights transition smoothly with the rate of change of alcohol content through the Sigmoid function, without hard threshold abrupt changes, thus avoiding weight abrupt changes and classification result oscillations caused by hard threshold switching.

[0142] The third step is to construct a binary classification SVM model. For each binary classification sub-task, samples corresponding to two sensory levels are selected as the training set, with the sample labels set to +1 and -1 respectively. The optimization objective function for the optimal classification hyperplane is then constructed as follows:

[0143] min(1 / 2)||w||²+C·Σξᵢ

[0144] The constraints are:

[0145] yᵢ·(w·zᵢ+b)≥1-ξᵢ,ξᵢ≥0

[0146] Where w is the normal vector of the classification hyperplane, b is the bias term of the classification hyperplane, C is the penalty coefficient, ξᵢ is the slack variable, yᵢ is the label value of the sample, zᵢ is the sample vector in the joint input matrix, and the optimization range of the penalty coefficient C is 1 to 10. The RBF kernel function is used to map the low-dimensional input data to a high-dimensional feature space, transforming the nonlinear classification problem in the low-dimensional space into a linear classification problem in the high-dimensional space. The expression of the RBF kernel function is:

[0147] K(zᵢ,zⱼ)=exp(-gamma·||zᵢ-zⱼ||²)

[0148] Where gamma is the kernel function parameter, the optimization range of gamma is from 0.001 to 100, and ||zᵢ-zⱼ|| is the Euclidean distance between two sample vectors.

[0149] The fourth step is to optimize the model parameters. The grid search method combined with 5-fold cross-validation is used to optimize the penalty coefficient C and the kernel function parameter gamma. The model set samples are randomly divided into 5 non-overlapping subsets. Each time, 4 subsets are selected as the training set and 1 subset is selected as the validation set. The training and validation process is repeated 5 times. The classification accuracy of the model under different parameter combinations is calculated, and the parameter combination with the highest classification accuracy is selected as the optimal parameters.

[0150] The fifth step is multi-class decision-making. For the sample to be classified, the sample is input into 15 binary SVM sub-models. Each binary sub-model outputs a classification result. The final classification level of the sample is determined by voting. The sensory level with the highest number of votes is the final sensory level judgment result of the sample.

[0151] In some embodiments, the sensory grade qualitative discrimination sub-model is constructed using a random forest multi-classification algorithm. The number of decision trees in the random forest is set to 100, the maximum tree depth is set to 10 layers, the Gini coefficient is used as the node splitting criterion, and the number of features selected when splitting a node is the square root of the total number of features. The model has better fitting ability for nonlinear features than the multi-classification SVM algorithm and is suitable for the classification scenario of baijiu grades with complex aromas and large differences in flavor components.

[0152] In some cases, considering that such models are sensitive to the quality of training data and the consistency of sensory evaluation, and that their generalization ability is uncertain across equipment and process scenarios, multi-expert blind evaluation calibration, batch data rolling updates, and same-process feature transfer mechanisms can be introduced during the model construction process. By unifying sensory evaluation standards, improving the representativeness of calibration samples, and strengthening feature reuse under the same process and fragrance type, the generalization and grading stability of the model across different equipment can be improved. For scenarios with large differences across fragrance types and processes, a retraining method based on an independent dataset of the target scenario can be adopted to ensure that the accuracy of the grade determination matches the actual working conditions.

[0153] S25: Set up a two-way closed-loop coupling rule for model training and inference, which corresponds completely to the hierarchical logic of the subsequent inference stage: In the training stage, the sample classification range is defined by the measured alcohol content as a hard constraint, and then the level training is completed by SVM with soft weights. Finally, the training bias of the alcohol content sub-model is corrected in reverse by the level calibration result; In the inference stage, the subsequent three-level logic is executed.

[0154] The specific execution process of the two-way closed-loop coupling rule is as follows: the first stage of the training phase is to define the classification range by hard constraints. The measured alcohol content data of the model set samples are compared with the preset alcohol content threshold ranges of each level. The corresponding candidate level range is defined for each sample, and non-candidate levels outside the threshold range are excluded. The classification boundary of each sample is limited, eliminating the ambiguity of sample classification and avoiding the blurring of classification boundaries caused by the overlap of spectral features of samples of different levels.

[0155] The second stage of the training phase is grade training with soft weights. Within the candidate grade range corresponding to each sample, a training subset of the multi-class SVM model is constructed. The multi-class SVM algorithm with dynamic soft weights is used to complete the model training, resulting in a sensory grade qualitative discrimination sub-model. The third stage of the training phase is inverse bias correction. The sensory grade calibration results of the training set samples are used as bias correction factors to systematically correct the prediction results of the alcohol content quantitative correction sub-model within the same grade range. The mean alcohol content prediction bias of all samples within the same grade is calculated, and this mean bias is used as the systematic correction amount for that grade. This is written into the graded correction table of the alcohol content quantitative correction sub-model. The graded correction table uses a key-value pair structure, where the key is a unique code for the six sensory grades, and the value is the corresponding systematic correction amount. The systematic correction amounts for each grade calculated during the model pre-training phase are initial fixed correction amounts. After the model completes pre-training, these are fixedly written into the graded correction table as the basic correction benchmark for the inference phase.

[0156] The system calibration quantity adopts an initial fixed benchmark combined with a rolling update mechanism during the production process. The update rules are as follows: First, single-batch rolling update: After the distillation of each batch, no less than 20 valid samples are randomly selected from the base spirits of each grade in that batch. The alcohol content is measured using laboratory benchmark methods, and at the same time, at least 3 qualified senior tasters complete the sensory grade review. The measured alcohol content value is compared with the alcohol content prediction value of the corresponding sample in the model, and the mean of the prediction deviation of all samples within the same grade is calculated as the batch calibration increment for that grade. The batch calibration increment is then weighted and fused with the currently effective system calibration quantity, with weighting coefficients of 0.3 and 0.7, respectively. The weighting coefficient for the batch calibration increment is 0.3, and the weighting coefficient for the current calibration quantity is 0.7, resulting in the updated system calibration quantity, which is then written into the graded calibration table. Single-batch updates are only triggered when there are no less than 20 valid samples for that grade. If the sample size is insufficient, no update is performed, and the batch calibration increment is accumulated to the next batch. Second, a full review and update is conducted quarterly. During each quarterly model iteration and optimization, the measured data of all valid wine samples accumulated in the current quarter and the review results of sommeliers are used to recalculate the system correction amount for each level, replace the current correction amount in the classification correction table, and complete the full update. The correction amount after the full update serves as the basis for the rolling update in the next quarter.

[0157] S26: Model training is performed using a modeling set. Model parameters are optimized using leave-one-out cross-validation. The model's generalization ability is verified using a validation set to ensure that the model accuracy meets the standards. Model accuracy must simultaneously meet the following indicators: alcohol content derivation error ≤ ±0.5% vol, sensory grade judgment accuracy ≥ 95%, abnormal wine identification accuracy ≥ 98%, validation set prediction bias ≤ ±2%, and model accuracy fluctuation ≤ 2% within a wine sample temperature fluctuation range of ±3℃.

[0158] S3: Pre-completion of hardware-level timing calibration of the integrated detection device: Calculate the flow path delay time Δt between the spectral acquisition flow cell and the alcohol content detection module, use the same clock source to trigger spectral acquisition and alcohol content detection, and after the spectral acquisition is completed, delay Δt to latch the corresponding alcohol content and temperature data to achieve acquisition timing synchronization; during the distillation process, through the calibrated integrated detection device, the near-infrared spectral data, alcohol content data and temperature data of the distilled liquor are acquired in real time and synchronously.

[0159] The specific steps for hardware-level timing calibration are as follows:

[0160] S31: Calculate the flow path volume L between the spectral acquisition flow cell and the alcohol content detection module. The flow path volume L is obtained by calculating the internal cavity volume of the flow path using 3D modeling software, or by actual measurement and calibration using a weighing method. The weighing method involves filling the flow path with deionized water with a resistivity ≥18.2 MΩ·cm, weighing the overall mass difference of the flow path before and after filling with water to obtain the mass m of the water in the flow path. Based on the density ρ of deionized water under experimental conditions, calculate the flow path volume L=m / ρ, with a volume measurement error ≤0.1ml. Combined with the design flow rate v of the distillate, which ranges from 1ml / s to 10ml / s and is determined according to the distillation flow rate of the baijiu distillation production line, calculate the theoretical delay time t0=L / v.

[0161] S32: Perform actual delay calibration using the step calibration method: Inject a standard concentration ethanol solution into the flow path, and record the time t1 when the spectral acquisition module detects the concentration change and the time t2 when the alcohol content detection module detects the concentration change, respectively. Calculate the actual flow path delay time Δt = t2 - t1. The specific execution process of the step calibration method is as follows: The inlet of the integrated detection device is connected to the standard ethanol solution storage tank and the deionized water storage tank via a peristaltic pump and a three-way solenoid valve, respectively. The flow rate of the peristaltic pump is set to be consistent with the design flow rate v of the distillation process. First, deionized water is continuously introduced into the flow path. After the liquid in the flow path is completely replaced and the detection data stabilizes, the flow path is quickly switched to 60% vol standard ethanol solution via the three-way solenoid valve. The switching time is ≤100 ms. Simultaneously, the absorbance data of the spectral acquisition module and the alcohol content data of the alcohol content detection module are synchronously recorded at a sampling frequency of 100 Hz through the control motherboard of the integrated detection device. The time t1 when the spectral acquisition module detects the absorbance change exceeding the preset threshold and the time t2 when the alcohol content detection module detects the alcohol content change exceeding the preset threshold are recorded. The actual flow path delay time Δt = t2 - t1 is calculated. The step calibration experiment is repeated 3 times, and the average value of Δt obtained from the 3 experiments is taken as the final actual delay time to ensure the measurement accuracy of the delay time.

[0162] S33: The Δt is written into the control system of the integrated detection device. The same 100Hz clock source is used to synchronously trigger the spectral acquisition module and the alcohol content detection module. After the spectral acquisition is completed, the corresponding alcohol content and temperature detection values ​​are latched after a delay of Δt, achieving hardware-level timing synchronization with a synchronization error ≤0.1s. During the distillation process, the calibrated integrated detection device synchronously acquires the near-infrared spectral data, alcohol content data, and temperature data of the distilled liquor in real time at a acquisition frequency of 1Hz. The acquired raw data is transmitted to the model calculation module in real time via industrial Ethernet.

[0163] In some embodiments, the hardware-level timing calibration adopts the cross-correlation calibration method, which continuously introduces an ethanol-water solution with a sinusoidal concentration into the flow path, collects spectral sequences and alcohol content sequences respectively, calculates the cross-correlation function of the two sequences, and the time offset corresponding to the peak value of the cross-correlation function is the flow path delay time Δt. This method is suitable for distillation production line scenarios with large flow rate fluctuations, and the calibration accuracy is improved by more than 20% compared with the step calibration method.

[0164] S4: Based on real-time temperature data, perform temperature-adaptive baseline correction on the spectral data. Input the corrected spectral data into the pre-trained bidirectional closed-loop coupled model, and synchronously complete alcohol content prediction and grade determination according to the following hierarchical logic:

[0165] Ⅰ Forward hard constraint: The alcohol content quantitative correction sub-model outputs the real-time alcohol content prediction value of the distilled liquid. Based on the preset alcohol content threshold range for each level, it defines the range of candidate sensory levels that meet the threshold requirements and excludes non-candidate levels outside the threshold range.

[0166] II. Sub-classification: Within the candidate sensory level range, the sensory level qualitative discrimination sub-model adopts a multi-class SVM algorithm with dynamic soft weights, using the corrected spectral data and real-time alcohol content prediction as joint inputs to complete the final sensory level sub-classification.

[0167] III. Reverse Bias Correction: The final sensory grade result is used as a bias correction factor to systematically correct the predicted alcohol content within the grade range, and the final alcohol content result is output.

[0168] The specific steps of temperature adaptive baseline correction are as follows:

[0169] S41: Near-infrared spectra of real base liquor calibration samples with different alcohol content gradients (5%-80% vol) within a temperature range of 20℃-60℃, different distillation fractions, and different fermentation batches were pre-collected to establish a quantitative correction model for temperature and spectral baseline shift. The temperature gradient was set to 2℃, and the alcohol content gradient to 5% vol. Three parallel spectra were collected for each temperature-alcohol content combination, and the average value was used as the standard spectral data for that combination, constructing a three-dimensional dataset of temperature-alcohol content-spectrum. Based on this three-dimensional dataset, a quantitative correction model for temperature and spectral baseline shift was established. A multiple linear regression algorithm was used, with the mean absorbance of the spectral data in the wavelength range of 1900nm to 2000nm as the dependent variable, and the liquor temperature and base liquor alcohol content as independent variables, to construct the regression equation:

[0170] A = a·T + b·C + c

[0171] Where A is the average absorbance, T is the temperature of the wine, C is the alcohol content of the base wine, and a, b, and c are regression coefficients. The regression coefficients are solved by the least squares method to obtain a quantitative correction model for the temperature and spectral baseline offset.

[0172] S42: For the raw near-infrared spectral data acquired in real time, baseline drift correction is performed using a quantitative correction model based on synchronously acquired temperature data, eliminating the interference of temperature fluctuations on the spectrum. The correction process involves calculating the spectral baseline shift based on the real-time acquired wine temperature data and the real-time predicted alcohol content using the quantitative correction model, and then subtracting the corresponding baseline shift from the full-wavelength absorbance value of the raw spectral data to complete the baseline drift correction.

[0173] S43: Preprocess the corrected spectral data by sequentially performing outlier removal, Savitzky-Golay convolutional smoothing and denoising, adaptive baseline correction, and wavelength normalization to obtain standard spectral feature data. The parameters for the preprocessing process are kept completely consistent with those used in the model training phase to ensure that the distribution of the preprocessed spectral data matches the distribution of the model training set, thus avoiding a decrease in model prediction accuracy due to data distribution shift.

[0174] The preprocessed standard spectral feature data is input into the pre-trained two-way closed-loop coupling model of spectrum-alcohol content-sensory level, and alcohol content prediction and level determination are completed synchronously according to the preset three-level logic.

[0175] In the first stage, forward hard constraints are implemented. The alcohol content quantitative correction sub-model outputs the real-time alcohol content prediction value of the distilled liquor. The real-time alcohol content prediction value is compared with the preset alcohol content threshold range for each level. The range of candidate sensory levels that meet the threshold requirements is defined, and non-candidate levels outside the threshold range are excluded. This limits the effective range of subsequent classification operations, eliminates the classification ambiguity caused by overlapping spectral features, and reduces the probability of misjudgment.

[0176] The second stage involves detailed classification and determination. Within the defined range of candidate sensory grades, the sensory grade qualitative discrimination sub-model employs a multi-class SVM algorithm with dynamic soft weights. The corrected standard spectral feature data and real-time alcohol content prediction are used as joint inputs to complete the final detailed sensory grade classification and determine the sensory grade of the distilled spirit.

[0177] The third stage performs reverse bias correction, using the final sensory grade result as the bias correction factor, retrieving the system correction amount corresponding to the grade in the grading correction table, adding the system correction amount to the alcohol content prediction value within the grade range, completing the system bias correction, and outputting the final alcohol content result.

[0178] S5: Based on the final sensory rating result, the intelligent diversion unit with anti-shake control logic completes the automatic grading and storage of wine of the corresponding grade.

[0179] The intelligent diversion unit includes an explosion-proof intelligent valve assembly and a multi-channel graded storage tank. The inlet of the explosion-proof intelligent valve assembly is connected to the outlet of the integrated detection device via a food-grade 304 stainless steel pipe and a flange. The outlet of the explosion-proof intelligent valve assembly is connected to the inlets of the six graded storage tanks via food-grade 304 stainless steel pipes and flanges. The six graded storage tanks correspond to the storage of base wines of six sensory levels. Each graded storage tank is equipped with an independent check valve at its inlet to prevent backflow and cross-contamination of the wine. The explosion-proof intelligent valve group uses six normally closed electric ball valves, each corresponding to a graded storage tank. The valve body of the electric ball valve is made of food-grade 304 stainless steel, and the valve seat is made of polytetrafluoroethylene. The nominal pressure of the valve is 1.6MPa, and the nominal diameter matches the diameter of the main conveying pipeline of the production line. The actuator of the electric ball valve is electrically connected to the intelligent control system through a shielded cable. The full opening and closing time of the valve is ≤0.5s, the protection level is IP67, and the explosion-proof level is ExdIIBT4GB, which is suitable for the explosion-proof environment requirements of the liquor distillation workshop.

[0180] The specific execution steps of the intelligent traffic splitting unit are as follows:

[0181] S51: The intelligent control system receives the final sensory level result output by the model calculation module and initiates the anti-shake verification logic.

[0182] S52: Only when the grade judgment results are completely consistent for 3 consecutive times with an interval of 0.5s, the electric ball valve of the corresponding grade is triggered to open, while the electric ball valves of other grades are closed. The valve operation is not executed when the judgment result changes once, so as to avoid frequent valve switching and mixing of wine grades caused by instantaneous fluctuations in the grade judgment result.

[0183] S53: By opening the explosion-proof intelligent valve group, the wine of the corresponding grade is diverted to the storage tank of the corresponding grade, and the automatic graded storage is completed.

[0184] In some embodiments, the valve of the intelligent diversion unit is a pneumatic diaphragm valve. The actuator of the pneumatic diaphragm valve is electrically connected to the intelligent control system through a pneumatic solenoid valve. The valve's opening and closing response time is ≤0.2s. The valve body flow channel adopts a straight-through, dead-angle-free structure with no liquid stagnation area, which is suitable for the rapid diversion requirements of high-flow distillation production lines.

[0185] After each round of distillation and alcohol collection is completed, the control system automatically starts the pipeline cleaning process: closes the corresponding valves of each graded storage tank, opens the cleaning branch and the sewage discharge channel, and sequentially uses deionized water and clean compressed air to circulate and dry the detection flow path, delivery pipeline and valve cavity to remove residual alcohol from the inner wall of the pipeline and avoid cross-interference between different batches and different grades of alcohol. After cleaning is completed, the sewage discharge channel is automatically closed and the system is reset to the ready-to-produce state.

[0186] In some embodiments, the system further includes S7: a full-process data closure and model adaptation step, which is specifically divided into the following steps:

[0187] The S7 system records the entire production process data for each batch of base liquor, creating a unique digital passport for each batch. Data integrity is 100%, and storage time is ≥5 years. The entire production process data includes near-infrared spectral data, alcohol content data, temperature data, sensory evaluation results, distillation time, valve action records, production shift information, production cycle information, and fermentation pit information for each sampling point. Each batch of base liquor's unique digital passport uses a unique batch code as its identifier. The batch code is composed of the production time, workshop number, shift number, and fermentation pit number, ensuring the uniqueness and traceability of each batch's digital passport. The digital passport data is stored on a local industrial server on the baijiu distillation production line. The industrial server uses redundant disk array storage to ensure data security and integrity, and the data storage time is no less than 5 years.

[0188] S72: Based on production operation data and standard sample verification results, fine-tune and optimize the parameters of the coupled model every quarter. Every quarter, randomly select no fewer than 30 valid samples from each grade of base liquor produced that quarter. At least three senior tasters with ≥10 years of experience and national-level tasting qualifications will conduct sensory grade verification using national-level standard samples as a benchmark. Simultaneously, laboratory benchmark methods will be used to measure alcohol content. The verified valid data will be added to the calibration dataset, and the model training and parameter optimization process will be re-executed to complete the quarterly iteration of the model. After iteration, the accuracy of the model's grade determination will improve by no less than 3%.

[0189] S73: To address the adaptation needs across wineries and aroma types, a transfer learning-based model adaptation method is adopted. This method maps the feature space of the pre-trained model to the feature space of the target scene. The target scene provides ≤50 sets of standard wine samples to complete the rapid model adaptation without the need for full remodeling. The specific execution process of the transfer learning-based model adaptation method is as follows: Spectral data, alcohol content data, and grade calibration data of no less than 50 sets of standard wine samples from the target scene are collected. Grade calibration is completed by at least 3 qualified senior sommeliers in the target scene according to a unified standard. A target domain dataset is constructed. A transfer component analysis algorithm is used to map the feature space of the source domain pre-trained model and the feature space of the target domain dataset to the same reproducing kernel Hilbert space, minimizing the maximum mean difference between the source and target domains, thus completing the feature space adaptation. Based on the adapted feature space, the classifier parameters of the model are fine-tuned to complete the rapid model adaptation without the need for full remodeling, significantly reducing the model deployment cycle and cost in new scenarios.

[0190] For the adaptation needs of different wineries with the same aroma type and process, a model adaptation method based on transfer learning is adopted to map the feature space of the pre-trained model to the target equipment scene. Only ≤50 sets of standard wine samples are required from the target scene to complete the rapid model adaptation without the need for full remodeling. For different wineries with different aroma types and processes, due to the significant differences in wine composition and spectral characteristics, the model is retrained based on the target scene's own calibration dataset to ensure the accuracy of grade determination and on-site applicability.

[0191] Example 2

[0192] A quantification and quality-based distillation system based on intelligent sensing is used to implement the aforementioned quantification and quality-based distillation method. Applied to a baijiu distillation production line, it includes a grading benchmark storage module, a calibration dataset management module, a model calculation module, an integrated online detection unit, an intelligent control unit, and a diversion execution unit. The grading benchmark storage module stores the sensory grade classification rules for six base liquors anchored to national standards, the corresponding alcohol content threshold ranges for each grade, and standardized sensory evaluation indicators. The calibration dataset management module stores a pre-constructed, time-aligned base liquor calibration dataset, which includes one-to-one corresponding near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data. The integrated online detection unit pre-completes hardware-level time-series calibration and integrates a near-infrared spectral acquisition module, a high-precision alcohol content detection module, and a temperature acquisition module triggered by the same clock source, used for real-time synchronous acquisition of near-infrared spectral data, alcohol content data, and temperature data of the distilled liquor. The model computation module incorporates a pre-trained, bidirectional closed-loop coupled model of spectral density, alcohol content, and sensory grade. This model integrates a quantitative alcohol content correction sub-model and a qualitative sensory grade discrimination sub-model. The module performs temperature-adaptive baseline correction of the spectral data and completes inference operations according to the following hierarchical logic: Ⅰ Forward hard constraints: The quantitative alcohol content correction sub-model outputs real-time alcohol content predictions to define the candidate sensory grade range; Ⅱ Fine-classification judgment: Within the candidate range, a multi-class SVM algorithm with dynamic soft weights performs fine-classification of sensory grades; Ⅲ Backward bias correction: Using the final grade result as a correction factor, the systemic bias of the alcohol content prediction value is corrected. The intelligent control unit incorporates anti-shake control logic to drive the distribution execution unit to automatically classify and store the corresponding grade of alcoholic beverages based on the final sensory grade result output by the model. The distribution execution unit is electrically connected to the intelligent control unit.

[0193] The integrated online detection unit completes hardware-level flow path timing calibration, with a synchronization error of ≤0.1s between spectral acquisition and alcohol content detection. The operating parameters of the integrated online detection unit meet the following requirements: spectral acquisition stability error ≤±0.02Abs, response time ≤1s; alcohol content detection error ≤±0.3%vol, response time ≤0.5s; temperature acquisition accuracy ±0.1℃; the device protection level is IP65, and the explosion-proof level meets the safety production standards of the liquor distillation workshop. The constant temperature water bath module of the integrated online detection unit controls the temperature of the liquor sample in the detection flow path at 40±2℃, forming a dual temperature compensation mechanism of "physical constant temperature + algorithm correction" with the temperature adaptive correction functions of the temperature acquisition module and the model calculation module. The model calculation module has a built-in spectral preprocessing subunit, which sequentially performs outlier removal, Savitzky-Golay convolution smoothing and denoising, multivariate scattering correction, and standard normal variable transformation of the spectral data, outputting standard spectral feature data. The diversion execution unit includes an explosion-proof intelligent valve group and a 6-channel graded storage tank. The total system response time of the intelligent control unit is ≤1.5s, the valve switching response time is ≤0.5s, and the diversion accuracy is ≥99%.

[0194] This intelligent sensing-based quantitative and qualitative distillation method and system achieves simultaneous acquisition of near-infrared spectral data, alcohol content data, and temperature data through hardware-level time-series calibration. This eliminates the time-series inaccuracy problem between the calibration dataset and real-time acquired data caused by flow path delays, improving data consistency during model training and inference. A dual-temperature compensation mechanism combining temperature adaptive baseline correction and physical isothermal control eliminates spectral baseline drift caused by temperature fluctuations during distillation, improving the model's operational stability in industrial production scenarios. A hierarchical logic that defines the candidate sensory level range through forward hard constraints, performs fine-grained classification within this range, and ultimately corrects the alcohol content prediction value using the level results eliminates the spectral feature ambiguity problem caused by the nonlinear correspondence between alcohol content and sensory levels, reducing the misclassification rate and improving the accuracy of alcohol content detection. A multi-classification algorithm with dynamic soft weights adapts to the differences in alcohol content change rates at different stages of distillation, improving the accuracy of level boundary determination. Anti-jitter control logic avoids frequent valve switching caused by instantaneous fluctuations in level determination results, eliminating the risk of mixed levels. By employing a transfer learning-based model adaptation method, rapid model adaptation to different wineries and aroma profiles was achieved without requiring a complete remodeling, reducing the cost and time required for technology implementation. A tiered calibration table combined with an initial fixed and rolling update calibration management mechanism enabled continuous optimization of alcohol content prediction accuracy, adapting to variations in base liquor quality across different production cycles and batches, thus improving the model's long-term operational stability. A full-process data recording and closed-loop optimization mechanism ensured traceability throughout the base liquor production process, continuously enhancing the model's accuracy and production stability. Standardized taster qualification requirements, blind tasting procedures, and consistency rules transformed traditional winemaking experience into quantifiable and reproducible calibration data, addressing the issue of traditional distillation techniques relying on personal experience and being difficult to replicate and pass on. This also ensured the accuracy and consistency of model training labels, improving the model's grading results' relevance to consumer taste preferences.

[0195] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0196] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0197] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0198] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0200] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0201] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

Claims

1. A method for graded distillation based on intelligent sensing, applied to the graded distillation of baijiu (Chinese liquor) distillation production line, characterized in that, Includes the following steps: S1: A pre-constructed time-aligned base wine calibration dataset, which includes one-to-one corresponding near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data based on preset wine standards. The sensory grades are divided into six levels: head, premium, first-grade, second-grade, third-grade, and tail, with each level corresponding to a preset alcohol content threshold range. S2: Based on the calibration dataset, a two-way closed-loop coupling model of spectrum-alcohol content-sensory level is pre-trained and constructed. The two-way closed-loop coupling model integrates the quantitative alcohol content correction sub-model and the qualitative sensory level discrimination sub-model. S3: Pre-completion of hardware-level timing calibration of the integrated detection device: Calculate the flow path delay time Δt between the spectral acquisition flow cell and the alcohol content detection module, use the same clock source to trigger spectral acquisition and alcohol content detection, and after the spectral acquisition is completed, delay Δt to latch the corresponding alcohol content and temperature data to achieve acquisition timing synchronization; during the distillation process, through the calibrated integrated detection device, the near-infrared spectral data, alcohol content data and temperature data of the distilled liquor are acquired in real time and synchronously. S4: Based on real-time temperature data, perform temperature-adaptive baseline correction on the spectral data. Input the corrected spectral data into the pre-trained bidirectional closed-loop coupled model, and synchronously complete alcohol content prediction and grade determination according to the following hierarchical logic: Ⅰ Forward hard constraint: The alcohol content quantitative correction sub-model outputs the real-time alcohol content prediction value of the distilled liquid. Based on the alcohol content threshold range of each level preset by S1, it defines the range of candidate sensory levels that meet the threshold requirements and excludes non-candidate levels outside the threshold range. II. Sub-classification: Within the candidate sensory level range, the sensory level qualitative discrimination sub-model adopts a multi-class SVM algorithm with dynamic soft weights, using the corrected spectral data and real-time alcohol content prediction as joint inputs to complete the final sensory level sub-classification. III. Reverse Bias Correction: The final sensory rating result is used as a bias correction factor to systematically correct the predicted alcohol content within the range of that rating, and the final alcohol content result is output. S5: Based on the final sensory rating result, the intelligent diversion unit with anti-shake control logic completes the automatic grading and storage of wine of the corresponding grade.

2. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S11: Based on the target preset alcohol standards, and using the corresponding aroma type preset standard alcohol samples as anchor benchmarks, establish the classification rules for 6 sensory levels, the corresponding alcohol content threshold range for each level, and standardized sensory evaluation indicators. S12: Collect base liquor calibration samples at the receiving port of the baijiu distillation production line, divided into three stages: head, middle and tail of liquor, covering different production shifts, batches and all working conditions of the cellars. The total number of samples shall not be less than 200, and the number of single-grade samples shall not be less than 15. S13: Complete synchronous collection for each calibration sample, obtain the near-infrared spectral data, measured alcohol content data and synchronous temperature data of the sample, and complete the sensory level calibration of the sample according to the standardized sensory evaluation indicators. S14: The initial data alignment is completed by hardware-level time-series calibration with the pre-completed integrated detection device. Then, the cross-correlation algorithm is used to perform secondary time-series alignment of the spectral sequence and the alcohol content sequence, and finally a one-to-one corresponding calibration dataset is formed.

3. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1, characterized in that, The specific steps of hardware-level timing calibration in step S3 are as follows: S31: Calculate the flow path volume L between the spectral acquisition flow cell and the alcohol content detection module, and determine the theoretical delay time t0 = L / v by combining the design flow rate v of the distillate; S32: Perform actual delay calibration using the step calibration method: Inject a standard concentration ethanol solution into the flow path, and record the time t1 when the spectral acquisition module detects the concentration change and the time t2 when the alcohol content detection module detects the concentration change, respectively. Calculate the actual flow path delay time Δt = t2 - t1. S33: Write Δt into the control system, use the same 100Hz clock source to synchronously trigger the spectral acquisition module and the alcohol content detection module, and after the spectral acquisition is completed, delay Δt time to latch the corresponding alcohol content and temperature detection values ​​to achieve hardware-level timing synchronization with a synchronization error ≤0.1s.

4. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1, characterized in that, The specific steps of temperature adaptive baseline correction in step S4 are as follows: S41: Pre-collect near-infrared spectra of real base wine standard samples with a concentration gradient of 5%-80%vol within a temperature range of 20℃-60℃, and establish a quantitative correction model for temperature and spectral baseline offset; S42: For the raw near-infrared spectral data acquired in real time, baseline drift correction is completed through a quantitative correction model using synchronously acquired temperature data, thereby eliminating the interference of temperature fluctuations on the spectrum; S43: Perform preprocessing on the corrected spectral data, sequentially performing outlier removal, Savitzky-Golay convolutional smoothing and denoising, adaptive baseline correction, and wavelength normalization to obtain standard spectral feature data.

5. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S21: Perform temperature adaptive correction and preprocessing on the spectral data of the calibration dataset to obtain a standard spectral feature dataset; S22: Using the Kennard-Stone algorithm, the standard spectral feature dataset, the corresponding measured alcohol content data, and the sensory grade data are divided into a modeling set and a validation set in a 9:1 ratio. S23: Construct a quantitative alcohol content calibration sub-model, using partial least squares regression (PLS) algorithm, with standard spectral characteristic data as input and measured alcohol content as output, and determine the optimal number of principal components to be 5-8 through cross-validation, with a cumulative contribution rate ≥95%; S24: Construct a qualitative sensory grade discrimination sub-model, using a multi-class SVM algorithm with dynamic soft weights, with standard spectral feature data and the synchronous output of the alcohol content quantitative correction sub-model as joint inputs and the sensory grade of the base wine as the output; S25: Set up a two-way closed-loop coupling rule for model training and inference, which corresponds completely to the hierarchical logic of step S4: During the training phase, the sample classification range is defined by the measured alcohol content as a hard constraint, and then the level training is completed by SVM with soft weights. Finally, the training bias of the alcohol content sub-model is corrected in reverse by the level calibration result; During the inference phase, the logic of steps I-III of step S4 is executed. S26: The model is trained using a modeling set, the model parameters are optimized using leave-one-out cross-validation, and the model's generalization ability is verified using a validation set to ensure that the model's accuracy meets the standards.

6. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1 or 5, characterized in that, The specific rules for the dynamic soft weights in step S4Ⅱ are as follows: S61: The dynamic soft weighting only applies within the candidate sensory level range defined in step S4Ⅰ. The weighting transitions smoothly with the rate of change of alcohol content through the Sigmoid function, without any hard threshold abrupt changes. S62: The weight transition rule is as follows: When the rate of change of alcohol content v < 1.5%vol / min, the weight of alcohol content is 0.5, and the weight of spectral characteristics is 0.

5. When the rate of change of alcohol content is 1.5%vol / min≤v≤2.5%vol / min, the weights are smoothly transitioned through the Sigmoid function, with the alcohol content weight smoothly increasing from 0.5 to 0.6 and the spectral feature weight smoothly decreasing from 0.5 to 0.

4. When the rate of change of alcohol content v > 2.5% vol / min, the weight of alcohol content is 0.6 and the weight of spectral characteristics is 0.

4. S63: The rate of change of alcohol content is calculated using the slope of a linear fit with a 10-second sliding window. The input alcohol content data is first processed by a 5-point sliding window Kalman filter to eliminate detection noise. S64: The multi-class SVM algorithm uses the RBF kernel function and optimizes the penalty coefficient C and kernel function parameter gamma through grid search combined with 5-fold cross-validation. The optimization range of the penalty coefficient C is 1~10.

7. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S51: The intelligent control system receives the final sensory level result output by the model and initiates the anti-shake verification logic; S52: The valve switching action of the corresponding level is triggered only when the level judgment results are completely consistent for 3 consecutive times with an interval of 0.5s. The valve operation is not executed when the judgment result changes once. S53: Through the explosion-proof intelligent valve group, the wine of the corresponding grade is diverted to the storage tank of the corresponding grade to complete the automatic graded storage.

8. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 5, characterized in that, In step S26, the model accuracy must simultaneously meet the following indicators: alcohol content derivation error ≤ ±0.5%vol, sensory grade judgment accuracy ≥95%, abnormal wine identification accuracy ≥98%, validation set prediction deviation ≤ ±2%, and model accuracy fluctuation within the range of wine sample temperature fluctuation ±3℃ ≤2%.

9. The method for quantitative and qualitative wine extraction based on intelligent sensing according to claim 1, characterized in that, It also includes the S7 end-to-end data closure and model adaptation steps, which are specifically divided into the following steps: S71: Record the entire production process data of each batch of liquor, build a unique digital passport for each batch of base liquor, with 100% data integrity and storage time ≥ 5 years; S72: Based on production operation data and standard wine sample verification results, complete the parameter fine-tuning and optimization of the coupled model every quarter; S73: To meet the adaptation needs across wineries and aroma types, a model adaptation method based on transfer learning is adopted to map the feature space of the pre-trained model to the feature space of the target scene. The target scene provides ≤50 sets of standard wine samples to complete the rapid model adaptation.

10. A quantitative and qualitative distillation system based on intelligent sensing, used to implement the quantitative and qualitative distillation method according to any one of claims 1-9, applied to a baijiu distillation production line, characterized in that, It includes a hierarchical benchmark storage module, a calibration dataset management module, a model computation module, an integrated online detection unit, an intelligent control unit, and a traffic splitting execution unit; The grading benchmark storage module is used to store the sensory grade classification rules of six base liquors anchored to national standards, the alcohol content threshold range corresponding to each grade, and standardized sensory evaluation indicators. The calibration dataset management module is used to store pre-built time-aligned base wine calibration datasets, which contain one-to-one corresponding near-infrared spectral data, measured alcohol content data, synchronous temperature data, and sensory grade data. The integrated online detection unit has completed hardware-level timing calibration in advance and has built-in near-infrared spectral acquisition module, high-precision alcohol content detection module and temperature acquisition module triggered by the same clock source, which are used to synchronously acquire near-infrared spectral data, alcohol content data and temperature data of distilled liquor in real time. The model computation module incorporates a pre-trained bidirectional closed-loop coupled model of spectral-alcohol content-sensory level. This model integrates a quantitative alcohol content correction sub-model and a qualitative sensory level discrimination sub-model. The model computation module is used to perform temperature-adaptive baseline correction of the spectral data and completes inference operations according to the following hierarchical logic: Ⅰ Forward hard constraint: The real-time alcohol content prediction value is output through the quantitative alcohol content correction sub-model to define the candidate sensory level range; Ⅱ Fine classification determination: Within the candidate range, the sensory level is finely classified using a multi-class SVM algorithm with dynamic soft weights; Ⅲ Backward bias correction: The systematic bias of the alcohol content prediction value is corrected using the final level result as a correction factor. The intelligent control unit has built-in anti-shake control logic, which is used to drive the diversion execution unit to complete the automatic grading and storage of wine of the corresponding grade according to the final sensory level result output by the model. The shunt execution unit is electrically connected to the intelligent control unit.

Images