CN-121999909-A - Method, device, equipment and medium for predicting bulk density of base wine

CN121999909ACN 121999909 ACN121999909 ACN 121999909ACN-121999909-A

Abstract

The invention provides a base wine body density prediction method, device, equipment and medium, which comprise the steps of obtaining multi-mode sensing data influencing the change of the base wine body density, sequentially carrying out data preprocessing and feature engineering processing on the multi-mode sensing data to obtain multi-mode sensing features, constructing a deep learning model integrating physical information constraint based on the multi-mode sensing features, training the deep learning model through a total loss function to obtain a base wine body density prediction model, and inputting data to be predicted into the base wine body density prediction model to obtain target base wine body density prediction data. Compared with the prior art, the method realizes non-contact, automatic and high-precision dynamic tracking and prediction of the density of the base wine body.

Inventors

LU YIN
CHEN ZIYANG
CAI YIJIE

Assignees

天翼物联科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260123

Claims (10)

1. A method for predicting the bulk density of a base wine is characterized by comprising the following steps: Acquiring multi-mode sensing data affecting the density change of the wine body of the base wine; sequentially carrying out data preprocessing and feature engineering processing on the multi-modal sensing data to obtain multi-modal sensing features; Based on the multi-modal sensing characteristics, constructing a deep learning model integrating physical information constraint, wherein the deep learning model at least comprises a graph convolution network module for capturing the dependency relationship among multiple variables, a time transform module for capturing the long-range time dependency and a decoding network module for fusing space-time information; training the deep learning model through a total loss function based on a training set containing historical multi-modal sensing data and corresponding actual measurement density labels to obtain a base wine body density prediction model; and acquiring data to be predicted related to the target base wine body density, and inputting the data to be predicted into the base wine body density prediction model to obtain the target base wine body density prediction data.
2. The method for predicting the bulk density of a base wine according to claim 1, wherein the sequentially performing data preprocessing and feature engineering processing on the multi-modal sensing data to obtain multi-modal sensing features comprises: Performing normalization processing on continuous numerical variables in the multi-modal awareness data, wherein the continuous numerical variables comprise alcohol concentration, ambient temperature and storage time; performing principal component analysis dimension reduction processing on high-dimensional variables representing chemical components in the multi-modal sensing data to extract low-dimensional feature vectors retaining main information components, wherein the high-dimensional time sequence variables comprise acid ester data representing chemical components of the wine body; Performing hierarchical embedded coding processing on classification variables in the multi-mode perception data, and mapping classification identifiers into low-dimensional dense vector representations, wherein the classification variables comprise materials of containers; And extracting the characteristics of time sequence data in the multi-mode sensing data to obtain time sequence characteristics.
3. The method for predicting the bulk density of a base wine according to claim 1, wherein the constructing a deep learning model integrating physical information constraint based on the multi-modal perceptual features comprises: Combining the graph convolution network module with the time transducer module to form a deep neural network backbone; constructing a physical information loss term in a loss function of the deep neural network trunk, wherein the physical information loss term calculates partial derivatives of a density value predicted by the network on time and components through an automatic differentiation technology, and minimizes residual errors meeting a preset simplified reaction diffusion equation; constructing a boundary condition loss term in a loss function of the deep neural network trunk, wherein the boundary condition loss term is used for restricting the rationality of model prediction under the known boundary condition; Constructing a supervised learning loss term based on the difference between the multi-modal perceptual features and the actual wine body density value measured by manual sampling; And carrying out weighted summation on the physical information loss term, the boundary condition loss term and the supervised learning loss term to jointly form a total loss function of the deep learning model.
4. A method for predicting bulk density of a base wine according to claim 3, wherein training the deep learning model to obtain a base wine bulk density prediction model comprises: Dividing the multi-mode perception characteristics of the historical time period and the corresponding real wine density values into a training set and a verification set; Performing iterative training on the deep learning model by adopting the training set, calculating gradient according to the total loss function in each iteration, and updating model parameters; In the training process, dynamically adjusting the weight coefficients of the physical information loss item and the boundary condition loss item by adopting a meta learning strategy; and monitoring the performance of the deep learning model in the training process by adopting the verification set, stopping training when the prediction error of the model on the verification set reaches a preset standard and is not reduced any more, and outputting the model with the fixed parameters as the base wine body density prediction model.
5. The base wine body density prediction method according to claim 4, wherein the iterative training of the deep learning model using the training set comprises: Taking the actual measurement environment and component parameters of each moment in the training set as graph node characteristics Constructing an adjacency matrix by calculating the pearson correlation coefficient or mutual information between the component parameters ; Sampling the historical time series data in the training set by adopting a sliding window, and constructing an input size of Wherein T is the time step, N is the number of parameter variables, and D is the feature dimension; Graph node characteristics for each time step t Performing graph rolling operation by using a graph rolling network module to obtain a characteristic sequence fusing space structure information ; The time transducer module is adopted for the characteristic sequence Coding processing is carried out to obtain global coding feature vector of comprehensive variable space-time information ; The global coding feature vector Fusing through a full connection layer, then sending into a multi-layer perceptron decoding network for decoding, and finally outputting the density predicted value of the next time step t+1 。
6. The method for predicting the bulk density of a base wine according to claim 4, wherein training the deep learning model to obtain the bulk density prediction model of the base wine further comprises: performing forward propagation for multiple times in a model reasoning stage by adopting a Bayesian deep learning technology, and introducing random discarding to calculate the mean value and variance of multiple prediction results of the deep learning model for the same input; Taking the average value of the prediction result as a final density prediction value, and converting the variance into an uncertainty confidence interval of the prediction; and outputting the density predicted value and the corresponding uncertainty confidence interval.
7. The method for predicting the bulk density of a base wine according to claim 4, wherein training the deep learning model to obtain the bulk density prediction model of the base wine further comprises: The deep learning model is expanded into a multi-task learning framework with simultaneous prediction density and auxiliary variables, the multi-task learning framework is shared with an underlying network based on the relevance of tasks, and the total loss of the multi-task learning framework is set as a weighted sum of the task losses.
8. A base wine bulk density prediction device, comprising: The data acquisition module is used for acquiring multi-mode sensing data affecting the density change of the wine body of the base wine; the feature acquisition module is used for sequentially carrying out data preprocessing and feature engineering processing on the multi-mode sensing data to obtain multi-mode sensing features; The deep learning module is used for constructing a deep learning model integrating physical information constraint based on the multi-modal sensing characteristics, and at least comprises a graph convolution network module used for capturing the dependency relationship among multiple variables, a time transform module used for capturing the long-range time dependency and a decoding network module used for fusing space-time information; The training module is used for training the deep learning model through a total loss function based on a training set containing historical multi-modal perception data and corresponding actual measurement density labels to obtain a base wine body density prediction model; The prediction module is used for collecting to-be-predicted data related to the target base wine body density, and inputting the to-be-predicted data into the base wine body density prediction model to obtain the target base wine body density prediction data.
9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the base wine bulk density prediction method according to any one of claims 1 to 7.
10. A computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the base wine body density prediction method according to any one of claims 1 to 7.

Description

Method, device, equipment and medium for predicting bulk density of base wine Technical Field The invention relates to the technical field of liquor density detection, in particular to a method, a device, equipment and a medium for predicting liquor density of base liquor. Background In the white spirit industry chain, accurate metering and quality monitoring of base wine (namely original wine for long-term storage) are key links of warehouse management, asset estimation and subsequent blending processes. The density of the wine body is taken as a core parameter for calculating the quality, the dynamic evolution of the density directly reflects the ageing processes such as alcohol volatilization, acid ester reaction and the like, and the density is an important index for measuring the quality and stability of the wine body. Currently, the mainstream detection method for the density of base wine in the industry relies on traditional manual contact sampling. The method requires operators to open the jar periodically, use tools such as a siphon pipe to extract wine samples, and send the wine samples to a laboratory for analysis by using a densitometer. The technical scheme has the main defects that firstly, the labor intensity is high, the efficiency is low, the frequent inventory requirement of a large-scale wine warehouse is difficult to deal with, secondly, the original sealing microenvironment of a wine body can be destroyed by opening a jar for sampling, the normal aging chemical reaction can be interfered, the quality risk is introduced, and finally, the method can only provide static data at discrete time points, can not continuously capture the dynamic change track of the density in the storage period, so that the quality monitoring has hysteresis and is difficult to discover abnormality in time. Disclosure of Invention The invention provides a method, a device, equipment and a medium for predicting the bulk density of base wine. By collecting multi-mode perception data, a deep learning model fused with physical constraints is constructed, and the deep learning model is combined with a physical chemistry principle in the white spirit storage process, so that high-precision and interpretable dynamic tracking and prediction of the density of the base wine body can be realized without opening a jar for sampling. The method comprises the steps of obtaining multi-modal sensing data affecting base liquor body density change, sequentially carrying out data preprocessing and feature engineering processing on the multi-modal sensing data to obtain multi-modal sensing features, constructing a deep learning model integrating physical information constraint based on the multi-modal sensing features, wherein the deep learning model at least comprises a graph convolution network module used for capturing multi-variable dependency, a time transform module used for capturing long-range time dependency and a decoding network module used for fusing space-time information, training the deep learning model through a total loss function based on a training set containing historical multi-modal sensing data and corresponding actual measurement density labels to obtain a base liquor body density prediction model, collecting data to be predicted related to target base liquor body density, and inputting the data to be predicted into the base liquor body density prediction model to obtain the target base liquor body density prediction data. According to one embodiment of the invention, the multi-modal sensing data is subjected to data preprocessing and feature engineering processing in sequence to obtain multi-modal sensing characteristics, wherein the multi-modal sensing data comprises the steps of performing normalization processing on continuous numerical variables in the multi-modal sensing data, wherein the continuous numerical variables comprise alcohol concentration, ambient temperature and storage time, performing principal component analysis dimension reduction processing on high-dimensional variables representing chemical components in the multi-modal sensing data to extract low-dimensional feature vectors retaining main information components, wherein the high-dimensional time sequence variables comprise acid ester data representing chemical components of a wine body, performing hierarchical embedded coding processing on classification variables in the multi-modal sensing data to map classification identifications into low-dimensional dense vector representations, wherein the classification variables comprise materials of containers, and performing feature extraction on time sequence data in the multi-modal sensing data to obtain time sequence characteristics. According to one embodiment of the invention, a deep learning model integrating physical information constraint is constructed based on the multi-modal sensing characteristics, the deep learning model comprises a deep neural network trunk formed by combining th