Tagatose milk tea processing technology evaluation method and system

Abstract

The present invention relates to the technical field of processing technology evaluation, and discloses a tagatose milk tea processing technology evaluation method and system, the method comprising: collecting and preprocessing multidimensional parameter data in the tagatose milk tea production process to obtain a multidimensional time series data set; inputting a TSVT-PNNT model for data reconstruction to obtain reconstructed data; performing multi-feature complementary fusion to obtain full-band fusion features; performing process parameter constraint analysis to obtain a process parameter constraint range; using a Bayesian network and a reinforcement learning method to generate multiple candidate process parameter adjustment schemes; in a variety of production scenarios, respectively verifying and evaluating multiple candidate process parameter adjustment schemes, selecting the optimal candidate process parameter adjustment scheme as a target process parameter adjustment scheme and outputting it. The present invention can adaptively generate optimized process parameter adjustment schemes according to different production scenarios to improve the production efficiency and product quality stability of tagatose milk tea.

Description

Technical Field

[0001] The present invention relates to the technical field of processing technology evaluation, and in particular to a tagatose milk tea processing technology evaluation method and system. Background Art

[0002] As a new sweetener, tagatose is increasingly used in the production of sugar-free milk tea. However, due to the difference in the solubility characteristics of tagatose and traditional sucrose, its process control in the production of milk tea faces many challenges. Traditional milk tea production process evaluation methods are difficult to accurately capture the subtle changes in the dissolution process of tagatose, resulting in unstable product quality and high energy consumption.

[0003] Existing process evaluation methods often ignore the complex interactions between multi-dimensional parameters in the tagatose milk tea production process and cannot effectively deal with data noise and missing problems in the production process. In addition, traditional methods lack sufficient adaptability and robustness in the face of changing production environments, making it difficult to provide optimized process parameter adjustment solutions for different production scenarios. Summary of the invention

[0004] The present invention provides a tagatose milk tea processing technology evaluation method and system, which can adaptively generate optimized process parameter adjustment schemes according to different production scenarios to improve the production efficiency and product quality stability of tagatose milk tea.

[0005] In a first aspect, the present invention provides a tagatose milk tea processing technology evaluation method, the tagatose milk tea processing technology evaluation method comprising:

[0006] The multidimensional parameter data of tagatose milk tea production process was collected and preprocessed to obtain a multidimensional time series data set;

[0007] Inputting the multidimensional time series data set into the TSVT-PNNT model to reconstruct the data to obtain reconstructed data;

[0008] Based on the reconstructed data, multiple features are complementary fused to obtain full-band fusion features;

[0009] Inputting the full-band fusion feature structure into a multi-objective integer linear programming model to perform process parameter constraint analysis to obtain a process parameter constraint range;

[0010] Based on the process parameter constraint range, a plurality of candidate process parameter adjustment schemes are generated using a Bayesian network and a reinforcement learning method;

[0011] In various production scenarios, the multiple candidate process parameter adjustment schemes are verified and evaluated respectively, and the optimal candidate process parameter adjustment scheme is selected and output as the target process parameter adjustment scheme.

[0012] In a second aspect, the present invention provides a tagatose milk tea processing technology evaluation system, the tagatose milk tea processing technology evaluation system comprising:

[0013] The acquisition module is used to collect and preprocess the multidimensional parameter data in the tagatose milk tea production process to obtain a multidimensional time series data set;

[0014] A reconstruction module, used for inputting the multidimensional time series data set into the TSVT-PNNT model to reconstruct the data and obtain reconstructed data;

[0015] A fusion module, used for performing complementary fusion of multiple features based on the reconstructed data to obtain full-band fusion features;

[0016] An analysis module is used to input the full-band fusion feature structure into a multi-objective integer linear programming model to perform process parameter constraint analysis to obtain a process parameter constraint range;

[0017] A generation module, for generating a plurality of candidate process parameter adjustment schemes based on the process parameter constraint range by using a Bayesian network and a reinforcement learning method;

[0018] The evaluation module is used to verify and evaluate the multiple candidate process parameter adjustment schemes in multiple production scenarios, select the best candidate process parameter adjustment scheme as the target process parameter adjustment scheme and output it.

[0019] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned tagatose milk tea processing technology evaluation method.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned tagatose milk tea processing technology evaluation method.

[0021] In the technical solution provided by the present invention, through the collection and preprocessing of multidimensional parameter data, combined with the TSVT-PNNT model for data reconstruction, the data quality is effectively improved, the influence of noise and missing values ​​on subsequent analysis is reduced, and the multi-feature complementary fusion technology is adopted to comprehensively utilize high-frequency and low-frequency feature information to comprehensively capture various dynamic characteristics in the tagatose milk tea production process, thereby improving the comprehensiveness and accuracy of process evaluation. A multi-objective integer linear programming model is introduced to perform process parameter constraint analysis, taking into account multiple production objectives and constraints, which helps to optimize energy consumption and production efficiency while ensuring product quality. Combining Bayesian networks and reinforcement learning methods to generate candidate process parameter adjustment schemes, historical data and expert knowledge are fully utilized to improve the feasibility and adaptability of the scheme. The candidate schemes are verified and evaluated in a variety of production scenarios, and simulation and decision-making methods are adopted to enhance the robustness of process evaluation and ensure the performance of the selected scheme under different production environments. The present invention can adaptively generate optimized process parameter adjustment schemes according to different production scenarios to improve the production efficiency and product quality stability of tagatose milk tea. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0023] Figure 1 Schematic diagram of the steps of the tagatose milk tea processing technology evaluation method in an embodiment of the present invention;

[0024] Figure 2 It is a structural schematic diagram of a tagatose milk tea processing technology evaluation system in an embodiment of the present invention;

[0025] Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] Embodiments of the present invention provide a tagatose milk tea processing technology evaluation method and system. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0027] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , one embodiment of the tagatose milk tea processing technology evaluation method in the embodiment of the present invention includes:

[0028] Step S1, collecting and preprocessing multidimensional parameter data in the tagatose milk tea production process to obtain a multidimensional time series data set;

[0029] It is understandable that the execution subject of the present invention may be a tagatose milk tea processing technology evaluation system, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as an execution subject as an example.

[0030] Specifically, high-precision sensors are installed at the key process nodes of raw material ratio, dissolution, heating, cooling, blending and packaging of the tagatose milk tea production line to collect a variety of key process parameters, including equipment temperature, stirring speed, solubility, raw material ratio, equipment cleanliness, tea extraction temperature and tea extraction time. Multidimensional parameters jointly describe the state and changes in the tagatose milk tea production process. The collected data is transmitted to the central processing unit through the industrial Internet of Things to obtain the original data set representing the production process. The original data set is divided into time series to obtain the initial time series data. The original data is arranged in chronological order to ensure that the data can reflect the dynamic changes in the production process. The outliers in the initial time series data are identified, and the dynamic threshold is set in combination with the dissolution characteristics of tagatose in the production process. The dynamic threshold is adjusted according to the normal fluctuation range of the parameters in the production process to achieve accurate judgment of the data points that deviate from the normal range. For data points whose deviation exceeds the set threshold, they are marked as outliers to obtain a data set containing outlier markers. Based on the outlier labeled data, local outlier processing is performed on the data points marked as outliers to retain the characteristics of normal data to the greatest extent, while minimizing the impact of outliers on the overall data distribution, and obtaining a data set after removing outliers. Multiple interpolation is performed on the missing values ​​in the outlier-removed data set. Multiple interpolation is a method based on statistics and machine learning, which is used to reasonably estimate and complete missing data to ensure the integrity of the data set and obtain a completed data set. Adaptive standardization is performed on the data of each parameter in the completed data set. The standardization method is dynamically adjusted according to the data distribution of each parameter to ensure that the data of all parameters are compared and analyzed at the same scale to obtain a multidimensional time series data set.

[0031] Step S2, inputting the multidimensional time series data set into the TSVT-PNNT model to reconstruct the data to obtain reconstructed data;

[0032] Specifically, a multidimensional time series data set is subjected to tensor quantization processing to convert the original data set into a tensor structure suitable for subsequent processing. The tensor quantization processing converts the multidimensional time series data set into a three-order tensor structure, in which the three dimensions represent time, parameter type and process stage, respectively, and describes the complex relationship and multidimensional characteristics in the tagatose milk tea production process, thereby obtaining the initial tensor data. The initial tensor data is subjected to tensor singular value decomposition, which is an effective method for processing high-order data structures. The tensor is represented as a combination of a set of low-order matrices by decomposing it. The process of tensor singular value decomposition is to calculate the left singular matrix, the right singular matrix and the singular value matrix of the initial tensor data to form the decomposed tensor components. Based on the decomposed tensor components, the truncated p-shrinkage norm is used as the convex envelope of the tensor rank to calculate the approximate value of the tensor nuclear norm, and the estimated value of the tensor rank is obtained. According to the estimated value of the tensor rank, the singular value matrix is ​​subjected to soft threshold processing, and the singular values ​​less than the preset threshold are set to zero, so as to effectively remove the noise components in the data and obtain the sparse singular value matrix. The tensor is reconstructed using the sparse singular value matrix and the original left singular matrix and right singular matrix to obtain the preliminary reconstructed tensor data. In order to capture the high-order correlation of the data in the tagatose milk tea production process, the p-norm neural tensor network is used to process the preliminary reconstructed tensor data. Through multi-layer nonlinear transformation, the neural tensor network can effectively capture the complex patterns and high-order features in the data and obtain the optimized tensor data. In this process, the multi-layer structure of the neural tensor network enables the high-order correlation of the data to be extracted and represented layer by layer, so that the final optimized tensor can better reflect the potential complex relationship in the data. Based on the optimized tensor data, the alternating direction multiplier method is used for iterative optimization to minimize the weighted sum of the reconstruction error and the regularization term. The alternating direction multiplier method is an optimization algorithm that gradually reduces the reconstruction error by iteration and adds regularization terms in the reconstruction process to ensure that the reconstructed tensor data is not only highly consistent with the original data, but also has good generalization performance. After multiple iterative optimizations, the final reconstructed tensor data is obtained. The final reconstructed tensor data is subjected to inverse tensor quantization processing, and the tensor data is converted back to the format of a multidimensional time series to obtain the reconstructed data.

[0033] Step S3: performing multi-feature complementary fusion based on the reconstructed data to obtain full-band fusion features;

[0034] Specifically, the reconstructed data is subjected to multi-scale wavelet decomposition, and the db4 wavelet basis function is used to perform a five-layer decomposition to capture the changing characteristics of important parameters in the tagatose milk tea production process from different time scales. For the key parameters in the production process: equipment temperature, stirring speed and solubility, the approximate coefficients and detail coefficients are calculated by wavelet decomposition to obtain the wavelet coefficient matrix. Wavelet decomposition decomposes complex time series signals into components of different frequency bands, so that the system can process and analyze high-frequency and low-frequency information separately to obtain richer feature representations. Based on the wavelet coefficient matrix, the improved maximum absolute value rule is applied to the high-frequency components. Combined with the dissolution kinetic model of tagatose, different frequency band weights are set for each parameter, fully considering the characteristics of each process parameter in different frequency bands and its impact on the quality of tagatose milk tea. In the weighting process, the high-frequency coefficient with the largest absolute value after weighting is selected as the representative feature to form a high-frequency feature vector. Integrated empirical mode decomposition is performed on the reconstructed data. Integrated empirical mode decomposition adds white noise and averages the results of multiple decompositions to obtain multiple intrinsic mode functions and residuals, and then extracts low-frequency signal components. Through multiple iterations, the integrated empirical mode decomposition can extract low-frequency features in the data more stably and reliably, and is suitable for complex production data. The low-frequency signal component contains trend information in the production process, which helps to describe the long-term changes in process parameters and their impact on product quality. The low-frequency signal component is input into the K-SVD sparse representation model. K-SVD is a sparse representation method based on dictionary learning. It captures the main features in the data by designing an over-complete dictionary that adapts to the production characteristics of tagatose milk tea. In the process of dictionary learning, the orthogonal matching pursuit algorithm is used for dictionary learning and sparse coding to obtain low-frequency feature vectors. The orthogonal matching pursuit algorithm is an iterative optimization method that achieves efficient sparse representation of low-frequency components by gradually selecting the dictionary elements that best match the signal, thereby ensuring that the key features in the low-frequency components are retained while reducing the data dimension. The high-frequency feature vectors and low-frequency feature vectors are adaptively normalized to standardize the features of different scales so that the influence of each feature on the model is at the same level in the subsequent fusion process. Considering the dynamic range of each parameter in the production process of tagatose milk tea, the piecewise linear mapping method is used to map the eigenvalue to the [0,1] interval to obtain a standardized eigenvector. The standardized eigenvector is input into the nonlinear feature mapping function, and the fusion and dimension reduction of features are achieved through multi-layer nonlinear transformation and feature reorganization. The nonlinear relationship between features is captured by a deep network structure, the key information related to the quality of tagatose milk tea is retained, and the redundant information is eliminated as much as possible to obtain the preliminary fusion features. The preliminary fusion features are adaptively weighted based on the attention mechanism, and a multi-head self-attention module is constructed to capture the different contributions of each feature to the quality of tagatose milk tea.The attention mechanism dynamically calculates the weight of each feature, allowing the model to pay more attention to features that have an important impact on the quality of the final product and achieve refined weighting of features. The multi-head self-attention module can capture the complex interactions between features in different subspaces and improve the effect of feature fusion. Through adaptive weighting, full-band fusion features are obtained.

[0035] Step S4, inputting the full-band fusion feature structure into a multi-objective integer linear programming model to perform process parameter constraint analysis to obtain a process parameter constraint range;

[0036] Specifically, feature importance analysis is performed on the full-band fusion features to identify the key features that have the greatest impact on the process results and obtain the importance ranking of the features. Based on the feature importance ranking, the top K most important features are selected as decision variables in multi-objective optimization, including key process parameters closely related to the production process such as equipment temperature, stirring speed, solubility and raw material ratio. At the same time, in order to comprehensively consider various factors in the production process, virtual variables are introduced to represent the potential impact of equipment cleanliness on the process, and to enhance the adaptability and description ability of the model to complex production conditions. Multiple optimization objectives are set according to the production process requirements of tagatose milk tea. The production process of tagatose milk tea involves multiple competing objectives, such as how to maximize product quality scores while minimizing energy consumption, production time and raw material loss. A multi-objective function is constructed to describe these objectives. In order to effectively balance the conflicts between objectives, a target weight allocation mechanism based on fuzzy set theory is introduced. Fuzzy set theory makes the weight of each objective adaptively adjusted according to the actual production situation by fuzzifying the target weight, ensuring that the model can reasonably allocate the priorities of different objectives according to the specific scenario during the multi-objective optimization process, making the multi-objective function more flexible and adaptable. The constraint relationship between process parameters is established to ensure that the solution meets the basic requirements of the production process. The constraint relationship covers the nonlinear relationship between tagatose solubility and equipment temperature, the relationship between stirring speed and energy consumption, and the response surface model between raw material ratio and product quality. The relationship between tagatose solubility and equipment temperature is nonlinear, and this nonlinear characteristic needs to be reflected in the model. Similarly, the relationship between stirring speed and energy consumption has a complex dependence. Too fast stirring speed will increase energy consumption, but too slow stirring speed will lead to insufficient dissolution, thus affecting product quality. The response surface model of raw material ratio and product quality reflects the impact of different ratios on the quality of the final product, ensuring that the produced milk tea reaches the expected quality. Since there are many nonlinear relationships in the constraint relationship between process parameters, the nonlinear constraints in the constraint condition set are adaptively piecewise linearized. Considering the solubility characteristics of tagatose, the segmentation density is increased in the key change area to more accurately describe the changes in parameters in these areas. By piecewise approximation of the nonlinear function, it is converted into several linear constraints, and the entire problem is converted into an integer linear programming problem. Discretize continuous variables into several gears so that they can adapt to the solution requirements of integer programming. Input the decision variable vector, multi-objective function and constraint set into the multi-objective integer linear programming solver to perform mixed integer programming solution. Mixed integer programming is an optimization method that processes some variables as continuous variables and some variables as discrete variables. It is suitable for solving optimization problems involving discrete equipment states and continuous process parameters in tagatose milk tea production. Through mixed integer programming, the optimal process parameter settings can be found under the consideration of multiple optimization objectives and constraints, and the constraint range of the process parameters can be obtained.

[0037] Step S5: Based on the process parameter constraint range, a plurality of candidate process parameter adjustment schemes are generated using a Bayesian network and a reinforcement learning method;

[0038] Specifically, based on the constraint range of process parameters and historical production data, the initial Bayesian network model of tagatose milk tea production process is constructed. Bayesian network is a probability-based model that can effectively describe the causal relationship and correlation between variables. In the scenario of tagatose milk tea production process, the Bayesian network constructs a systematic probabilistic graph model by utilizing the interdependence between process parameters. The initial model can outline the main influencing factors in the production process and their relationship with each other. Parameter learning is performed on the initial Bayesian network model to improve its accuracy and adaptability. In the parameter learning process, the expectation maximization algorithm is used to optimize the parameters of the Bayesian network. The expectation maximization algorithm is a classic method for processing probabilistic models with hidden variables. By iteratively performing the expectation step (E step) and the maximization step (M step), the fitting degree of the network model is gradually improved. Combined with the dissolution kinetics model of tagatose, the prior probability of the Bayesian network is adjusted to ensure that the model can accurately reflect the characteristics of the tagatose dissolution process, and the target Bayesian network model is obtained. Based on the target Bayesian network model, the state space, action space and reward function in reinforcement learning are defined. In the process optimization scenario of tagatose milk tea, the state space includes the current process parameters and quality indicators. These states describe the current state of the production process, while the action space is the adjustment amount of the process parameters, which is used to change the process parameters in the production process to achieve better results. At the same time, the reward function is constructed based on the quality score and production efficiency of the product, which is used to evaluate the results after each action. The design of the reward function determines the learning direction of the reinforcement learning agent-by setting appropriate rewards, the agent can learn how to improve production efficiency without reducing quality, and finally generate a better process parameter adjustment plan. In order to improve the effect of reinforcement learning in the optimization of tagatose milk tea production process, the deep Q network (DQN) algorithm is used to construct the reinforcement learning agent. The deep Q network is an algorithm that combines deep learning and reinforcement learning, which can perform effective strategy learning in complex high-dimensional state space. The target Bayesian network model is embedded into the network structure of DQN as prior knowledge to obtain the DQN model with Bayesian prior, which enables DQN to make full use of the causal relationship and statistical characteristics of historical data contained in the Bayesian network during the learning process, accelerate the convergence speed of reinforcement learning, and improve the effectiveness of the final generated solution. The introduction of Bayesian prior enables the intelligent agent to have stronger causal reasoning ability in the decision-making process, so that the generated process parameter adjustment plan is more in line with the actual production situation and target requirements. When using the DQN model with Bayesian prior for reinforcement learning, the ε-greedy strategy is used for action selection. The ε-greedy strategy is an exploration and utilization strategy that achieves a balance between comprehensive exploration of the environment and local optimization by selecting random actions (exploration) with a certain probability and selecting the optimal action (utilization) with a certain probability.At the same time, the Boltzmann exploration mechanism is introduced to make probabilistic choices based on the Q-value distribution of actions, so that when faced with multiple actions with similar Q-values, the reinforcement learning agent can make choices in a more reasonable way and better balance the process of exploration and utilization. In the balance between exploration and utilization, the introduction of the Boltzmann mechanism enables the agent to try more different combinations of process parameters in the early stage, and gradually tend to choose those better adjustment plans in the later stage, thereby ensuring the diversity and quality of the generated plans. By integrating the DQN model with Bayesian priors, and combining the ε-greedy strategy and the Boltzmann exploration mechanism, multiple candidate process parameter adjustment plans are generated.

[0039] Step S6: verify and evaluate multiple candidate process parameter adjustment schemes in various production scenarios, select the best candidate process parameter adjustment scheme as the target process parameter adjustment scheme and output it.

[0040] Specifically, based on historical production data and expert knowledge, a multi-scenario simulation model for tagatose milk tea production is constructed. The scenarios cover a variety of production conditions such as normal production, high temperature environment, and slight equipment pollution, so as to evaluate the performance of candidate process parameters under different production conditions. In order to build a multi-scenario simulation environment, the Latin hypercube sampling method is used to generate scenario parameter distribution, which effectively covers various situations that may be encountered in the production process. At the same time, combined with the dissolution kinetics model and thermodynamic model of tagatose, a simulation environment that can realistically simulate different production scenarios is constructed, so that the simulation can accurately reflect the impact of changes in process parameters in the actual production process on product quality and production efficiency. Multiple candidate process parameter adjustment schemes are input into the multi-scenario simulation environment for parallel simulation experiments. In order to accelerate the simulation process, the Kriging proxy model is used. Kriging is an effective meta-model construction method that can build a high-precision proxy model based on fewer simulation experiments, significantly reducing the computing time and resource costs required for simulation. Through the acceleration of the Kriging proxy model, the performance indicator data of each candidate process parameter adjustment scheme in different scenarios can be quickly obtained, including key parameters such as product quality score, production efficiency, and energy consumption, which fully reflects the performance of each candidate scheme under various production conditions. The performance index data is evaluated in multiple dimensions, and an evaluation index system based on fuzzy analytic hierarchy process is constructed. Through fuzzy analytic hierarchy process, the performance indexes of multiple different dimensions are comprehensively analyzed, and the initial comprehensive scores of each candidate process parameter adjustment scheme in different scenarios are calculated. Based on the initial comprehensive score, the relative proximity of each candidate process parameter adjustment scheme to the ideal solution is calculated. In order to dynamically adjust the weight of each index in different scenarios, the coefficient of variation method is introduced. Through the coefficient of variation method, the weight of each index is dynamically adjusted according to the degree of change of each index in different scenarios, and a weighted comprehensive score that is more in line with the actual situation is obtained. A global sensitivity analysis is performed on the weighted comprehensive score to obtain the sensitivity analysis results, and the sensitivity of the performance of each candidate scheme to different factors and the interaction effect between different factors are determined. Based on the sensitivity analysis results, the robustness of each candidate process parameter adjustment scheme is evaluated, and the stability and risk indicators of each scheme in dealing with uncertainty and fluctuations in the production process are obtained. The weighted comprehensive score and risk assessment index of each candidate process parameter adjustment scheme are input into the multi-attribute decision-making model to rank the schemes. The multi-attribute decision-making model comprehensively considers the performance score and risk index of each solution, and can determine the priority of each solution through weighting and sorting. According to the solution sorting results, the candidate process parameter adjustment solution with the best comprehensive performance and controllable risk is selected as the final target process parameter adjustment solution and output.

[0041] In the embodiment of the present invention, by collecting and preprocessing multidimensional parameter data and combining the TSVT-PNNT model for data reconstruction, the data quality is effectively improved, the influence of noise and missing values ​​on subsequent analysis is reduced, and the multi-feature complementary fusion technology is adopted to comprehensively utilize high-frequency and low-frequency feature information to comprehensively capture various dynamic characteristics in the tagatose milk tea production process, thereby improving the comprehensiveness and accuracy of the process evaluation. The multi-objective integer linear programming model is introduced to perform process parameter constraint analysis, and multiple production objectives and constraints are considered, which helps to optimize energy consumption and production efficiency while ensuring product quality. The candidate process parameter adjustment scheme is generated by combining the Bayesian network and the reinforcement learning method, which makes full use of historical data and expert knowledge, and improves the feasibility and adaptability of the scheme. The candidate schemes are verified and evaluated in a variety of production scenarios, and the simulation and decision-making methods are adopted to enhance the robustness of the process evaluation and ensure the performance of the selected scheme under different production environments. The present invention can adaptively generate optimized process parameter adjustment schemes according to different production scenarios to improve the production efficiency and product quality stability of tagatose milk tea.

[0042] In a specific embodiment, the process of executing step S1 may specifically include the following steps:

[0043] Install high-precision sensors at the key process nodes of raw material ratio, dissolution, heating, cooling, blending and packaging of tagatose milk tea production line to collect multi-dimensional parameter data of equipment temperature, stirring speed, solubility, raw material ratio, equipment cleanliness, tea extraction temperature and tea extraction time, and transmit the collected data to the central processing unit through industrial Internet of Things to obtain the original data set;

[0044] The original data set is divided into time series to obtain initial time series data, and outliers in the initial time series data are identified. A dynamic threshold is set in combination with the solubility characteristics of tagatose, and data points with deviations exceeding the threshold are marked as outliers to obtain outlier marked data;

[0045] Based on the outlier labeled data, local outlier processing is performed on the data points marked as outliers to obtain a data set without outliers, and multiple interpolation is performed on the missing values ​​in the data set without outliers to obtain a completed data set;

[0046] The parameter data in the completed data set are adaptively standardized to obtain a multidimensional time series data set.

[0047] Specifically, high-precision sensors are installed at the key process nodes of raw material ratio, dissolution, heating, cooling, blending and packaging of tagatose milk tea production line to collect multi-dimensional parameter data, including equipment temperature, stirring speed, solubility, raw material ratio and equipment cleanliness, tea extraction temperature and tea extraction time. The equipment temperature sensor is used to monitor the equipment temperature changes in real time during the dissolution, heating and cooling of raw materials, the stirring speed sensor records the rate during the stirring process, the solubility sensor monitors the dissolution of tagatose in the liquid, the raw material ratio sensor ensures that each raw material is added in the prescribed proportion, and the equipment cleanliness sensor is used to monitor the cleanliness of the equipment to ensure that it meets the hygiene requirements. The data of multi-dimensional parameters are transmitted to the central processing unit through the industrial Internet of Things to form an original data set. The original data set is divided into time series to obtain the initial time series data. The multi-dimensional data at each time point in the production process are organized into a time series form, so that these data clearly reflect the dynamic changes of process parameters over time in the production process. The outliers in the initial time series data are identified. The dynamic threshold is set in combination with the solubility characteristics of tagatose to make reasonable judgments on abnormal changes in the data. The setting of dynamic threshold is based on the solubility characteristics of tagatose under different equipment temperature and stirring speed conditions. For example, assuming that the solubility of tagatose is positively correlated with the equipment temperature and there is a saturated solubility curve, the dynamic threshold is set by formula to determine whether there are abnormal values. The specific formula for setting the dynamic threshold is:

[0048] ;

[0049] in, Indicates the device temperature The permissible solubility deviation is is a proportional coefficient used to adjust the allowed fluctuation range. Tagatose at the equipment temperature Theoretical solubility under this condition. Combined with this formula, for deviations exceeding The data points are marked as outliers to form outlier marked data. Based on the outlier marked data, the outlier data points are processed locally to obtain a data set without outliers. The local outliers are smoothed by using interpolation methods, such as linear interpolation or polynomial interpolation of outliers using data from adjacent time points. Suppose the outlier value at a certain moment is , through its two normal values ​​before and after and Perform linear interpolation:

[0050] ;

[0051] Through interpolation processing, the influence of outliers on the overall trend of the data is removed to form a data set without outliers. Multiple interpolation is performed on the missing values ​​in the data set without outliers to obtain a complete data set. Multiple interpolation is a method based on statistics and machine learning. It repeatedly estimates the missing parts to obtain reasonable filling values ​​to ensure the integrity and continuity of the data and obtain a completed data set. The completed data set is adaptively standardized. Data of different dimensions are mapped to the same scale so that the model will not be affected by dimensional differences when processing these data. All parameter data are standardized to a distribution range with a mean of 0 and a standard deviation of 1 to eliminate the influence of different dimensions between different parameters and obtain a multidimensional time series data set.

[0052] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0053] Performing tensorization processing on the multidimensional time series data set, converting the multidimensional time series data set into a three-order tensor structure, where the three dimensions represent time, parameter type and process stage, respectively, to obtain initial tensor data;

[0054] Perform tensor singular value decomposition on the initial tensor data, calculate the left singular matrix, the right singular matrix and the singular value matrix, and form the decomposed tensor components;

[0055] Based on the decomposed tensor components, the truncated p-shrinkage norm is used as the convex envelope of the tensor rank, and the approximate value of the tensor nuclear norm is calculated to obtain the estimated value of the tensor rank;

[0056] According to the estimated value of the tensor rank, the singular value matrix is ​​soft-thresholded, singular values ​​less than the preset threshold are set to zero, and a sparse singular value matrix is ​​obtained. The sparse singular value matrix and the original left singular matrix and right singular matrix are used to reconstruct the tensor to obtain preliminary reconstructed tensor data;

[0057] Applying the p-norm neural tensor network to the initially reconstructed tensor data, capturing the high-order correlation of the data through multi-layer nonlinear transformation, and obtaining the optimized tensor data, and based on the optimized tensor data, using the alternating direction multiplier method for iterative optimization, minimizing the weighted sum of the reconstruction error and the regularization term, and obtaining the final reconstructed tensor data;

[0058] The final reconstructed tensor data is subjected to inverse tensor quantization and converted back to a multidimensional time series format to obtain the reconstructed data.

[0059] Specifically, the multidimensional time series data set is tensorized to convert the original multidimensional time series data set into a three-order tensor structure. A tensor is a high-order array that can effectively represent high-dimensional data. In the process production of tagatose milk tea, the multidimensional time series data set contains multiple parameters that change over time, such as equipment temperature, stirring speed, solubility, etc. These parameters have their own change characteristics in different process stages. The multidimensional time series data set is converted into a three-order tensor, in which the three dimensions represent time, parameter type, and process stage respectively. For example, assuming that the time series data has time points, involving different parameter types, and process stages, the obtained tensor structure is expressed as , that is, the initial tensor data, can fully describe the multi-dimensional dynamic characteristics of the production process. The initial tensor data is subjected to tensor singular value decomposition to decompose the original tensor into several low-order tensors to reduce data redundancy and facilitate subsequent processing. Decomposed into left singular matrix, right singular matrix and singular value matrix, specifically expressed as:

[0060] ;

[0061] in, is the original third-order tensor, is a left singular matrix, which represents the correlation between the time dimension and the process stage; is a singular value matrix whose elements represent the "strength" or "weight" of each mode of the tensor, is a right singular matrix, which represents the relationship between process stages and parameter types. Based on the decomposed tensor components, truncation is adopted The norm is used as the convex envelope of the tensor rank, and the approximate value of the tensor nuclear norm is calculated to obtain the estimated value of the tensor rank. The approximate value of the tensor nuclear norm is used to evaluate the amount of effective information in the tensor, truncating The norm is defined as:

[0062] ;

[0063] in, is the singular value matrix The singular values, is the cutoff threshold, Represents the truncated nuclear norm. The rank of the tensor is estimated through this norm, and the singular values ​​are found to be the main contributors, so as to screen and sparse the singular value matrix. According to the estimated value of the tensor rank, the singular value matrix is ​​soft-thresholded, and the singular values ​​less than the preset threshold are set to zero to obtain a sparse singular value matrix. Let the singular value matrix A singular value in , the process of soft threshold processing is expressed as:

[0064] ;

[0065] in, is the preset threshold, The symbol function representing the singular value. Through soft threshold processing, the noise component is effectively removed, making the singular value matrix more sparse. Using the sparse singular value matrix and the original left singular matrix , right singular matrix , perform tensor reconstruction, and obtain preliminary reconstructed tensor data. Apply the preliminary reconstructed tensor data Norm neural tensor network, which captures high-order correlations in data through multiple layers of nonlinear transformations. The norm neural tensor network is a deep learning method that can process the input tensor through a multi-layer neural network to mine the implicit patterns and high-order dependencies. In this way, an optimized tensor data is obtained. Based on the optimized tensor data, the alternating direction multiplier method is used for iterative optimization to minimize the weighted sum of the reconstruction error and the regularization term to obtain the final reconstructed tensor data. In the alternating direction multiplier method, by decomposing the optimization problem into multiple sub-problems and optimizing them successively, the complexity of the solution can be effectively reduced while ensuring that the reconstruction error is minimized. The final reconstructed tensor data is inversely quantized to convert the tensor data back to the format of a multidimensional time series. The third-order tensor is remapped back to the original multidimensional structure of time, parameter type, and process stage to facilitate application and interpretation in the actual production process. Through inverse tensor quantization, a reconstructed multidimensional time series data set is obtained.

[0066] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0067] The reconstructed data is subjected to multi-scale wavelet decomposition, and the db4 wavelet basis function is used to perform a 5-layer decomposition. The approximate coefficient and detail coefficient are calculated for the equipment temperature, stirring speed, and solubility parameter, respectively, to obtain the wavelet coefficient matrix.

[0068] Based on the wavelet coefficient matrix, the improved maximum absolute value rule is applied to the high-frequency components. Combined with the tagatose dissolution kinetic model, different frequency band weights are set for each parameter, and the high-frequency coefficient with the largest absolute value after weighting is selected to form a high-frequency feature vector.

[0069] Perform integrated empirical mode decomposition on the reconstructed data, extract multiple intrinsic mode functions and residuals by adding white noise and calculating the average of multiple decompositions to obtain low-frequency signal components;

[0070] The low-frequency signal components are input into the K-SVD sparse representation model, an overcomplete dictionary adapted to the production characteristics of tagatose milk tea is designed, and the orthogonal matching pursuit algorithm is used for dictionary learning and sparse coding to obtain the low-frequency feature vector.

[0071] Adaptively normalize the high-frequency eigenvectors and low-frequency eigenvectors. According to the dynamic range of each parameter in the tagatose milk tea production process, the eigenvalues ​​are mapped to the [0,1] interval using a piecewise linear mapping method to obtain a standardized eigenvector.

[0072] The standardized feature vector is input into the nonlinear feature mapping function, and feature fusion and dimension reduction are performed through multi-layer nonlinear transformation and feature reorganization, while retaining the key information related to the quality of tagatose milk tea to obtain preliminary fusion features;

[0073] The preliminary fusion features are adaptively weighted based on the attention mechanism, and a multi-head self-attention module is constructed. The attention weight is dynamically calculated according to the contribution of each feature to the quality of tagatose milk tea to obtain the full-band fusion features.

[0074] Specifically, multi-scale wavelet decomposition is performed on the reconstructed data, and the db4 wavelet basis function is selected and five-layer decomposition is performed to separate the high-frequency and low-frequency information in the data. In the production process of tagatose milk tea, different parameters such as equipment temperature, stirring speed, solubility, etc. have different effects on product quality, and wavelet decomposition is an effective method to decompose signals into different scales and frequency bands. Through 5-layer decomposition, the detailed change characteristics and long-term trends of each parameter at different frequencies are obtained. For each parameter, the approximate coefficient (low-frequency information) and the detail coefficient (high-frequency information) are calculated separately to obtain a complete wavelet coefficient matrix. , which contains information describing each parameter in different frequency bands. Based on the wavelet coefficient matrix, an improved maximum absolute value rule is applied to the high-frequency component in order to extract the most important features for product quality from the high-frequency information. In the improved maximum absolute value rule, different frequency band weights are set for each parameter in combination with the dissolution kinetic model of tagatose to ensure that their physical properties are fully considered when evaluating the effects of different frequency bands on solubility, equipment temperature or stirring speed. Let the high-frequency component be , whose band weight is , the weighted value of each high-frequency coefficient is expressed as:

[0075] ;

[0076] in, represents the weighted high-frequency coefficient, is the original high frequency component, is the frequency band weight. The high-frequency coefficient with the largest absolute value after weighting is selected to form the high-frequency feature vector , which contains the most important high-frequency features of the different frequency bands that affect the quality of tagatose milk tea. At the same time, for the reconstructed data, integrated empirical mode decomposition is performed. By adding white noise and calculating the average of multiple decompositions, multiple intrinsic mode functions (IMFs) and residuals are extracted. These IMFs and residuals describe the low-frequency components in the signal. Let IMF be , then the low-frequency signal component after multiple decompositions is expressed as:

[0077] ;

[0078] in, represents the final low-frequency signal, is the number of decompositions, For the The intrinsic mode function obtained by the decomposition. The low-frequency signal component is input into the K-SVD sparse representation model. K-SVD is a dictionary learning method used to construct a dictionary suitable for a specific application from data. An over-complete dictionary adapted to the production characteristics of tagatose milk tea is designed to describe the characteristics of low-frequency signals in the production process. In the dictionary learning process, the orthogonal matching pursuit algorithm (OMP) is used for dictionary learning and sparse coding to obtain the low-frequency feature vector . The OMP algorithm constructs a sparse representation by gradually selecting a part of the dictionary elements that best explain the signal, so that the low-frequency feature vector retains the main information of the signal while reducing redundancy. In order to ensure the scale consistency of high-frequency feature vectors and low-frequency feature vectors, they are adaptively normalized and data of different dimensions are mapped to the same interval, so that different features are comparable in subsequent analysis. In the production process of tagatose milk tea, the dynamic range of each parameter is different, and the piecewise linear mapping method is used to map the eigenvalue to the interval [0,1]. Assuming that the original eigenvalue is , and its mapped value is , the normalized formula is expressed as:

[0079] ;

[0080] in, and are the minimum and maximum eigenvalues ​​respectively. Through this normalization formula, the standardized high-frequency eigenvector and low-frequency eigenvector are obtained, which are recorded as and . The standardized feature vector is input into the nonlinear feature mapping function, and feature fusion and dimensionality reduction are performed through multi-layer nonlinear transformation and feature reorganization. A deep neural network is used to map the input feature vector through multi-layer nonlinear activation functions, and feature reorganization is performed to capture the complex nonlinear relationship between high-frequency and low-frequency features to obtain preliminary fusion features. The preliminary fusion features are adaptively weighted based on the attention mechanism. A multi-head self-attention module is constructed, and the multi-head attention mechanism is used to capture the complex interactive relationship between features in different subspaces, and the attention weight is dynamically calculated according to the contribution of each feature to the quality of tagatose milk tea. In this way, different attention can be allocated to different features to form a full-band fusion feature representation.

[0081] In a specific embodiment, the process of executing step S4 may specifically include the following steps:

[0082] The feature importance analysis of the full-band fusion features was performed to obtain the feature importance ranking. Based on the feature importance ranking, the top K most important features were selected as decision variables to construct a decision variable vector, including key process parameters such as equipment temperature, stirring speed, solubility, and raw material ratio. At the same time, virtual variables were introduced to represent the impact of equipment cleanliness.

[0083] According to the production process requirements of tagatose milk tea, multiple optimization goals were set, including maximizing product quality score, minimizing energy consumption, minimizing production time and minimizing raw material loss, constructing a multi-objective function, and introducing a target weight allocation mechanism based on fuzzy set theory;

[0084] Establish the constraint relationship between process parameters, including the nonlinear relationship between tagatose solubility and equipment temperature, the relationship between stirring speed and energy consumption, and the response surface model between raw material ratio and product quality, to form a set of constraint conditions;

[0085] The nonlinear constraints in the constraint set are processed by adaptive piecewise linearization. According to the solubility characteristics of tagatose, the segment density is increased in the key change area, and the continuous variables are discretized into several gears, which are converted into integer linear programming problems.

[0086] The decision variable vector, multi-objective function and constraint condition set are input into the multi-objective integer linear programming solver to perform mixed integer programming and obtain the constraint range of process parameters.

[0087] Specifically, the importance of the full-band fusion features is analyzed. Through methods such as random forest or gradient boosting tree, the importance contribution of each feature in the model to the final prediction result is calculated to obtain the importance score of each feature, which reflects the role of each feature in describing and predicting the production results. The importance ranking of the features is obtained by sorting the scores. Based on this ranking, the top The most important features are used as decision variables, including key process parameters closely related to the production process, such as equipment temperature, stirring speed, solubility and raw material ratio. At the same time, in order to fully reflect the complex factors in the production process, a virtual variable is introduced to represent the impact of equipment cleanliness to ensure that the potential impact of equipment status on the final production results can also be included in the model for optimization. According to the specific requirements of the tagatose milk tea production process, multiple optimization goals are set, including maximizing product quality score, minimizing energy consumption, minimizing production time and minimizing raw material loss. Construct a multi-objective optimization function, expressed by the following formula:

[0088] ;

[0089] in, It is a multi-objective optimization function; represents the product quality score, which needs to be maximized; It represents energy consumption, which needs to be minimized; It indicates the production time, which needs to be minimized; It indicates the loss of raw materials, which needs to be minimized; are the weights of each objective, which are used to measure the importance of different objectives. In order to reasonably allocate these weights, a target weight allocation mechanism based on fuzzy set theory is introduced. Fuzzy set theory can adaptively allocate weights to each optimization objective according to the fuzzy membership between different objectives, thereby ensuring that each objective can receive proper attention in the overall optimization. Model the constraint relationship between production process parameters. There are complex interactions between process parameters in the production of tagatose milk tea. For example, the solubility of tagatose has a nonlinear relationship with the equipment temperature, which is described by the following equation:

[0090] ;

[0091] in, Indicates the solubility of tagatose, Indicates the device temperature. and The parameters are obtained by empirical fitting. The equation reflects the effect of equipment temperature change on solubility. At the same time, there is a direct correlation between stirring speed and energy consumption. Too fast stirring speed will lead to a significant increase in energy consumption, while too slow stirring speed will lead to insufficient dissolution. The relationship between stirring speed and energy consumption is described by a linear function, for example:

[0092] ;

[0093] in, Indicates energy consumption, Indicates the stirring speed, and is the linear coefficient. Similarly, a response surface model between the raw material ratio and the product quality is established to determine the specific impact on the product quality under different ratios. These relationships together constitute the constraint condition set of the entire process parameters. The nonlinear constraints in the constraint condition set are adaptively piecewise linearized. According to the solubility characteristics of tagatose, the segmentation density is increased in the key change area of ​​the nonlinear function, so as to approximate the nonlinear function piecewise, and the continuous nonlinear variables are discretized into several gears, so as to convert the entire problem into an integer linear programming problem. For example, for the nonlinear relationship between equipment temperature and solubility, the equipment temperature range is divided into multiple discrete intervals, and a linear function is used to approximate the original nonlinear curve in each interval, which can effectively reduce the complexity of the solution and enable the solver to converge quickly. The decision variable vector, multi-objective function and constraint condition set are input into the multi-objective integer linear programming solver for mixed integer programming solution. Mixed integer programming is an optimization method that contains both integer variables and continuous variables. It is suitable for dealing with complex optimization problems involving discrete equipment states and continuous process parameters in tagatose milk tea production. Through mixed integer programming, the optimal process parameter settings are found on the basis of satisfying all constraints, and the constraint range of the process parameters is obtained.

[0094] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0095] Based on the process parameter constraints and historical production data, an initial Bayesian network model of tagatose milk tea production process was constructed;

[0096] The initial Bayesian network model was trained with parameters, the expectation maximization algorithm was used to optimize the network parameters, and the prior probability was adjusted in combination with the tagatose dissolution kinetics model to obtain the target Bayesian network model.

[0097] Based on the target Bayesian network model, the state space, action space and reward function of reinforcement learning are defined. The state space includes the current process parameters and quality indicators, the action space is the adjustment amount of the process parameters, and the reward function is constructed based on the product quality score and production efficiency.

[0098] The deep Q network algorithm is used to build a reinforcement learning agent, and the target Bayesian network model is embedded into the DQN network structure as prior knowledge to obtain a DQN model that integrates Bayesian priors.

[0099] Based on the DQN model integrating Bayesian prior, an ε-greedy strategy is designed for action selection. Meanwhile, the Boltzmann exploration mechanism is introduced to balance exploration and utilization, and multiple candidate process parameter adjustment schemes are generated.

[0100] Specifically, based on the constraint range of process parameters and historical production data, an initial Bayesian network model of tagatose milk tea production process is constructed. Bayesian network is a probabilistic graphical model that can reflect the causal relationship between variables. It describes various process parameters and their dependencies through a directed acyclic graph. In the production of tagatose milk tea, different process parameters such as equipment temperature, stirring speed, solubility and raw material ratio interact with each other and have a direct or indirect impact on the final product quality. By inputting historical production data into the Bayesian network, an initial model is obtained to characterize the relationship between different parameters in the production process. Parameter learning is performed on the initial Bayesian network model to improve the accuracy and reliability of the model. The expectation maximization (EM) algorithm is used to optimize the parameters of the Bayesian network. The EM algorithm is a method for processing probabilistic models with hidden variables. The parameters in the model are estimated in an iterative manner so that the model can better fit the observed data. In this process, the EM algorithm is divided into two steps: the expectation step (E step) and the maximization step (M step). In the E step, the expected value of each latent variable is calculated, and then in the M step, the model parameters are updated by maximizing the log-likelihood function. The goal of the EM algorithm is to maximize the likelihood of the Bayesian network model for historical production data, so that the model is closer to the actual production situation. In order to better describe the solubility characteristics of tagatose, the prior probability in the Bayesian network is adjusted in combination with the tagatose dissolution kinetic model. For example, the solubility of tagatose Expressed as time Function, combined with the dissolution kinetics equation:

[0101] ;

[0102] in, is the saturated solubility of tagatose, is the dissolution rate constant, is time. By integrating the model into the Bayesian network and setting the prior probability of parameters such as equipment temperature and stirring speed for solubility, a more accurate target Bayesian network model is obtained. Based on the target Bayesian network model, the basic framework of reinforcement learning is defined, which includes state space, action space and reward function. The state space of reinforcement learning contains current process parameters and product quality indicators, which are used to describe the state of the production system at a certain point in time, such as the current values ​​of variables such as equipment temperature, stirring speed, solubility, and product quality scores. The action space is a collection of adjustments to process parameters, such as adjusting the equipment temperature by 2 degrees or decreasing it by 1 degree, which is used to make the production process develop in the desired direction. The reward function is the core part of reinforcement learning. It determines the reward obtained by the system after each action and is used to guide the learning of the intelligent agent. In the production of tagatose milk tea, the reward function is based on the product quality score. and production efficiency The goal is to maximize product quality and production efficiency. The reward function is defined as follows:

[0103] ;

[0104] in, Indicates in status Next action The rewards received, is the new state after executing the action, and are the weights of quality score and production efficiency, To control the penalty coefficient of cost, Indicates action The energy consumption or resource consumption caused by this. The reward function enables the reinforcement learning agent to learn how to improve production efficiency and reduce resource consumption without reducing product quality. In order to achieve optimal control of the process, the deep Q network (DQN) algorithm is used to construct a reinforcement learning agent. DQN is a method that combines deep learning and reinforcement learning for policy learning in high-dimensional state space. The target Bayesian network model is embedded into the network structure of DQN as prior knowledge to obtain a DQN model that integrates Bayesian priors. By embedding Bayesian prior knowledge in DQN, the agent can utilize the causal relationship and parameter dependency contained in the Bayesian network in the initial learning stage to improve learning efficiency and accelerate convergence. The Bayesian network provides prior knowledge of the dependency relationship between various process parameters. This information is embedded in the network weights of DQN, giving the agent better guidance when exploring the environment. When learning based on the DQN model that integrates Bayesian priors, the The greedy strategy performs action selection. The greedy strategy is an exploration and exploitation balance strategy, in which the agent Choose random actions (exploration) with probability Select the current optimal action (utilization) to effectively balance the contradiction between exploring new strategies and utilizing the current best strategy. In order to improve the efficiency and diversity of exploration, the Boltzmann exploration mechanism is introduced. The Boltzmann mechanism selects the Q value of the action probabilistically, so that a better action selection strategy can be achieved at different exploration stages. The action selection probability under the Boltzmann strategy is expressed as:

[0105] ;

[0106] in, Indicates in status Next select action The probability of For Action In Status The Q value under is the device temperature parameter, which is used to control the randomness of exploration. When it is high, the selection probability of all actions tends to be equal, allowing more exploration; when When it is lower, the selection probability is more concentrated on actions with higher Q values ​​to make more use of them. By integrating the DQN model with Bayesian prior and combining the ε-greedy strategy and Boltzmann exploration mechanism, the agent can select appropriate process parameter adjustment strategies under different production conditions and generate multiple candidate process parameter adjustment schemes.

[0107] In a specific embodiment, the process of executing step S6 may specifically include the following steps:

[0108] Based on historical production data and expert knowledge, a multi-scenario simulation model for tagatose milk tea production was constructed, including scenarios of normal production, high temperature environment, and slight equipment pollution. Latin hypercube sampling was used to generate scenario parameter distribution, and a multi-scenario simulation environment was obtained by combining the tagatose dissolution kinetic model and thermodynamic model.

[0109] Multiple candidate process parameter adjustment schemes are input into a multi-scenario simulation environment for parallel simulation experiments. The Kriging proxy model is used to accelerate the simulation process to obtain the performance index data of each candidate process parameter adjustment scheme in different scenarios.

[0110] Conduct multi-dimensional evaluation of performance indicator data, build an evaluation indicator system based on fuzzy analytic hierarchy process, and calculate the initial comprehensive score of each candidate process parameter adjustment scheme in different scenarios;

[0111] Based on the initial comprehensive score, the relative closeness between each candidate process parameter adjustment scheme and the ideal solution is calculated, and the coefficient of variation method is introduced to dynamically adjust the indicator weight to obtain the weighted comprehensive score of each candidate process parameter adjustment scheme;

[0112] Perform global sensitivity analysis on the weighted comprehensive score to obtain sensitivity analysis results, which are used to represent the sensitivity index and interaction effect of the scheme performance to each factor;

[0113] Based on the sensitivity analysis results, the robustness of each candidate process parameter adjustment scheme is evaluated to obtain the risk assessment index of each candidate process parameter adjustment scheme;

[0114] The weighted comprehensive score and risk assessment index are input into the multi-attribute decision-making model to sort the schemes and obtain the scheme sorting results. Based on the scheme sorting results, the candidate process parameter adjustment scheme with the best comprehensive performance and controllable risks is selected as the target process parameter adjustment scheme and output.

[0115] Specifically, based on historical production data and expert knowledge, a multi-scenario simulation model for tagatose milk tea production is constructed, covering multiple scenarios such as normal production, high temperature environment, and slight equipment contamination, to ensure reasonable process optimization under various production conditions. In order to generate a variety of possible production scenarios, the Latin hypercube sampling method is adopted. Latin hypercube sampling can perform uniform sampling in high-dimensional space to generate representative scenario parameter distributions. Combined with the dissolution kinetic model and thermodynamic model of tagatose, the dissolution behavior and energy changes of tagatose under different production conditions are described to obtain a complete multi-scenario optimal environment. For example, the dissolution behavior of tagatose is described by the following formula:

[0116] ;

[0117] in, Indicates the device temperature The solubility of tagatose under is the prefactor, is a parameter related to dissolution characteristics, is the activation energy, is the gas constant, is the absolute equipment temperature. The dissolution kinetics model is used to simulate the dissolution process under different equipment temperatures and stirring conditions for multi-scenario simulation analysis. Multiple candidate process parameter adjustment schemes are input into the simulation environment for parallel simulation experiments to evaluate the performance of different process parameter schemes under various production conditions. Since the simulation process involves complex physical and chemical reactions and has a high computational cost, the Kriging proxy model is used to accelerate the simulation process. Kriging is an effective meta-modeling method. By fitting a small amount of simulation experimental data, the system response is quickly estimated without actual simulation to obtain the performance indicator data of each candidate process parameter adjustment scheme in different scenarios, such as product quality score, energy consumption, production time, etc. The performance indicator data is evaluated in multiple dimensions. An evaluation index system based on fuzzy hierarchical analysis method is constructed. Fuzzy hierarchical analysis method is an evaluation method that combines subjective judgment with objective data. Through fuzzy membership function, the qualitative judgment of each performance indicator is converted into a quantitative score, and the initial comprehensive score of each candidate process parameter adjustment scheme in different scenarios is calculated, which represents the overall performance of each candidate scheme in each scenario. Based on the initial comprehensive score, the relative closeness of each candidate process parameter adjustment scheme to the ideal solution is calculated to determine the relative advantages and disadvantages of each scheme. In this process, the coefficient of variation method is introduced to dynamically adjust the weight of each indicator. The coefficient of variation method is a weight allocation method based on the degree of variation of indicator data. By calculating the coefficient of variation of each performance indicator, the weight is dynamically adjusted according to its importance. Suppose the standard deviation of a certain performance indicator is , the mean is , then the coefficient of variation It is expressed as:

[0118] ;

[0119] By coefficient of variation Determine the weight of each indicator to ensure that those indicators with a larger degree of variation receive a larger weight in the final score, and obtain the weighted comprehensive score of each candidate process parameter adjustment scheme. The weighted comprehensive score is used for global sensitivity analysis to determine the sensitivity index of different factors to the overall performance and the interaction effect between these factors. Through global sensitivity analysis, the sensitivity index of each process parameter is obtained to measure the robustness of the scheme. Based on the results of sensitivity analysis, the robustness of each candidate process parameter adjustment scheme is evaluated to evaluate the stability and risk of each scheme in different production scenarios, and to evaluate whether it can maintain high product quality and production efficiency when dealing with uncertainties in the production process (such as equipment temperature fluctuations, slight equipment contamination, etc.). The robustness evaluation obtains the risk assessment index of each candidate scheme, which is used to describe the reliability of the scheme in the face of uncertainty. The weighted comprehensive score and risk assessment index of each candidate process parameter adjustment scheme are input into the multi-attribute decision model to rank the schemes. The multi-attribute decision model comprehensively evaluates each scheme by comprehensively considering the weighted comprehensive score and risk assessment index to determine the priority of each scheme. Based on the scheme ranking results, the candidate process parameter adjustment scheme with the best comprehensive performance and controllable risks is selected as the final target process parameter adjustment scheme and output.

[0120] The tagatose milk tea processing technology evaluation method in the embodiment of the present invention is described above. The tagatose milk tea processing technology evaluation system in the embodiment of the present invention is described below. Figure 2 In one embodiment of the present invention, a tagatose milk tea processing technology evaluation system includes:

[0121] The acquisition module is used to collect and preprocess the multidimensional parameter data in the tagatose milk tea production process to obtain a multidimensional time series data set;

[0122] The reconstruction module is used to input the multidimensional time series data set into the TSVT-PNNT model for data reconstruction to obtain the reconstructed data;

[0123] The fusion module is used to perform complementary fusion of multiple features based on the reconstructed data to obtain full-band fusion features;

[0124] An analysis module is used to input the full-band fusion feature structure into a multi-objective integer linear programming model to perform process parameter constraint analysis and obtain the process parameter constraint range;

[0125] A generation module, for generating multiple candidate process parameter adjustment schemes based on the process parameter constraint range by using a Bayesian network and a reinforcement learning method;

[0126] The evaluation module is used to verify and evaluate multiple candidate process parameter adjustment schemes in various production scenarios, select the best candidate process parameter adjustment scheme as the target process parameter adjustment scheme and output it.

[0127] Through the collaborative cooperation of the above-mentioned components, through the collection and preprocessing of multidimensional parameter data, combined with the TSVT-PNNT model for data reconstruction, the data quality is effectively improved, the impact of noise and missing values ​​on subsequent analysis is reduced, and the multi-feature complementary fusion technology is adopted to comprehensively utilize high-frequency and low-frequency feature information to fully capture various dynamic characteristics in the tagatose milk tea production process, and improve the comprehensiveness and accuracy of process evaluation. The multi-objective integer linear programming model is introduced to perform process parameter constraint analysis, taking into account multiple production goals and constraints, which helps to optimize energy consumption and production efficiency while ensuring product quality. Combining Bayesian networks and reinforcement learning methods to generate candidate process parameter adjustment schemes, historical data and expert knowledge are fully utilized to improve the feasibility and adaptability of the scheme. The candidate schemes are verified and evaluated in a variety of production scenarios, and simulation and decision-making methods are used to enhance the robustness of process evaluation and ensure the performance of the selected scheme under different production environments. The present invention can adaptively generate optimized process parameter adjustment schemes according to different production scenarios to improve the production efficiency and product quality stability of tagatose milk tea.

[0128] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0129] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0130] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0131] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0132] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0133] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0134] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A tagatose milk tea processing technology evaluation method, characterized in that: The method comprises: The multidimensional parameter data of tagatose milk tea production process was collected and preprocessed to obtain a multidimensional time series data set; Input the multidimensional time series data set into the TSVT-PNNT model for data reconstruction to obtain reconstructed data; specifically including: performing tensorization processing on the multidimensional time series data set, converting the multidimensional time series data set into a three-order tensor structure, wherein the three dimensions represent time, parameter type and process stage, respectively, to obtain initial tensor data; performing tensor singular value decomposition on the initial tensor data, calculating a left singular matrix, a right singular matrix and a singular value matrix, to form decomposed tensor components; based on the decomposed tensor components, using the truncated p-shrinkage norm as the convex envelope of the tensor rank, calculating an approximate value of the tensor nuclear norm, to obtain an estimated value of the tensor rank; according to the estimated value of the tensor rank , performing soft threshold processing on the singular value matrix, setting the singular values ​​less than the preset threshold to zero, obtaining a sparse singular value matrix, and using the sparse singular value matrix and the original left singular matrix and right singular matrix to perform tensor reconstruction to obtain preliminary reconstructed tensor data; applying a p-norm neural tensor network to the preliminary reconstructed tensor data, capturing the high-order correlation of the data through multi-layer nonlinear transformation, obtaining optimized tensor data, and based on the optimized tensor data, using the alternating direction multiplier method for iterative optimization, minimizing the weighted sum of the reconstruction error and the regularization term, obtaining the final reconstructed tensor data; performing inverse tensor quantization processing on the final reconstructed tensor data, converting it back to a multidimensional time series format, obtaining reconstructed data; Based on the reconstructed data, multiple features are complementary fused to obtain full-band fusion features; specifically, the following steps are performed: multi-scale wavelet decomposition is performed on the reconstructed data, db4 wavelet basis function is used to perform 5-layer decomposition, approximate coefficients and detail coefficients are calculated for equipment temperature, stirring speed, and solubility parameter, respectively, to obtain a wavelet coefficient matrix; based on the wavelet coefficient matrix, an improved maximum absolute value rule is applied to high-frequency components, and a tagatose dissolution kinetic model is combined to set different frequency band weights for each parameter, and the high-frequency coefficient with the largest absolute value after weighting is selected to form a high-frequency feature vector; integrated empirical mode decomposition is performed on the reconstructed data, and multiple intrinsic mode functions and residuals are extracted by adding white noise and calculating the average value of multiple decompositions to obtain low-frequency signal components; the low-frequency signal components are input into the K-SVD sparse representation model, and an appropriate According to the overcomplete dictionary of tagatose milk tea production characteristics, an orthogonal matching pursuit algorithm is used for dictionary learning and sparse coding to obtain a low-frequency feature vector; the high-frequency feature vector and the low-frequency feature vector are adaptively normalized, and according to the dynamic range of each parameter in the tagatose milk tea production process, a piecewise linear mapping method is used to map the feature value to the [0,1] interval to obtain a standardized feature vector; the standardized feature vector is input into a nonlinear feature mapping function, and feature fusion and dimension reduction are performed through multi-layer nonlinear transformation and feature reorganization, while retaining key information related to the quality of tagatose milk tea to obtain a preliminary fusion feature; the preliminary fusion feature is adaptively weighted based on the attention mechanism, a multi-head self-attention module is constructed, and the attention weight is dynamically calculated according to the contribution of each feature to the quality of tagatose milk tea to obtain a full-band fusion feature; Inputting the full-band fusion feature structure into a multi-objective integer linear programming model to perform process parameter constraint analysis to obtain a process parameter constraint range; Based on the process parameter constraint range, a plurality of candidate process parameter adjustment schemes are generated using a Bayesian network and a reinforcement learning method; In various production scenarios, the multiple candidate process parameter adjustment schemes are verified and evaluated respectively, and the optimal candidate process parameter adjustment scheme is selected and output as the target process parameter adjustment scheme.

2. The tagatose milk tea processing technology evaluation method according to claim 1, characterized in that: The multidimensional parameter data in the tagatose milk tea production process is collected and preprocessed to obtain a multidimensional time series data set, including: Install high-precision sensors at the key process nodes of raw material ratio, dissolution, heating, cooling, blending and packaging of tagatose milk tea production line to collect multi-dimensional parameter data of equipment temperature, stirring speed, solubility, raw material ratio, equipment cleanliness, tea extraction temperature and tea extraction time, and transmit the collected data to the central processing unit through industrial Internet of Things to obtain the original data set; Dividing the original data set into time series to obtain initial time series data, identifying outliers in the initial time series data, setting a dynamic threshold based on the solubility characteristics of tagatose, marking data points with deviations exceeding the threshold as outliers, and obtaining outlier marked data; Based on the outlier labeled data, local outlier processing is performed on the data points marked as outliers to obtain an outlier-removed data set, and multiple interpolation is performed on the missing values ​​in the outlier-removed data set to obtain a completed data set; Adaptively standardize each parameter data in the supplemented data set to obtain a multidimensional time series data set.

3. The tagatose milk tea processing technology evaluation method according to claim 1, characterized in that: The full-band fusion feature structure is input into a multi-objective integer linear programming model to perform process parameter constraint analysis to obtain a process parameter constraint range, including: Performing feature importance analysis on the full-band fusion features to obtain feature importance ranking, and based on the feature importance ranking, selecting the top K most important features as decision variables, constructing a decision variable vector, including key process parameters such as equipment temperature, stirring speed, solubility, and raw material ratio, while introducing virtual variables to represent the impact of equipment cleanliness; According to the production process requirements of tagatose milk tea, multiple optimization goals were set, including maximizing product quality score, minimizing energy consumption, minimizing production time and minimizing raw material loss, constructing a multi-objective function, and introducing a target weight allocation mechanism based on fuzzy set theory; Establish the constraint relationship between process parameters, including the nonlinear relationship between tagatose solubility and equipment temperature, the relationship between stirring speed and energy consumption, and the response surface model between raw material ratio and product quality, to form a set of constraint conditions; Adaptively performing piecewise linearization processing on the nonlinear constraints in the constraint condition set, increasing the segment density in the key change area according to the solubility characteristics of tagatose, discretizing the continuous variables into several gears, and converting them into an integer linear programming problem; The decision variable vector, the multi-objective function and the constraint condition set are input into a multi-objective integer linear programming solver to perform mixed integer programming to obtain a process parameter constraint range.

4. The tagatose milk tea processing technology evaluation method according to claim 3, characterized in that: Based on the process parameter constraint range, a plurality of candidate process parameter adjustment schemes are generated using a Bayesian network and a reinforcement learning method, including: Based on the process parameter constraint range and historical production data, an initial Bayesian network model of tagatose milk tea production process is constructed; Performing parameter learning on the initial Bayesian network model, optimizing the network parameters using an expectation maximization algorithm, and adjusting the prior probability in combination with a tagatose dissolution kinetic model to obtain a target Bayesian network model; Based on the target Bayesian network model, the state space, action space and reward function of reinforcement learning are defined, wherein the state space includes current process parameters and quality indicators, the action space is the adjustment amount of the process parameters, and the reward function is constructed based on product quality score and production efficiency; A deep Q network algorithm is used to construct a reinforcement learning agent, and the target Bayesian network model is embedded into the DQN network structure as prior knowledge to obtain a DQN model integrating Bayesian priors; Based on the DQN model integrating Bayesian prior, an ε-greedy strategy is designed for action selection. Meanwhile, a Boltzmann exploration mechanism is introduced to balance exploration and utilization, and multiple candidate process parameter adjustment schemes are generated.

5. The tagatose milk tea processing technology evaluation method according to claim 4, characterized in that: The method of verifying and evaluating the plurality of candidate process parameter adjustment schemes in various production scenarios, selecting and outputting the optimal candidate process parameter adjustment scheme as the target process parameter adjustment scheme, comprises: Based on historical production data and expert knowledge, a multi-scenario simulation model for tagatose milk tea production was constructed, including scenarios of normal production, high temperature environment, and slight equipment pollution. Latin hypercube sampling was used to generate scenario parameter distribution, and a multi-scenario simulation environment was obtained by combining the tagatose dissolution kinetic model and thermodynamic model. Input the multiple candidate process parameter adjustment schemes into a multi-scenario simulation environment for parallel simulation experiments, use a Kriging proxy model to accelerate the simulation process, and obtain performance indicator data of each candidate process parameter adjustment scheme in different scenarios; Conduct multi-dimensional evaluation on the performance indicator data, construct an evaluation indicator system based on fuzzy analytic hierarchy process, and calculate the initial comprehensive score of each candidate process parameter adjustment scheme in different scenarios; Based on the initial comprehensive score, the relative closeness of each candidate process parameter adjustment scheme to the ideal solution is calculated, and the coefficient of variation method is introduced to dynamically adjust the indicator weight to obtain a weighted comprehensive score for each candidate process parameter adjustment scheme; Performing a global sensitivity analysis on the weighted comprehensive score to obtain a sensitivity analysis result, wherein the sensitivity analysis result is used to represent the sensitivity index and interaction effect of the scheme performance to each factor; Based on the sensitivity analysis results, a robustness assessment is performed on each candidate process parameter adjustment scheme to obtain a risk assessment index for each candidate process parameter adjustment scheme; The weighted comprehensive score and the risk assessment index are input into a multi-attribute decision model to sort the schemes, and a scheme sorting result is obtained. Based on the scheme sorting result, a candidate process parameter adjustment scheme with the best comprehensive performance and controllable risk is selected as the target process parameter adjustment scheme and output.

6. A tagatose milk tea processing technology evaluation system, characterized in that: Used to execute the tagatose milk tea processing technology evaluation method according to any one of claims 1 to 5, the tagatose milk tea processing technology evaluation system comprises: The acquisition module is used to collect and preprocess the multidimensional parameter data in the tagatose milk tea production process to obtain a multidimensional time series data set; A reconstruction module, used for inputting the multidimensional time series data set into the TSVT-PNNT model to reconstruct the data and obtain reconstructed data; A fusion module, used for performing complementary fusion of multiple features based on the reconstructed data to obtain full-band fusion features; An analysis module is used to input the full-band fusion feature structure into a multi-objective integer linear programming model to perform process parameter constraint analysis to obtain a process parameter constraint range; A generation module, for generating a plurality of candidate process parameter adjustment schemes based on the process parameter constraint range by using a Bayesian network and a reinforcement learning method; The evaluation module is used to verify and evaluate the multiple candidate process parameter adjustment schemes in multiple production scenarios, select the best candidate process parameter adjustment scheme as the target process parameter adjustment scheme and output it.

7. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and wherein the processor implements the tagatose milk tea processing technology evaluation method according to any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor is enabled to execute the tagatose milk tea processing technology evaluation method according to any one of claims 1 to 5.

Images