Vacuum impregnation and gradient drying parameter optimization control method for fruit leather preparation

Abstract

The application provides a vacuum sugar permeation and gradient drying parameter optimization control method for fruit preserve preparation, and belongs to the technical field of intelligent production control; the application dynamically collects multi-source process parameters and product quality parameters by deploying a high-precision sensor array at key workstations of a sugar permeation and drying device; a deep learning method is used to extract process deep features; an intelligent decision-making model based on deep reinforcement learning is constructed to realize real-time optimization output of each control parameter; the deep reinforcement learning model constructed by the application comprehensively considers multiple targets such as preparation efficiency, product comprehensive quality, energy consumption and process constraints in a reward function. The application improves the automation and precision level of fruit preserve preparation process control, realizes collaborative intelligent regulation and control of the influence of multi-parameter coupling, and realizes multi-target global optimization decision-making.

Description

Technical Field

[0001] This invention belongs to the field of intelligent production control technology, and in particular relates to a method for optimizing and controlling the parameters of vacuum sugar infiltration and gradient drying in the preparation of dried fruit. Background Technology

[0002] Vacuum sugar infiltration and gradient drying are core steps in the processing of dried fruit, and the precision of their parameter control has a decisive impact on the uniformity of sugar content, texture, and color stability of the final product. However, traditional dried fruit preparation methods rely heavily on manual experience for control, making it difficult to adapt to natural fluctuations in raw material characteristics and changes in environmental conditions. This has led to persistent problems such as inconsistent product quality within the same batch or even within the same batch, low energy efficiency, and long processing cycles.

[0003] Developing optimized control methods for vacuum sugar infiltration and gradient drying parameters in dried fruit preparation is an inevitable choice for promoting technological upgrading and improving the precision and intelligence of the traditional dried fruit industry. This research not only helps solve the long-standing pain points of quality fluctuations and high energy consumption in the industry, but also lays a key technological foundation for achieving standardized, efficient, and sustainable development of dried fruit manufacturing, with significant economic benefits and industry promotion value.

[0004] Chinese invention patent application number CN202511088290.4 provides an integrated equipment for making dried fruit and a method for processing low-sugar black fungus dried fruit. The integrated equipment for making dried fruit includes a closed vacuum sugar infusion tank and a dried fruit loading basket horizontally suspended inside the vacuum sugar infusion tank. A detachable sealing cover is provided on the top of the vacuum sugar infusion tank. The sealing cover is provided with a closable air outlet pipe and an air inlet pipe that communicate with the inside of the vacuum sugar infusion tank. The lower part of the side wall of the vacuum sugar infusion tank is also provided with horizontally arranged sugar inlet pipe and sugar outlet pipe. This integrated equipment for making dried fruit effectively integrates vacuum sugar infusion, gradient drying and surface treatment processes, is energy-saving and efficient, and avoids the problem of dried fruit exposure and contamination caused by multi-process processing.

[0005] Chinese invention patent application number CN202310146635.1 relates to a method for preparing natural pumpkin preserves based on plant-extracted gum, belonging to the field of food processing technology. The technical solution of this invention mainly includes processes such as gum boiling, pretreatment, hardening, gradient drying, gradient sugar infiltration, primary drying, shaping, secondary drying, and packaging. This invention proposes a gradient sugar infiltration method during the boiling process, which allows the pumpkin to complete the sugar infiltration process in a short time, simplifying the processing steps and reducing production costs.

[0006] However, the above method has the following limitations:

[0007] 1) Insufficient intelligence and precision in process parameter control. Existing technologies mostly rely on preset, fixed process parameters or simple gradient changes, lacking the ability to perceive and dynamically adjust key parameters during sugar infiltration and drying, and are unable to adaptively optimize based on the actual state of the material;

[0008] 2) Limited ability to control the coupled effects of multiple parameters. Existing methods fail to fully consider the nonlinear interactions between multiple parameters such as vacuum level, temperature, sugar concentration, and gas flow rate, and their combined impact on the quality of the final product. The control strategies are relatively isolated, making it difficult to achieve global optimization.

[0009] 3) Lack of precise perception of drying uniformity and process status. Existing technologies usually control the drying process indirectly by controlling the ambient temperature and humidity. They lack real-time, online monitoring methods for the temperature and humidity distribution at different points in the drying chamber, as well as the surface and internal moisture status of the dried fruit. This makes it difficult to effectively solve problems such as uneven drying, crusting, or over-drying.

[0010] 4) Lack of adaptive and cross-condition migration capabilities. The control parameters of existing technical solutions are often set for specific raw materials or fixed formulas. When the raw material type, initial state or environmental conditions change, it is necessary to rely on human experience to readjust the process. The system itself does not have the ability to quickly learn from small sample data and update itself, which limits its promotion and application.

[0011] 5) The overall optimization objective of the system decision-making is singular. Existing control methods mostly focus on completing a certain process as the main objective, failing to integrate multiple objectives such as preparation efficiency, overall product quality, energy consumption, and raw material loss into a unified optimization framework for collaborative decision-making. Summary of the Invention

[0012] To address the problems of the above methods, this invention proposes a method for optimizing and controlling the parameters of vacuum sugar infiltration and gradient drying in the preparation of dried fruit. In this invention, a high-precision sensor array is deployed at key stations of the sugar infiltration and drying equipment to dynamically collect multi-source process parameters and product quality parameters; deep learning methods are used to extract deep process features; and an intelligent decision-making model based on deep reinforcement learning is constructed to achieve real-time optimization output of each control parameter, ultimately achieving the goals of improving product quality, shortening the preparation cycle, and reducing energy consumption.

[0013] This invention provides a method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters in the preparation of dried fruit, comprising the following steps:

[0014] S1, Collect parameters of the sugar infiltration process, drying process, and quality parameters during the preparation of dried fruit; the sugar infiltration process parameters include time-series data of vacuum pressure inside the tank, time-series data of sugar concentration inside the tank, and time-series data of temperature distribution inside the tank; the drying process parameters include time-series data of humidity in the drying chamber and time-series data of airflow velocity in the drying chamber; the quality parameters include time-series data of moisture content of dried fruit, real-time data of sugar content of dried fruit, time-series data of firmness of dried fruit, and time-series data of color parameters of dried fruit.

[0015] S2, preprocess the three parameters respectively, and extract the dynamic features of the sugar infiltration process through a multi-head self-attention network based on the sugar infiltration process parameters, extract the spatiotemporal features of the drying process through a spatiotemporal graph convolutional network based on the drying process parameters, and extract the temporal features of the quality parameters through a gated recurrent unit network based on the quality evolution of the quality parameters.

[0016] S3. The obtained dynamic features of the sugar infiltration process, the spatiotemporal features of the drying process, and the temporal features of the quality parameters are input into the cross-modal attention fusion network for fruit preserve preparation. Multimodal feature fusion is achieved through cross-modal attention weight calculation to obtain the system state representation vector. Subsequently, the system state representation vector is used as input to construct a proximal policy optimization reinforcement learning model for fruit preserve preparation. The optimal control parameters, including vacuum pump power, sugar solution supply, heater power, and fan speed, are output through the policy network and value network.

[0017] Preferably, the multi-head self-attention network for the fruit candy sugar infusion process consists of two layers of multi-head self-attention modules connected in sequence, each layer containing a sugar infusion parameter specific encoding module and an 8-head self-attention calculation layer;

[0018] The sugar infiltration parameter specific encoding module first has an input projection layer that encodes the standardized sugar infiltration process parameter vector. The data is mapped to an initial feature space, followed by a parameter-specific embedding layer. Independent embedding vectors are assigned to the vacuum level, sugar concentration, temperature, and time parameters, and the parameter type information is encoded into the features through vector addition. Finally, a linear transformation layer maps the features incorporating parameter type information to a unified high-dimensional feature space, generating sugar infiltration parameter embedding features. ;

[0019] The eight-head self-attention computation layer is used to capture parameter interaction relationships in different dimensions; each attention head embeds a unified saccharification parameter into the feature through three independent weight matrices. The query vector, key vector, and value vector are linearly mapped to the query vector, key vector, and value vector, respectively. The correlation strength between parameters at different positions in the input sequence is measured by calculating the dot product of the query vector and all key vectors. The dot product result is scaled by dividing by the square root of the key vector dimension to stabilize the gradient during training. Finally, the scaled score is normalized using the Softmax function to obtain an attention distribution with a weight sum of 1.

[0020] The output features of the eight attention heads are concatenated and then passed through a linear transformation layer and a residual connection to obtain the dynamic feature vector of the sugar infusion process. .

[0021] Preferably, the drying process parameters are first normalized during preprocessing, and then the normalized drying parameters are constructed into a spatiotemporal diagram structure:

[0022] ;

[0023] in Represents the set of nodes in a dry indoor space. For the number of sensors, Represents spatial connection edges between nodes. Represents parameters of the gradient drying process , , The graph node feature matrix formed after normalization.

[0024] Preferably, the spatiotemporal graph convolutional network for the dried fruit process includes a spatial feature extraction layer and a temporal feature extraction layer. The spatial feature extraction layer uses a graph convolutional network to capture the spatial dependencies between nodes, thereby constructing a good graph structure feature matrix. The input is given to the spatial feature extraction layer, and the output is the spatial feature vector at time t. The temporal feature extraction layer uses a one-dimensional convolutional network to capture the temporal evolution of parameters, arranging the spatial features at different times into a three-dimensional tensor in a time series. Input one-dimensional convolutional layer to compute time features ,in Represents the temporal convolution weight matrix. Indicates the first The temporal feature extraction layer outputs the data; spatial and temporal features are fused through residual connections and batch normalization layers to obtain the spatiotemporal features of the drying process. .

[0025] Preferably, the fruit preserve quality evolution gated recurrent unit network adds a quality parameter attention gating module to the traditional gated recurrent unit to enhance the temporal characteristic response of key quality parameters;

[0026] The fruit preserve quality evolution gated loop unit network regulates the historical quality time sequence state through reset gates and update gates. The degree of forgetting and retention, combined with the standardized quality parameter sequence at the current moment. The candidate quality time-series state is calculated; the quality parameter attention gating module dynamically calculates the attention weight vector based on the importance of moisture, sugar content, hardness, and color to the final product. The weight is then used to optimize the update of the candidate quality time series state; the final quality time series state is... The calculation integrates the update gate output and the weighted candidate quality time-series states, and its output is the time-series feature of the quality parameters. .

[0027] Preferably, the cross-modal attention fusion network for fruit preserve preparation specifically comprises:

[0028] First, linear mapping is performed on the three modal features respectively, and the original features are transformed to a unified feature dimension through weight matrix and bias vector, eliminating the differences in dimensions and scale of different modal features;

[0029] Then, cross-modal attention weights are calculated, which are used to quantify the importance of different modal features to the representation of the current system state. The weight calculation is based on the similarity between the mapped feature vector and the modal feature mean vector. The correlation strength between each modal feature and the overall fused feature is measured by the dot product operation, and the softmax function is used for normalization to ensure that the sum of all weights is 1.

[0030] Finally, the system state representation vector is obtained by fusing the three modal features through attention weighting. .

[0031] Preferably, the fruit preserve preparation proximal policy optimization reinforcement learning model first obtains a high-quality initial policy by supervising the policy network of the model. The supervised pre-training part uses the control parameter target value labeled in the dataset as the supervision signal, takes the system state representation vector as the input, and takes the control parameter target value as the target output to minimize the error between the predicted value and the control parameter target value.

[0032] The state space of the model is defined as That is, the set of multimodal fused state features at all times; the action space is defined as... ,in Indicates the first Action vectors at each time step; probability distribution of actions output by the model's policy network after supervised pre-training. ,in The policy network parameters are represented by a reparameterization technique to ensure gradient differentiability; the model's value network is used to estimate state values. ,in Represents the value network parameters used to calculate the advantage function. ,in Indicates the first Instant rewards for each moment This represents the discount factor, used to balance immediate rewards with long-term rewards.

[0033] Preferably, the reward function of the fruit preserve preparation proximal strategy optimization reinforcement learning model includes: preparation efficiency reward. Used to incentivize shorter preparation cycles; product quality rewards Used to incentivize products to meet quality standards; energy consumption rewards Used to incentivize energy reduction; process constraint rewards It is used to restrain violations of regulations.

[0034] The model continuously interacts with the fruit preserve preparation simulation platform, collecting state-action-reward data and updating network parameters using a proximal policy optimization objective function. By alternately updating the policy network and value network, the model learns the optimal adaptive control strategy and outputs the optimal control parameters. .

[0035] Preferably, after the reinforcement learning model for optimizing the proximal strategy in the dried fruit preparation process is deployed, it continuously collects real-time quality parameter data and control output data during the dried fruit preparation process, and calculates the deviation between the real-time quality parameters and the preset target values. ,when When the preset deviation threshold is exceeded, or when changes in raw material type or environmental conditions are detected, a dynamic data acquisition mechanism is triggered to collect multi-source parameter data and status characteristics under this operating condition. Controlling actions and reward value This forms a small-sample fine-tuning dataset. ;

[0036] When obtaining a small sample fine-tuning dataset Then, the outer loop of meta-learning is fine-tuned, first based on the learning rate of the inner loop. The model parameters are initially updated, and then the learning rate is adjusted via the outer loop. The parameters are updated a second time to obtain the fine-tuned model parameters. .

[0037] Compared with the prior art, the present invention has the following innovative features and beneficial effects:

[0038] (1) Improved the automation and precision of the fruit preserve preparation process control: A high-precision sensor array is used to collect multi-source parameters in real time and from all directions during the sugar infiltration and drying process, and a deep reinforcement learning model is used to automatically generate the optimal control strategy, which completely changes the traditional mode that relies on human experience and fixed formula settings. Traditional methods cannot make real-time dynamic adjustments to the process, resulting in large fluctuations in quality. However, this technology, through an intelligent decision model, can automatically and accurately output control commands such as vacuum degree, sugar liquid replenishment amount, and fan speed based on the real-time perceived system status, which greatly reduces human intervention and ensures the stability and consistency of the production process;

[0039] (2) Achieved collaborative intelligent control of the coupling effects of multiple parameters: By innovatively introducing a multi-head self-attention mechanism and a spatiotemporal graph convolutional network, this technology can deeply explore and understand the complex nonlinear interactions between multiple parameters such as vacuum degree, temperature, concentration, and airflow, and their joint impact on the final quality. Traditional methods often control a single parameter in isolation, making it difficult to achieve global optimization. However, the model of this technology can adaptively capture these coupling relationships and consider them comprehensively when making decisions, thereby achieving collaborative optimization of sugar penetration efficiency and drying uniformity, effectively solving the problems of insufficient sugar penetration, drying crusting, or cracking caused by parameter mismatch in traditional methods;

[0040] (3) Achieving global optimization decision-making for multiple objectives: The deep reinforcement learning model constructed in this technology comprehensively considers multiple objectives such as preparation efficiency, overall product quality, energy consumption, and process constraints in its reward function. Traditional methods often focus on a single objective and are difficult to balance. The agent in this technology learns the long-term optimal strategy that can balance multiple competing objectives by interacting with the environment. Thus, it can simultaneously achieve energy saving and efficiency improvement while ensuring high quality, thereby maximizing the overall benefits of dried fruit preparation. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0042] Figure 2 This is a structural diagram of an intelligent decision-making model based on cross-modal attention and deep reinforcement learning.

[0043] Figure 3 This is a diagram of the overall system framework of the present invention.

[0044] Figure 4 This is a graph showing the time consumption of feature processing in the embodiment.

[0045] Figure 5 This is a comparison chart of overall quality deviations in the examples.

[0046] Figure 6 The image shows a heatmap illustrating the overall performance in this embodiment. Detailed Implementation

[0047] This invention proposes a method for optimizing and controlling parameters of vacuum sugar infiltration and gradient drying in the preparation of dried fruit. The overall process is as follows: Figure 1 As shown:

[0048] S1. To address the need for multimodal data construction in the fruit preserve preparation process, this invention deploys high-precision sensor arrays at key stations in the vacuum sugar infusion equipment and gradient drying equipment to collect multidimensional parameters such as vacuum degree, sugar solution concentration, temperature, humidity, wind speed, moisture content, sugar content, hardness, and color. Simultaneously, it collects control parameters such as vacuum pump power, sugar solution supply, heater power, and fan speed through the equipment control system interface. Outlier removal and timestamp alignment are performed on the collected multi-source parameter data. Finally, based on process standards, the preparation status is labeled and the target values ​​of control parameters are labeled to construct a multimodal dataset for training an intelligent control model.

[0049] S2 inputs the multimodal dataset obtained in S1 into the multi-source parameter deep preprocessing and feature extraction module. First, outlier removal, filtering, and standardization are performed on the sugar infusion process parameters. Dynamic features of the sugar infusion process are extracted through a multi-head self-attention network for the fruit preserve sugar infusion process. Second, spatial alignment and normalization are performed on the drying process parameters. Spatiotemporal features of the drying process are extracted through a spatiotemporal graph convolutional network for the fruit preserve drying process. Finally, outlier removal and standardization are performed on the quality parameters. Temporal features of the quality parameters are extracted through a gated recurrent unit network for fruit preserve quality evolution.

[0050] S3 inputs the dynamic features of the sugar infiltration process, the spatiotemporal features of the drying process, and the temporal features of the quality parameters output from S2 into the cross-modal attention fusion network for fruit preserve preparation. Multimodal feature fusion is achieved through cross-modal attention weight calculation to obtain the system state representation vector. Subsequently, using the system state representation vector as input, a proximal policy optimization reinforcement learning model for fruit preserve preparation is constructed. The optimal control parameters are output through the policy network and the value network, and the reward function comprehensively considers preparation efficiency, product quality, energy consumption, and process constraints.

[0051] S4. After the intelligent decision-making model is deployed online, a real-time deviation monitoring module is built to collect real-time parameter data and control output data, and calculate the comprehensive quality deviation. When the deviation exceeds the threshold or a change in working conditions is detected, dynamic data collection is triggered to form a small sample fine-tuning dataset. The model is rapidly fine-tuned with small samples through the fruit preserve preparation meta-learning framework combined with the sugar infiltration-drying coupling adaptation module. The model parameters are dynamically updated based on performance evaluation.

[0052] In the system integration phase, high-precision sensor arrays are installed at key workstations of the vacuum sugar infiltration and gradient drying equipment according to the deployment plan, and connected to the control system via industrial Ethernet. The trained and optimized fruit preserve preparation proximal strategy optimization reinforcement learning model and related networks are deployed to the industrial control computer. After the system starts, it receives sensor data in real time, obtains system state representation through feature extraction and fusion network, and outputs optimized control parameters from the fruit preserve preparation proximal strategy optimization reinforcement learning model, which are then sent to the actuators via the device bus. At the same time, closed-loop control and target achievement of the preparation process are realized through real-time deviation monitoring and meta-learning adaptive optimization.

[0053] The specific implementation process of the present invention will be further described below with reference to specific embodiments.

[0054] S1. Construction of a multimodal dataset for the dried fruit preparation process.

[0055] S1-1 Multi-source sensor deployment and parameter acquisition

[0056] High-precision sensor arrays are deployed at key stations in the vacuum sugar infiltration equipment and gradient drying equipment to achieve synchronous acquisition of multi-dimensional parameters of the preparation process. Inside the sugar infiltration tank of the vacuum sugar infiltration equipment, vacuum sensors, sugar concentration sensors, and temperature sensors are uniformly deployed along the circumference of the tank wall. The vacuum sensors collect real-time vacuum pressure data, which is recorded as real-time data. The sugar concentration sensors collect real-time sugar concentration data, which is recorded as real-time data. The temperature sensors collect real-time temperature data for different areas within the tank, which is recorded as real-time temperature distribution data. A time recording device is also configured synchronously to accurately record the time parameters of each stage of sugar infiltration. Inside the drying chamber of the gradient drying equipment, a humidity sensor array and anemometer are deployed in a spatial grid distribution. The humidity sensor array collects real-time humidity data for different areas within the drying chamber; the anemometer collects real-time airflow velocity data for drying. At the finished product testing station in the preparation process, an online moisture analyzer, a sugar analyzer (Near-Infrared Spectrometer, NIRS), a texture analyzer, and a color sensor are deployed. The online moisture analyzer collects real-time data on the moisture content of the dried fruit, the sugar analyzer collects real-time data on the sugar content, the texture analyzer collects real-time data on the firmness, and the color sensor collects real-time data on the color parameters. These together serve as core indicators for quality evaluation. Simultaneously, control parameters related to the preparation process are collected synchronously through the equipment control system interface. These control parameters include real-time data on vacuum pump power, sugar solution replenishment, heater power, and fan speed.

[0057] S1-2 Data Preprocessing and Timestamp Alignment

[0058] The collected multi-source parameter data underwent systematic preprocessing. First, box plots were used to remove outliers from all real-time data collected in S1.1. This was done by determining whether the data fell within the range of outliers. The interval, in which It is the first quartile. It is the third quartile. For the interquartile range, data exceeding this range are marked as outliers and filled using linear interpolation. Secondly, to address the differences in sampling frequencies among different sensors, a unified timestamp synchronization mechanism is established. Using the highest sampling frequency as a benchmark, the sampling time points of each parameter are adjusted using linear interpolation to ensure a one-to-one correspondence between all parameters and control parameters at the same timestamp, resulting in a time-aligned multi-source time-series dataset. This multi-source time-series dataset specifically includes time-series data on tank vacuum pressure, tank sugar concentration, tank temperature distribution, drying chamber humidity, drying airflow velocity, fruit moisture content, fruit sugar content, fruit firmness, fruit color parameters, vacuum pump power, sugar supply, heater power, and fan speed.

[0059] S1-3 Data Labeling and Multimodal Dataset Integration

[0060] Based on the GB / T standard, the time-aligned multi-source time-series datasets were labeled. The labeling included preparation status annotation and control parameter target value annotation. The preparation status annotation process was as follows: At the critical process endpoints, if the instantaneous values ​​of the time-series data for fruit moisture content, fruit sugar content, fruit firmness, and fruit color parameters are all within the acceptable range specified in the GB / T standard, and the fluctuations of the time-series data for tank vacuum pressure, tank sugar concentration, tank temperature distribution, drying chamber humidity, and drying airflow velocity are all within a stable range during the process stage, then the preparation status for that time period is considered complete. The state is marked as meeting quality standards, otherwise it is marked as failing quality standards. For the control parameter target value labeling, different preparation states are identified. First, the corresponding vacuum pump power time-series data, sugar solution replenishment time-series data, heater power time-series data, and fan speed time-series data are extracted from the data samples marked as meeting quality standards. Then, statistical analysis is performed on the extracted control parameter data sequences. The mode of the data distribution of each parameter during the stable and efficient operation phase is used as the benchmark for its set value. The mean and variance of the data during this phase are calculated for verification, ultimately determining the optimal vacuum pump power. Sugar solution replenishment set value Heater power setting value Fan speed setting value To obtain the target value of the control parameter Finally, the preprocessed multi-source time-series parameter data, labeled preparation state information, and target values ​​of control parameters are integrated to construct a multimodal dataset for training the intelligent control model. The multimodal dataset specifically includes time-series data on vacuum pressure inside the tank, sugar concentration inside the tank, temperature distribution inside the tank, humidity in the drying chamber, airflow velocity in the drying chamber, moisture content of the dried fruit, sugar content of the dried fruit, firmness of the dried fruit, color parameters of the dried fruit, target values ​​of control parameters, and labeled preparation state information.

[0061] S2, Multi-source parameter deep preprocessing and feature extraction module

[0062] S2-1 Dynamic Feature Extraction of the Glucose Infusion Process Based on Parameter-Specific Multi-Head Self-Attention

[0063] For multimodal datasets The time-series data of tank vacuum pressure, tank sugar concentration, and tank temperature distribution were preprocessed, and the data were smoothed using a moving average filter. The calculation formula is as follows:

[0064] ;

[0065] in Indicates the filtered first... The osmotic parameters at any given time, Indicates the size of the sliding window. Indicates the first The original osmotic parameters at each time point. The filtered data undergoes standardization to map the parameters to a uniform scale; the calculation formula is as follows:

[0066] ;

[0067] in Represents the standardized first The osmotic parameters at any given time, This represents the mean value of the corresponding osmotic parameters. This represents the standard deviation of the corresponding osmotic parameters.

[0068] The standardized osmotic parameters are integrated into a vector. The input process for sugar infusion in candied fruit is described using a multi-head self-attention network (MSA). This network consists of two sequentially connected MSA modules, each containing a sugar infusion parameter-specific encoding module and an 8-head self-attention computation layer. The MSA adds a sugar infusion parameter-specific encoding module to the traditional MSA mechanism. This module is based on the physical characteristics and process correlations of parameters such as vacuum level, sugar concentration, temperature, and time. The specific structure of the sugar infusion parameter-specific encoding module is as follows: First, there is an input projection layer that projects the vector... The data is mapped to the initial feature space; then a parameter-specific embedding layer is applied, where independent embedding vectors are assigned to vacuum level, sugar concentration, temperature, and time parameters, encoding parameter type information into the features through vector addition; finally, a linear transformation layer maps the features incorporating parameter type information to a unified high-dimensional feature space, generating sugar infiltration parameter embedding features. .

[0069] The fruit candy sugar infusion process utilizes a multi-head self-attention network (MIAN) to capture parameter interactions across different dimensions through eight self-attention computation layers in each layer. Each attention head embeds uniform sugar infusion parameters into the features using three independent weight matrices. These are linearly mapped to query vector, key vector, and value vector, respectively. The core of attention weight calculation is the query-key matching mechanism, which measures the correlation strength between parameters at different positions in the input sequence by calculating the dot product of the query vector and all key vectors. The dot product result is scaled by dividing by the square root of the key vector dimension to stabilize the gradient during training. Finally, the Softmax function is used to normalize the scaled score, resulting in an attention distribution where the sum of the weights is 1.

[0070] The output features of the eight attention heads are concatenated and then passed through a linear transformation layer and a residual connection to obtain the dynamic feature vector of the sugar infusion process. The dynamic feature vector of this sugar infiltration process comprehensively captures the nonlinear interaction between vacuum degree, sugar concentration, temperature and time and their combined influence on sugar infiltration efficiency.

[0071] S2-2 Spatiotemporal Feature Extraction of Drying Process Based on Spatiotemporal Map Convolution of Drying Field

[0072] This step involves the multimodal dataset. The time-series data of humidity in the drying chamber and drying airflow velocity in the parameters are preprocessed, and the parameters are mapped to the [0,1] interval by min-max normalization. The calculation formula is as follows:

[0073] ;

[0074] in Represents the normalized i-th Time of the first Drying parameter values ​​for each spatial node, This represents the original drying parameter values. This represents the minimum value of the corresponding drying parameter. This indicates the maximum value of the corresponding drying parameter.

[0075] The normalized drying parameters are constructed into a spatiotemporal diagram structure:

[0076] ;

[0077] in Represents the set of nodes in the dry indoor space ( (Number of sensors) Represents spatial connection edges between nodes. Represents parameters of the gradient drying process , , The normalized graph node feature matrix is ​​then constructed. The spatiotemporal graph structure is input into a spatiotemporal graph convolutional network for the dried fruit processing. This network includes spatial feature extraction layers and temporal feature extraction layers. The spatial feature extraction layer uses a graph convolutional network to capture the spatial dependencies between nodes, thus constructing the graph structure feature matrix. The input is given to the spatial feature extraction layer, and the output is the spatial feature vector at time t. The temporal feature extraction layer uses a one-dimensional convolutional network to capture the temporal evolution of parameters, arranging the spatial features at different times into a three-dimensional tensor in a time series. ( (where the time window length is the input to the one-dimensional convolutional layer to calculate temporal features) ,in Represents the temporal convolution weight matrix. Indicates the first The output of the temporal feature extraction layer is shown. Spatial and temporal features are fused through residual connections and batch normalization layers to obtain the spatiotemporal features of the drying process. ,in This indicates a batch normalization operation. Indicates the first The spatial-temporal feature vector of the drying process at any given moment captures both the spatial distribution dependence and the temporal evolution of the drying field.

[0078] S2-3 Extraction of Temporal Features of Quality Parameters Based on Quality Evolution Gated Recurrent Units

[0079] This step involves the multimodal dataset. The time-series data of dried fruit moisture content, sugar content, firmness, and color parameters were standardized, and the calculation formula is as follows:

[0080] ;

[0081] in Represents the standardized first Moment quality parameter values, Indicates the original quality parameter value. This represents the mean value of the corresponding quality parameter. This represents the standard deviation of the corresponding quality parameter.

[0082] The standardized quality parameter sequence is used as the input sequence. The input is a gated recurrent unit network for fruit preserve quality evolution. This network adds a quality parameter attention gating module to the traditional gated recurrent unit to enhance the temporal characteristic response of key quality parameters.

[0083] The fruit preserve quality evolution gated loop unit network regulates the historical quality time sequence state through reset gates and update gates. The degree of forgetting and retention, combined with the input sequence at the current moment. The candidate quality time-series state is calculated. The quality parameter attention gating module dynamically calculates the attention weight vector based on the importance of different quality parameters such as moisture, sugar content, firmness, and color to the final product. This weight is then used to optimize the update of candidate quality time series states. Final quality time series state. The calculation integrates the update gate output and the weighted candidate quality time-series states, and its output is the time-series feature of the quality parameters. .

[0084] S3. Construction of an intelligent decision-making model based on cross-modal attention and deep reinforcement learning

[0085] Model structure as follows Figure 2 As shown, the dynamic features of the sugar infiltration process, the spatiotemporal features of the drying process, and the temporal features of the quality parameters are input into the cross-modal attention fusion network for fruit preserve preparation. Multimodal feature fusion is achieved through cross-modal attention weight calculation to obtain the system state representation vector. Subsequently, using the system state representation vector as input, a proximal policy optimization reinforcement learning model for fruit preserve preparation is constructed. The optimal control parameters are output through the policy network and the value network. The reward function comprehensively considers preparation efficiency, product quality, energy consumption, and process constraints.

[0086] S3-1 Multimodal Feature Fusion Based on Fruit Preserves for Cross-Modal Attention

[0087] This step will output the dynamic characteristics of the sugar infusion process from S2-1. Spatiotemporal characteristics of the drying process output by S2-2 and the timing characteristics of the quality parameters output by S2-3 As input, the Preserved Fruit Preparation Cross-Modal Attention Fusion Network (PFPC-MAFN) is used. This network aims to establish explicit associations between features of different modalities and dynamically adjust the contribution weights of each modal feature. First, linear mapping is performed on the three modal features respectively, transforming the original features to a unified feature dimension through weight matrices and bias vectors. This eliminates the differences in dimensions and scales between the different modal features, providing a consistent representation space for subsequent cross-modal attention calculations. The calculation formula is as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] in , , They represent the first The time-mapping features of sugar infiltration, drying, and quality. , , These represent the mapping weight matrices for the corresponding modes. , , These represent the mapping bias vectors for the corresponding modes.

[0092] Cross-modal attention weights are calculated, which quantify the importance of different modal features to the representation of the current system state. The weights are calculated based on the similarity between the mapped feature vector and the mean vector of the modal features. A dot product operation is used to measure the correlation strength between each modal feature and the overall fused features, and a softmax function is used for normalization to ensure that the sum of all weights is 1. The weight calculation formula is:

[0093] ;

[0094] ;

[0095] ;

[0096] in , , They represent the first The attentional weight given to sugar infiltration, drying, and quality characteristics at all times should satisfy:

[0097] ;

[0098] ;

[0099] This represents the modal feature mean vector.

[0100] By fusing the three modal features using attention weights, a system state representation vector is obtained:

[0101] ;

[0102] in Indicates the first The multimodal fusion state feature vector of the fruit preserve preparation system comprehensively and accurately reflects the overall state of the preparation process.

[0103] S3-2 Construction of a Reinforcement Learning Intelligent Decision-Making Model Based on Proximal Strategy Optimization in Dried Fruit Production

[0104] This step constructs a Preserved Fruit Preparation Proximal Policy Optimization Reinforcement Learning Model (PFPP-PPORM). First, a high-quality initial policy is obtained by supervising the pre-training of the model's policy network. The supervised pre-training part utilizes the target values ​​of control parameters labeled in the S1-3 multimodal dataset as supervision signals. These target control parameters include the optimal vacuum pump power. Sugar solution replenishment set value Heater power setting value Fan speed setting value The system state representation vector output by S3-1 Using the target value of the control parameter as input and the target value of the control parameter as output, the error between the predicted value and the target value of the control parameter is minimized, providing a high-performance starting point for subsequent reinforcement learning and accelerating convergence.

[0105] The state space of the model is defined as That is, the set of multimodal fused state features at all times; the action space is defined as... ,in Indicates the first The action vector at each time step. The probability distribution of actions output by the model's policy network after supervised pre-training. ,in The policy network parameters are represented by a parameterization technique, where action sampling is implemented to ensure gradient differentiability. The model's value network is used to estimate state values. ,in Represents the value network parameters used to calculate the advantage function. ,in Indicates the first Instant rewards for each moment This represents the discount factor, used to balance immediate rewards with long-term rewards.

[0106] The reward function comprehensively considers preparation efficiency, product quality, energy consumption, and process constraints. The calculation formula is as follows:

[0107] ;

[0108] in , , , This represents the weight coefficient of each reward item, satisfying... .

[0109] Preparation efficiency bonus:

[0110] ;

[0111] in Indicates the first Accumulate preparation time at all times. This indicates the preset maximum preparation time, used to incentivize shorter preparation cycles;

[0112] Product quality awards:

[0113] ;

[0114] in , , , These represent the preset target values ​​for moisture content, sugar content, hardness, and color, respectively. , , , This represents the quality deviation penalty coefficient, used to incentivize products to meet quality standards.

[0115] Energy consumption bonus:

[0116] ;

[0117] in Indicates the first Accumulated energy consumption over time This indicates the preset maximum allowable energy consumption, used to incentivize energy reduction.

[0118] Process constraint rewards:

[0119] (The quality of the preparation state information corresponding to the multimodal dataset annotation meets the standards) or (The quality of the preparation state information labeled in the multimodal dataset is substandard), used to constrain violations.

[0120] The model continuously interacts with the fruit preparation simulation platform, collecting state-action-reward data and updating network parameters using a proximal policy optimization objective function. By alternately updating the policy network and value network, the model learns the optimal adaptive control strategy and outputs the optimal control parameters. .

[0121] S4. Construction of an Adaptive Optimization Platform Based on Meta-Learning

[0122] S4-1 Real-time Deviation Monitoring and Dynamic Data Acquisition

[0123] After the intelligent decision-making model is deployed online, this step involves building a real-time deviation monitoring module to continuously collect real-time parameter data and control output data for the dried fruit preparation process. The real-time parameter data includes parameters for the sugar infiltration process after S2 pretreatment, drying process parameters, and quality parameters. The control output data consists of the optimal control parameters output by the PFPP-PPORM model. Calculate the deviation between the real-time quality parameters and the preset target values. The deviation calculation formula is as follows:

[0124] ;

[0125] in Indicates the first Constantly consider overall quality deviations. This indicates a deviation in moisture content. This indicates a deviation in sugar content. Indicates hardness deviation. Indicates color deviation. When When the preset deviation threshold is exceeded, or when changes in raw material type or environmental conditions are detected, a dynamic data acquisition mechanism is triggered to collect multi-source parameter data and status characteristics under that operating condition. Controlling actions and reward value This forms a small-sample fine-tuning dataset. .

[0126] S4-2 Fast Fine-Tuning Based on Meta-Learning in Dried Fruit Preparation (Small Samples)

[0127] This step constructs the Preserved Fruit Preparation Meta-Learning Framework (PFPM-LF). This framework introduces a sugar infiltration-drying coupled adaptation module based on the MAML (Model-Agnostic Meta-Learning) algorithm to rapidly adapt to changes in raw materials and environmental drift. The meta-training phase of the framework uses a multimodal dataset constructed using S1. Based on this, the initial parameters of the PFPP-PPORM model are trained through multi-task learning. The meta-training objective function is:

[0128] ;

[0129] in Indicates the task. Indicates task distribution. The training set representing the task. This represents the test set for the task. This represents the learning rate of the meta-learning inner loop. This represents the mean square error loss.

[0130] When obtaining a small sample fine-tuning dataset Then, the outer loop of meta-learning is fine-tuned, first based on the learning rate of the inner loop. Initial updates to the model parameters:

[0131] ;

[0132] in Indicates the current model parameters. This represents the parameters updated in the inner loop. Subsequently, the learning rate is used in the outer loop. The parameters are updated a second time, and the outer loop objective function is:

[0133] ;

[0134] The parameter update formula is:

[0135] ;

[0136] in This represents the model parameters after fine-tuning. This represents the learning rate of the outer loop. The sugar infiltration-drying coupling adaptation module dynamically adjusts the weights of each parameter in the loss function by analyzing the correlation changes between the sugar infiltration parameters and the drying parameters under new operating conditions.

[0137] S4-3 Model Parameter Dynamic Update and Continuous Optimization

[0138] This step establishes a dynamic update mechanism for model parameters, monitors the performance of the fine-tuned model in real time, and uses the overall quality deviation as the performance evaluation metric. and cumulative rewards When the fine-tuned model operates over multiple consecutive time steps... Below the deviation threshold and During continuous improvement, The current optimal parameters are determined and updated to the online PFPP-PPORM model. If the performance does not meet expectations, small sample data is collected again and the fine-tuning process in S4-2 is repeated until the model performance meets the requirements. At the same time, a parameter update document is established to record the operating conditions, parameter changes, and performance improvements for each update.

[0139] S5, System Integration and Intelligent Control Implementation

[0140] The overall system framework diagram is as follows: Figure 3 As shown.

[0141] S5-1 Sensor Array Deployment and System Interconnection

[0142] This step follows the deployment plan in S1-1, installing various high-precision sensors at designated key positions in the vacuum sugar infusion equipment and gradient drying equipment. This ensures that vacuum sensors, sugar concentration sensors, and temperature sensors are evenly distributed within the sugar infusion tank; humidity sensor arrays and wind speed sensors are arranged in a spatial grid within the drying chamber; and online moisture analyzers, sugar analyzers, texture analyzers, and color sensors are precisely deployed at the finished product testing station. All sensors are connected to the core controller of the fruit preserve preparation control system via an industrial Ethernet data interface, using the Modbus TCP communication protocol to achieve data transmission, ensuring the real-time performance and stability of data acquisition. Simultaneously, the controller establishes communication with underlying actuators such as the vacuum pump, sugar supply system, heaters, and fans via the device bus, enabling reliable issuance of control commands.

[0143] S5-2 Intelligent Control Model Deployment and Operating Environment Configuration

[0144] This step involves deploying the trained and S4-optimized PFPP-PPORM model, the PFPC-MAFN fusion network, and the S2 feature extraction network to an industrial control computer. This computer meets the requirements for real-time model computation and data storage. By writing device drivers and data interface programs, data interaction and command transmission between the control computer and the sensor array and underlying actuators are achieved.

[0145] S5-3 Real-time Data Processing and Intelligent Control Execution

[0146] After the intelligent control system starts up in this step, it receives multi-source parameter data from various sensors in real time. This data is first transmitted to the feature extraction module in S2. Following the preprocessing procedures and feature extraction methods of S2-1, S2-2, and S2-3, outlier removal, filtering, alignment, and normalization operations are performed sequentially to extract the dynamic features of the sugar infiltration process. Spatiotemporal characteristics of the drying process and quality parameters and time series characteristics Subsequently, these feature vectors are input into the PFPC-MAFN fusion network, and after cross-modal attention weighted fusion, the system state representation vector is obtained. The PFPP-PPORM model uses... As input, the system outputs optimal control parameters based on the learned optimal control strategy. The control computer converts these setpoints into standardized control commands, which are then sent to the corresponding underlying actuators via the device bus.

[0147] S5-4 Closed-Loop Management and Target Achievement in the Preparation Process

[0148] During this step of the control execution process, the system continuously collects real-time preparation parameters and product quality parameters through sensors, and calculates the overall quality deviation according to the real-time deviation monitoring method in S4. Dynamically evaluate the control effect. When changes in raw material type or environmental condition drift are detected, leading to... When the threshold is exceeded, the S4 meta-learning adaptive optimization process is automatically initiated, rapidly fine-tuning and updating model parameters through small samples to ensure the adaptability of the control strategy. Simultaneously, the system has a built-in condition monitoring module that monitors equipment operating status and parameter changes in real time. In the event of sensor failure, actuator malfunction, or other issues, an alarm signal is immediately triggered, and the relevant preparation process is paused.

[0149] Experimental verification and analysis:

[0150] This experiment, based on a constructed intelligent control system for candied fruit preparation, was tested in an industrial-grade candied fruit production workshop and a simulated laboratory environment, covering various operating conditions including standard raw materials, raw material variety switching, and environmental temperature and humidity drift. Data acquisition employed a sensor array deployed according to the S1-1 scheme (vacuum sensor accuracy ±0.01MPa, sugar solution concentration sensor accuracy ±0.1Brix, temperature and humidity sensor accuracy ±0.5℃ / ±2% RH), simultaneously collecting parameters related to sugar infiltration, drying, and quality. Based on the GB / T 10782-2021 standard, six types of true values ​​for the preparation states were labeled, constructing a multimodal dataset containing 8000 time-series samples, divided into training, validation, and test sets in a 7:1:2 ratio. A traditional PID control scheme was selected as a baseline for comparison to verify the core performance of this system.

[0151] 1. Multimodal feature extraction and fusion performance

[0152] The performance of the S2 feature extraction module and the S3-1 cross-modal fusion network was verified from three dimensions: feature representation effectiveness, cross-modal fusion rationality, and real-time performance.

[0153] The evaluation indicators are shown in Table 1:

[0154] Table 1 Experimental results of multimodal feature extraction and fusion performance.

[0155]

[0156] Feature discriminant score: F1-score for classifying the state based on feature vectors; the higher the value, the stronger the feature representation ability.

[0157] Fusion weight fluctuation coefficient: the coefficient of variation of cross-modal attention weights; the lower the value, the more reasonable the dynamic adjustment of the weights.

[0158] Feature processing time: e.g. Figure 4 As shown, the total computational latency (in milliseconds) from preprocessing to fusion completion for a single time-series sample.

[0159] 2. Intelligent control and meta-learning adaptive effect

[0160] To achieve the three core objectives of control accuracy, adaptive speed, and energy consumption optimization, the performance of the S3 reinforcement learning decision model and the S4 meta-learning adaptive module was verified, and the differences between this system and traditional PID control were compared:

[0161] Evaluation indicators:

[0162] Overall quality deviation: such as Figure 5 As shown, the overall deviation between the quality parameters and the target values ​​is such that the smaller the value, the higher the control accuracy.

[0163] Adaptive iteration count: The number of iterations required for the model to fine-tune to meet the target after changes in operating conditions; the fewer the iterations, the faster the adaptation.

[0164] Energy consumption reduction rate: The percentage of energy savings compared to traditional PID control; a higher value indicates better energy-saving effect.

[0165] Table 2. Experimental Results of Intelligent Control and Meta-learning Adaptive Effect

[0166]

[0167] The experimental results are shown in Table 2 and Figure 6As shown, this system achieves accurate characterization of the preparation process state through multimodal feature deep extraction and cross-modal attention fusion. The reinforcement learning-based intelligent decision-making model significantly improves control accuracy and reduces energy consumption. The meta-learning adaptive module can quickly respond to changes in raw materials and environmental drift, and can complete model fine-tuning without a large number of samples. Overall performance is superior to traditional PID control schemes. This system provides reliable technical support for the intelligent and adaptive optimization of dried fruit production. In the future, it can be further expanded to more dried fruit categories to improve the system's generalization ability.

[0168] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0169] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for optimizing and controlling parameters of vacuum sugar infiltration and gradient drying in the preparation of dried fruit, characterized in that, Includes the following steps: S1, Collect parameters of the sugar infiltration process, drying process, and quality parameters during the preparation of dried fruit; the sugar infiltration process parameters include time-series data of vacuum pressure inside the tank, time-series data of sugar concentration inside the tank, and time-series data of temperature distribution inside the tank; the drying process parameters include time-series data of humidity in the drying chamber and time-series data of airflow velocity in the drying chamber; the quality parameters include time-series data of moisture content of dried fruit, real-time data of sugar content of dried fruit, time-series data of firmness of dried fruit, and time-series data of color parameters of dried fruit. S2, preprocess the three parameters respectively, and extract the dynamic features of the sugar infiltration process through a multi-head self-attention network based on the sugar infiltration process parameters, extract the spatiotemporal features of the drying process through a spatiotemporal graph convolutional network based on the drying process parameters, and extract the temporal features of the quality parameters through a gated recurrent unit network based on the quality evolution of the quality parameters. S3. The obtained dynamic features of the sugar infiltration process, the spatiotemporal features of the drying process, and the temporal features of the quality parameters are input into the cross-modal attention fusion network for fruit preserve preparation. Multimodal feature fusion is achieved through cross-modal attention weight calculation to obtain the system state representation vector. Subsequently, the system state representation vector is used as input to construct a proximal policy optimization reinforcement learning model for fruit preserve preparation. The optimal control parameters, including vacuum pump power, sugar solution supply, heater power, and fan speed, are output through the policy network and value network. The cross-modal attention fusion network for dried fruit preparation is specifically as follows: First, linear mapping is performed on the three modal features respectively, and the original features are transformed to a unified feature dimension through weight matrix and bias vector, eliminating the differences in dimensions and scale of different modal features; Then, cross-modal attention weights are calculated, which are used to quantify the importance of different modal features to the representation of the current system state; The weight calculation is based on the similarity between the mapped feature vector and the modal feature mean vector. The correlation strength between each modal feature and the overall fused feature is measured by the dot product operation, and the softmax function is used for normalization to ensure that the sum of all weights is 1. Finally, the system state representation vector is obtained by weighting and fusing the three modal features through attention weights.

2. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 1, characterized in that: The multi-head self-attention network for the fruit candy sugar infusion process consists of two layers of multi-head self-attention modules connected in sequence. Each layer contains a sugar infusion parameter specific encoding module and an 8-head self-attention calculation layer. The sugar infiltration parameter specific encoding module first has an input projection layer that encodes the standardized sugar infiltration process parameter vector. The data is mapped to an initial feature space, followed by a parameter-specific embedding layer. Independent embedding vectors are assigned to the vacuum level, sugar concentration, temperature, and time parameters, and the parameter type information is encoded into the features through vector addition. Finally, a linear transformation layer maps the features incorporating parameter type information to a unified high-dimensional feature space, generating sugar infiltration parameter embedding features. ; The eight-head self-attention computation layer is used to capture parameter interaction relationships in different dimensions; each attention head embeds a unified saccharification parameter into the feature through three independent weight matrices. The query vector, key vector, and value vector are linearly mapped to the query vector, key vector, and value vector, respectively. The correlation strength between parameters at different positions in the input sequence is measured by calculating the dot product of the query vector and all key vectors. The dot product result is scaled by dividing by the square root of the key vector dimension to stabilize the gradient during training. Finally, the scaled score is normalized using the Softmax function to obtain an attention distribution with a weight sum of 1. The output features of the eight attention heads are concatenated and then passed through a linear transformation layer and a residual connection to obtain the dynamic feature vector of the sugar infusion process. .

3. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 1, characterized in that: The drying process parameters are first normalized during preprocessing, and then the normalized drying parameters are constructed into a spatiotemporal diagram structure: ； in Represents the set of nodes in a dry indoor space. For the number of sensors, Represents spatial connection edges between nodes. Represents parameters of the gradient drying process , , The graph node feature matrix formed after normalization.

4. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 3, characterized in that: The spatiotemporal graph convolutional network for the dried fruit process includes a spatial feature extraction layer and a temporal feature extraction layer. The spatial feature extraction layer uses a graph convolutional network to capture the spatial dependencies between nodes, thus constructing a graph structure feature matrix. The input is given to the spatial feature extraction layer, and the output is the spatial feature vector at time t. The temporal feature extraction layer uses a one-dimensional convolutional network to capture the temporal evolution of parameters, arranging the spatial features at different times into a three-dimensional tensor in a time series. Input one-dimensional convolutional layer to compute time features ,in Represents the temporal convolution weight matrix. Indicates the first The temporal feature extraction layer outputs the data; spatial and temporal features are fused through residual connections and batch normalization layers to obtain the spatiotemporal features of the drying process. .

5. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 1, characterized in that: The fruit preserve quality evolution gated recurrent unit network adds a quality parameter attention gated module to the traditional gated recurrent unit to enhance the temporal characteristic response of key quality parameters. The fruit preserve quality evolution gated loop unit network regulates the historical quality time sequence state through reset gates and update gates. The degree of forgetting and retention, combined with the standardized quality parameter sequence at the current moment. Calculate the temporal state of candidate qualities; The quality parameter attention gating module dynamically calculates the attention weight vector based on the importance of moisture, sugar content, firmness, and color to the final product. And use this weight to optimize the update of the candidate quality time series status; Final quality timing status The calculation integrates the update gate output and the weighted candidate quality time-series states, and its output is the time-series feature of the quality parameters. .

6. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 1, characterized in that: The fruit preserve preparation proximal policy optimization reinforcement learning model first obtains a high-quality initial policy by supervising the policy network of the model. The supervised pre-training part uses the control parameter target value labeled in the dataset as the supervision signal, takes the system state representation vector as the input, and takes the control parameter target value as the target output to minimize the error between the predicted value and the control parameter target value. The state space of the model is defined as That is, the set of multimodal fused state features at all times; Action space is defined as ,in Indicates the first Action vectors at each time step; probability distribution of actions output by the model's policy network after supervised pre-training. ,in The policy network parameters are represented by a reparameterization technique to ensure gradient differentiability; the model's value network is used to estimate state values. ,in Represents the value network parameters used to calculate the advantage function. ,in Indicates the first Instant rewards for each moment This represents the discount factor, used to balance immediate rewards with long-term rewards.

7. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 6, characterized in that: The reward function of the proximal strategy optimization reinforcement learning model for dried fruit preparation includes: preparation efficiency reward. Used to incentivize shorter preparation cycles; product quality rewards Used to incentivize products to meet quality standards; energy consumption rewards Used to incentivize energy reduction; process constraint rewards This is used to restrict violations of regulations; The model continuously interacts with the fruit preserve preparation simulation platform, collecting state-action-reward data and updating network parameters using a proximal policy optimization objective function. By alternately updating the policy network and value network, the model learns the optimal adaptive control strategy and outputs the optimal control parameters. .

8. The method for optimizing and controlling vacuum sugar infiltration and gradient drying parameters for dried fruit preparation as described in claim 1, characterized in that: After the reinforcement learning model for optimizing the proximal strategy in the dried fruit preparation process is deployed, it continuously collects real-time quality parameter data and control output data during the dried fruit preparation process, and calculates the deviation between the real-time quality parameters and the preset target values. ,when When the preset deviation threshold is exceeded, or when changes in raw material type or environmental conditions are detected, a dynamic data acquisition mechanism is triggered to collect multi-source parameter data and status characteristics under this operating condition. Controlling actions and reward value This forms a small-sample fine-tuning dataset. ; When obtaining a small sample fine-tuning dataset Then, the outer loop of meta-learning is fine-tuned, first based on the learning rate of the inner loop. The model parameters are initially updated, and then the learning rate is adjusted via the outer loop. The parameters are updated a second time to obtain the fine-tuned model parameters. .

Images