A drying scheme decision method and system for improving the quality of agricultural products

Abstract

The present application relates to the technical field of agricultural products, and in particular to a drying scheme decision-making method and system for improving the quality of agricultural products, comprising: collecting product state information and environmental parameters of target agricultural products in a drying process in real time, fusing the product state information and the environmental parameters, and generating quality metabolic flow data; inputting the quality metabolic flow data into a preset digital twin model, performing multi-step forward deduction, and obtaining a decision space; performing optimization in the decision space to generate optimized drying control parameters at the current time; and controlling a drying device to perform drying operations on the target agricultural products according to the optimized drying control parameters, collecting new product state information of the target agricultural products after the drying operations, and updating the quality metabolic flow data and the preset digital twin model according to the new product state information, thereby solving the problem that the existing drying process cannot be dynamically and adaptively adjusted according to the agricultural products, resulting in a mismatch between the control parameters and the actual state of the agricultural products, and making it difficult to continuously achieve quality optimization.

Description

Technical Field

[0001] This invention relates to the field of agricultural product technology, and specifically to a method and system for making decisions on drying schemes to improve the quality of agricultural products. Background Technology

[0002] In the field of agricultural product processing, drying is a key link that affects the quality, energy consumption and efficiency of the final product. Its core role is to preserve the nutritional and functional qualities of agricultural products to the greatest extent possible while achieving long-term preservation by precisely controlling the dehydration process.

[0003] Currently, decision-making methods for improving agricultural product quality through drying mainly rely on setting static process parameters based on fixed mathematical models or historical experience, and using sensors to monitor temperature and humidity for feedback control during the drying process. However, because the decisions are based on pre-set static models or empirical rules, it is impossible to adjust the decisions according to the real-time multi-dimensional quality changes of agricultural products during the drying process. This leads to a mismatch between control parameters and the actual state of agricultural products, making it difficult to continuously achieve the goal of optimizing quality under dynamic external environments and individual differences in agricultural products. Summary of the Invention

[0004] To address the technical problem that existing drying decision-making methods cannot dynamically and adaptively adjust based on the agricultural products during the drying process, resulting in a mismatch between control parameters and the actual state of the agricultural products, and making it difficult to continuously achieve optimal quality, this application provides a drying scheme decision-making method and system to improve the quality of agricultural products.

[0005] The drying scheme decision-making method and system for improving the quality of agricultural products provided in this application adopts the following technical solution:

[0006] A method for deciding on drying schemes to improve the quality of agricultural products, comprising:

[0007] Real-time collection of product status information and environmental parameters of target agricultural products during the drying process, and fusion of product status information and environmental parameters to generate quality metabolic flow data;

[0008] The quality metabolic flow data is input into a preset digital twin model, and a multi-step forward extrapolation is performed to obtain the decision space;

[0009] The optimal drying control parameters for the current moment are generated by searching the decision space.

[0010] Based on optimized drying control parameters, the drying equipment is controlled to perform drying operations on the target agricultural products. After the drying operation, the new product status information of the target agricultural products is collected, and the quality metabolic flow data and preset digital twin model are updated based on the new product status information.

[0011] Furthermore, the steps for fusing product status information and environmental parameters to generate quality metabolic flow data include:

[0012] Based on a pre-defined quality association knowledge graph, spatiotemporal alignment and association analysis of product status information and environmental parameters are performed to obtain a multidimensional physical property matrix.

[0013] By performing weighted fusion and dimensionality reduction on the multidimensional property matrix, quality metabolic flow data that characterizes the quality migration and transformation state of the target agricultural product during the drying process is generated.

[0014] Furthermore, based on a pre-defined quality association knowledge graph, the steps for performing spatiotemporal alignment and association analysis on product status information and environmental parameters to obtain a multidimensional property matrix include:

[0015] Based on spatiotemporal mapping rules, timestamp alignment and spatial coordinate unification are performed on product status information and environmental parameters to obtain multi-source data streams;

[0016] Based on product association rules, graph structure association reasoning is performed on multi-source data streams to obtain an association feature set;

[0017] The associated feature set is embedded into a preset graph embedding model for processing to obtain a multidimensional property matrix.

[0018] Furthermore, the process of inputting quality metabolic flow data into a pre-defined digital twin model and performing multi-step forward inference to obtain the decision space includes the following steps:

[0019] After synchronizing the state of the preset digital twin model based on the quality metabolic flow data, the preset digital twin model is driven to explore multiple preset drying parameter adjustment sequences in parallel in the future time period, starting from the current drying parameters, and generate multiple initial inference paths according to the preset collaborative inference strategy.

[0020] By using the meta-decision maker in the preset digital twin model, multiple initial deduction paths are simulated and verified for constraints. After obtaining candidate deduction paths that meet the preset feasibility conditions, several candidate deduction paths are constructed into a decision space.

[0021] Furthermore, starting with the current drying parameters, the steps of exploring multiple preset drying parameter adjustment sequences in parallel to generate multiple initial derivation paths include:

[0022] Based on the current drying parameters and the model characteristics of the preset digital twin model, the integrated decision generator is invoked to output multiple preset drying parameter adjustment sequences in parallel.

[0023] After path enhancement of multiple preset drying parameter adjustment sequences, based on preset perturbation rules and historical drying parameter adjustment sequences, each path-enhanced preset drying parameter adjustment sequence is mutated and recombined to obtain optimized drying parameter adjustment sequences covering different optimization directions.

[0024] Each optimized drying parameter adjustment sequence is input into a preset digital twin model and extrapolated over a future time period to obtain an initial extrapolation path corresponding to each optimized drying parameter adjustment sequence.

[0025] Furthermore, the steps for optimizing the drying control parameters in the decision space to generate the optimal drying control parameters for the current moment include:

[0026] Based on a preset multi-objective evaluation function, each candidate deduction path in the decision space is evaluated to obtain the utility value of each candidate deduction path;

[0027] After each agent sorts the utility values ​​and generates multiple sets of path preference rankings, multiple rounds of iterative weighted voting are performed on the multiple sets of path preference rankings to update the consensus score of each agent for each candidate deduction path until the preset convergence condition is met. The candidate deduction path corresponding to the consensus score higher than the preset consensus score is determined as the target deduction path.

[0028] Extract the target drying parameter adjustment sequence corresponding to the current moment from the target simulation path, and determine the drying parameters in the target drying parameter adjustment sequence as the optimized drying control parameters.

[0029] Furthermore, the steps of updating the quality metabolic flow data and the preset digital twin model based on the new product status information include:

[0030] Calculate the multi-dimensional state deviation data between the new product state information and the deduced product state information represented by the target deduction path. Based on the pre-set reliability evaluation rules, the multi-dimensional state deviation data is weighted and processed to generate incremental quality metabolic flow data.

[0031] Integrate incremental quality metabolic flow data with quality metabolic flow data to generate updated quality metabolic flow data;

[0032] Based on incremental quality metabolic flow data, the preset digital twin model is corrected through incremental learning to obtain an updated preset digital twin model.

[0033] This application also provides a drying scheme decision system for improving the quality of agricultural products, including:

[0034] The data generation module is used to collect product status information and environmental parameters of the target agricultural products in real time during the drying process, and to fuse the product status information and environmental parameters to generate quality metabolic flow data.

[0035] The data extrapolation module is used to input quality metabolic flow data into a preset digital twin model, perform multi-step forward extrapolation, and obtain the decision space;

[0036] The data optimization module is used to perform optimization in the decision space and generate the optimal drying control parameters for the current moment.

[0037] The control update module is used to control the drying equipment to perform drying operations on the target agricultural products based on optimized drying control parameters, collect the new product status information of the target agricultural products after the drying operation, and update the quality metabolic flow data and preset digital twin model based on the new product status information.

[0038] Beneficial effects achieved:

[0039] This application provides a decision-making method for drying schemes to improve the quality of agricultural products, including: real-time collection of product status information and environmental parameters of the target agricultural product during the drying process, and fusion of the product status information and environmental parameters to generate quality metabolic flow data; inputting the quality metabolic flow data into a preset digital twin model, performing multi-step forward extrapolation to obtain a decision space; performing optimization in the decision space to generate optimized drying control parameters for the current moment; controlling the drying equipment to perform drying operations on the target agricultural product based on the optimized drying control parameters, collecting new product status information of the target agricultural product after the drying operation, and updating the quality metabolic flow data and the preset digital twin model based on the new product status information.

[0040] In this application, the problem of existing drying decision-making methods being unable to dynamically and adaptively adjust is solved because it collects and integrates product status information and environmental parameters in real time to generate quality metabolic flow data. This quality metabolic flow data represents the real state of the target agricultural product in real time, providing a dynamic perception basis for decision-making. By inputting the quality metabolic flow data into a preset digital twin model for multi-step forward extrapolation, it is possible to simulate and extrapolate the quality evolution under various future control strategies based on the current state of the target agricultural product, generating a decision space containing multiple decision paths. This transforms the decision basis from static preset to dynamic extrapolation based on real-time state. Optimization is performed in the decision space to generate optimized drying control parameters for the current moment, ensuring that the control parameters are dynamically generated optimal solutions based on the current state of the agricultural product and after evaluating future effects, achieving real-time matching between control parameters and the actual state of the agricultural product. After executing the drying operation based on the optimized drying control parameters, new product status information is collected and the quality metabolic flow data and preset digital twin model are updated. This closed-loop feedback enables the system to continuously correct the errors of the preset digital twin model and make the next round of decisions based on the latest state, thereby achieving continuous adaptive adjustment and quality optimization in the changing drying process. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the steps of a drying scheme decision-making method for improving the quality of agricultural products according to this application.

[0042] Figure 2 A flowchart illustrating the steps involved in generating quality metabolic stream data for this application;

[0043] Figure 3 A flowchart illustrating the steps involved in generating the decision space based on quality metabolic flow data for this application;

[0044] Figure 4 This is a schematic diagram illustrating the steps of obtaining the optimized drying control parameters at the current moment based on the decision space in this application;

[0045] Figure 5 This is a schematic diagram illustrating the steps involved in updating the quality metabolic flow data and the preset digital twin model in this application.

[0046] Figure 6 This is a schematic diagram of the decision-making system for improving the quality of agricultural products in this application.

[0047] Explanation of reference numerals in the attached figures:

[0048] 10. Data generation module; 20. Data deduction module; 30. Data optimization module; 40. Control update module. Detailed Implementation

[0049] The following combination Figures 1-6 This application will be described in further detail.

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0052] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0053] This application discloses a method for deciding on drying schemes to improve the quality of agricultural products.

[0054] Please refer to Figure 1 The drying scheme decision-making method for improving agricultural product quality proposed in this embodiment includes steps S10~S40:

[0055] Step S10: Collect product status information and environmental parameters of the target agricultural product in real time during the drying process, and fuse the product status information and environmental parameters to generate quality metabolic flow data.

[0056] This method collects real-time product state information and environmental parameters of target agricultural products during the drying process. This overcomes the limitations of traditional methods that rely on single or limited static parameters. By simultaneously acquiring product state information including internal moisture distribution, spatial distribution of surface and internal chemical components, and macroscopic morphological changes, and combining this with environmental parameters, a comprehensive data source is constructed that reflects the dynamic changes in the physical properties of the target agricultural product during drying. Based on this objective, product state information and environmental parameters are fused to generate quality metabolic flow data. This transforms the real-time monitored data into data that characterizes the dynamic migration and transformation of the target agricultural product's quality. This elevates the traditionally isolated product state information and environmental parameters to a continuous tracking of the target agricultural product's quality formation process. It provides dynamic quality state input for subsequent pre-designed digital twin models and decision-making processes, enabling the entire decision-making system to operate based on an understanding of the target agricultural product's real-time state rather than relying on external indirect parameters. This lays the data foundation for achieving real-time, accurate matching and adaptive optimization of control parameters and the actual state of the agricultural product.

[0057] The system collects product status information and environmental parameters by deploying an integrated multimodal sensor array on the drying equipment. Specifically, a low-field nuclear magnetic resonance sensor, deployed in a dedicated detection chamber adjacent to the agricultural product conveying path of the drying equipment, monitors the internal moisture distribution. A hyperspectral imaging system, deployed on a sealed observation window on the side wall of the drying equipment chamber and facing the surface of the agricultural product, scans and acquires the spatial distribution of chemical components on the surface and inside of the agricultural product. Multi-angle machine vision units, deployed at multiple observation points on the top and sides of the drying equipment chamber, capture macroscopic morphological change information. Environmental parameters are simultaneously measured by temperature and humidity sensors, wind speed sensors, and thermal radiation sensors deployed at the air inlet, agricultural product area, and return air inlet of the drying chamber.

[0058] Step S20: Input the quality metabolic flow data into the preset digital twin model, perform multi-step forward extrapolation, and obtain the decision space.

[0059] Quality metabolic flow data is input into a pre-defined digital twin model and subjected to multi-step forward extrapolation. By utilizing the pre-defined digital twin model, the future dynamic evolution process of the target agricultural product based on quality metabolic flow data is reproduced and pre-simulated in a virtual space, overcoming the limitation of traditional methods that can only react to the current instantaneous state. To this end, through multi-step forward extrapolation, the pre-defined digital twin model is driven to simulate the possible change trajectory of quality metabolic flow data caused by applying different drying control parameter sequences in future time periods, based on its embedded drying dynamics, quality evolution, and constraint rules, starting from the current moment, thereby generating a decision space containing multiple decision paths.

[0060] In this step, the decision-making basis is shifted from a passive response to the current state to an assessment of future possibilities. This allows the decision-making system to obtain the impact of different drying operations on the target agricultural product before the actual drying operation, thus providing a basis for subsequent optimization decisions.

[0061] Step S30: Optimize in the decision space to generate the optimized drying control parameters for the current moment.

[0062] From the decision space generated by the above steps, which contains multiple possible future evolution trajectories (i.e., decision paths), the optimal decision path is selected by comprehensively balancing multiple objectives such as quality, energy efficiency, and time, thereby generating the optimized drying control parameters for the current moment.

[0063] The specific role of optimization in a decision space composed of multiple decision paths lies in comprehensively evaluating and comparing the prediction results of each decision path through methods such as multi-agent negotiation or multi-objective optimization algorithms. This allows for the identification of the decision path that best meets the global optimization objective from among numerous possibilities, thus realizing the transformation from a predictive decision space to an executive decision path. This ensures that the final optimized drying control parameters are not simply selected based on a single indicator or fixed rule, but rather are the optimal solution derived from the simulation of the future state evolution of the target agricultural product and the trade-off between multiple objectives. As a result, the optimized drying control parameters at each moment can accurately adapt to the real-time state and process objectives of the target agricultural product, ensuring the stable improvement of the quality of the target agricultural product and the optimal achievement of the process objectives.

[0064] Step S40: Control the drying equipment to perform drying operation on the target agricultural product according to the optimized drying control parameters, collect the new product status information of the target agricultural product after the drying operation, and update the quality metabolic flow data and preset digital twin model according to the new product status information.

[0065] Controlling the drying equipment to perform the drying operation on the target agricultural product based on optimized drying control parameters transforms the optimized drying control parameters obtained from the previous step based on prediction and optimization into actual physical action, thereby achieving precise control of the drying process. Collecting new product status information of the target agricultural product after the drying operation is crucial for obtaining real-time feedback on the target agricultural product's status after the drying operation, in order to evaluate the control effect and capture the difference between the actual response and the prediction. Furthermore, updating the quality metabolic flow data and the preset digital twin model based on the new product status information forms a complete closed loop from perception, decision-making, execution to feedback and relearning. This closed loop can correlate the actual results of each drying operation with... The prediction results of the preset digital twin model are compared, and the generated deviation data is used to continuously calibrate the quality metabolic flow data to improve its representation accuracy. The preset digital twin model is driven to perform adaptive learning to correct its internal prediction bias, so that the entire decision-making system has dynamic evolution and self-optimization capabilities. It can continuously reduce the cognitive error of the actual state and future evolution of the target agricultural product, thereby ensuring that the decision-making system can make judgments based on more accurate cognition and more reliable models in each subsequent decision-making cycle. Finally, in the dynamically changing drying process, the drying control parameters are continuously and accurately matched with the actual state of the target agricultural product, and long-term stability of quality optimization is achieved.

[0066] It should be noted that the optimized drying control parameters are sent in real time to the underlying programmable logic controller (PLC) of the drying equipment via the standard communication interface. The PLC parses these parameters and converts them into control signals for various physical execution units, such as electric heating elements, steam humidification valves, and variable frequency fans. This drives the physical execution units to adjust their outputs in the next control cycle, thereby physically altering the thermal, humidity, and airflow environment within the drying chamber to achieve the drying operation of the target agricultural product. Following the execution of the control command, after a brief delay, the same integrated multimodal sensor array deployed on the drying equipment is triggered to perform a new, complete scan and measurement of the target agricultural product using the same synchronous acquisition mechanism as in step S10. The low-field nuclear magnetic resonance sensor acquires new internal moisture distribution patterns, the hyperspectral imaging system acquires new surface and internal chemical composition spatial distribution patterns, and the multi-angle machine vision unit acquires new macroscopic morphological change information. These newly acquired data, along with the latest environmental parameters measured by environmental sensors, are packaged together as new product status information.

[0067] In one feasible implementation, refer to Figure 2 As shown, step S10 may specifically include steps S11 to S12:

[0068] Step S11: Based on the preset quality association knowledge graph, perform spatiotemporal alignment and association analysis on product status information and environmental parameters to obtain a multidimensional physical property matrix.

[0069] It should be noted that the preset quality association knowledge graph includes spatiotemporal mapping rules and product association rules.

[0070] Based on a pre-defined quality association knowledge graph, spatiotemporal alignment and correlation analysis of product status information and environmental parameters are performed. The fundamental purpose is to solve the data fragmentation problem caused by asynchronous and mismatched collection time and spatial location of product status information and environmental parameters, and to deeply reveal the complex interaction mechanism between various quality dimensions of agricultural products hidden behind product status information and environmental parameters.

[0071] Among them, the spatiotemporal mapping rules are a set of rules used to align product status information and environmental parameters asynchronously collected by different sensors in time and space, and to uniformly register them in the spatial coordinate system. Its function is to build a spatiotemporally consistent observation framework to ensure that product status information and environmental parameters can accurately correspond to the same moment and the same agricultural product location. The product association rules are a set of logical rules in the pre-set quality association knowledge graph used to describe the mutual influence and evolution of different quality indicators, such as moisture, color, and nutritional components, as well as their interaction with environmental parameters, such as temperature and humidity. Its function is to perform graph structure association analysis on the aligned product status information and environmental parameters based on domain knowledge, thereby extracting a set of association features that reflect the coupling relationship between various agricultural product quality dimensions.

[0072] Specifically, spatiotemporal alignment of product status information and environmental parameters aims to unify product status information and environmental parameters from different sources and spatiotemporal benchmarks into a consistent spatiotemporal coordinate system. This ensures that product status information and environmental parameters can accurately and synchronously reflect the status of the same agricultural product at the same time and location. Correlation analysis, on the other hand, is based on predefined product association rules in a pre-defined quality association knowledge graph. It involves graph structure reasoning on the aligned product status information and environmental parameters to explicitly extract the association feature set that reflects the coupling relationship between various agricultural product quality dimensions.

[0073] In this step, product status information and environmental parameters are converted into a structured feature expression—a multidimensional property matrix—that is a standardized vector representation containing the potential laws of agricultural product quality evolution and reflecting the dynamic changes in its quality indicators. This provides input for subsequent data fusion and model calculation, establishing a computable data representation foundation for the entire decision-making system that has a deep understanding of the intrinsic mechanism of the drying process. This enables subsequent calculations based on a quantitative understanding of the intrinsic correlation laws of agricultural product quality, greatly enhancing the depth of state perception and the reliability of decision reasoning.

[0074] Furthermore, step S11 may also include steps S111 to S113:

[0075] Step S111: Based on the spatiotemporal mapping rules, the product status information and environmental parameters are timestamped and their spatial coordinates are unified to obtain a multi-source data stream.

[0076] First, a unified high-precision clock source is established for all sensors within the decision-making system, and product status information and environmental parameters are timestamped with the time from this clock source. Next, based on spatiotemporal mapping rules, the product status information and environmental parameters are synchronized using a time base. This means that data from different sampling frequencies are unified onto the same high-resolution time series through timestamping, ensuring that product status information and environmental parameters describing the same physical moment strictly correspond on the timeline. Then, regarding the unification of spatial coordinates, a global spatial coordinate system is established based on the core motion trajectory of agricultural products within the drying equipment, according to the spatiotemporal mapping rules. The local measurement coordinates defined by various sensors are mapped to the global spatial coordinate system through a pre-calibrated transformation matrix. This ensures that the product status information and environmental parameters describing the physical location of the same agricultural product also strictly correspond in spatial coordinates. This integrates the previously discrete and heterogeneous product status information and environmental parameters in time and space into a multi-source data stream that is strictly synchronized in time and space and accurately corresponds in location. This fundamentally eliminates the problem of data spatiotemporal misalignment and fragmentation caused by asynchronous sensor sampling and different observation perspectives and positions, providing a data foundation for mining the physical correlation and causal relationship between product status information and environmental parameters in subsequent steps.

[0077] Step S112: Based on the product association rules, perform graph structure association reasoning on the multi-source data stream to obtain the association feature set.

[0078] First, each data point in the spatiotemporally aligned multi-source data stream—for example, the moisture value at a specific spatial coordinate at a certain moment, the color spectral vector of a point, or the temperature value of a region—is instantiated as an entity node or attribute node in a predefined quality association knowledge graph. Then, based on the logical relationships in the product association rules—such as the effect of moisture migration rate on color change, or localized high temperatures causing changes in enzyme activity and thus affecting nutritional components—directed edges with clear semantics, such as "affects," "causes," and "related to," are established between nodes. This constructs a dynamic, temporary inference graph that reflects the relationships between each data point. Based on this temporary inference graph, the node information is propagated, aggregated, and iteratively updated along the directed edges through graph traversal or the message passing mechanism of a graph neural network. This ensures that the feature vector of each node not only contains its own observation value but also incorporates information from other nodes with which it has a rule-defined relationship. This process is known as graph-structured association inference. Finally, the updated feature vectors of each inference node are summarized to obtain the associated feature set. The product association rules are deeply integrated with real-time observation data, enabling the decision system to automatically derive high-order features that characterize the coupling relationship between the quality dimensions of various agricultural products, thereby providing the input of association semantics for the preset graph embedding model.

[0079] Step S113: The associated feature set is embedded into a preset map embedding model for processing to obtain a multidimensional property matrix.

[0080] It should be noted that the preset graph embedding model is a pre-trained computational model that can map the graph structure obtained in step S112, i.e. the associated feature set, into a low-dimensional dense vector, i.e., a multi-dimensional property matrix.

[0081] Each feature vector in the associated feature set, existing as graph nodes, is fed into a pre-defined graph embedding model. This model first uses a graph attention layer to calculate the attention weights between different node features based on the product association rules defined in the graph. This weighted aggregation of each node's feature vector ensures that it not only contains its own information but also incorporates information from its neighboring nodes. Then, the resulting sequence of node feature vectors is input into the encoding and transformation layer of the pre-defined graph embedding model. This layer uses a feedforward neural network to map all node feature vectors into a unified, low-dimensional continuous vector space. Finally, it aggregates all node feature vectors from the associated feature set and outputs a fixed-dimensional multidimensional property matrix.

[0082] Step S12 involves weighted fusion and dimensionality reduction of the multidimensional property matrix to generate quality metabolic flow data that characterizes the quality migration and transformation state of the target agricultural product during the drying process.

[0083] A weighted fusion layer based on an attention mechanism processes the multidimensional property matrix. This layer, using pre-defined or online-learned weight coefficients, weights and aggregates the feature vectors of different dimensions within the matrix based on their importance. This highlights the dominant features crucial to the quality evolution of agricultural products during the current drying stage while suppressing secondary or noisy features, resulting in a weighted, fused high-dimensional feature representation. Next, this weighted, fused high-dimensional feature representation is input into a dimensionality reduction encoder, mapping it to a lower-dimensional vector space. This ensures that the newly generated low-dimensional vector captures and retains the most critical and significant differences and changes in the high-dimensional feature representation. Finally, the low-dimensional vectors are concatenated in chronological order of generation to form quality metabolic flow data. This transforms the original high-dimensional, sparse, and potentially redundant multidimensional property matrix into a highly condensed quality metabolic flow data that can comprehensively and dynamically represent the core laws of intrinsic migration and transformation of agricultural product quality in real time. This provides highly efficient and semantically rich input data for the subsequent pre-defined digital twin model, significantly improving the model's computational efficiency and robustness of state representation.

[0084] In one feasible implementation, refer to Figure 3 As shown, step S20 may specifically include steps S21 to S22:

[0085] Step S21: After synchronizing the state of the preset digital twin model based on the quality metabolic flow data, the preset digital twin model is driven to explore multiple preset drying parameter adjustment sequences in parallel in the future time period, starting from the current drying parameters, and generate multiple initial inference paths according to the preset collaborative inference strategy.

[0086] It should be noted that the preset collaborative inference strategy is a set of control rules that are pre-set to coordinate the various sub-models in the preset digital twin model and generate multiple inference paths in parallel.

[0087] The state of the preset digital twin model is synchronized based on the quality metabolic flow data to calibrate the real-time state of the preset digital twin model and the target agricultural product, so as to ensure the absolute accuracy of the starting point of the extrapolation, overcome the problem of prediction distortion caused by the initial state deviation of the preset digital twin model, and lay a reliable initial condition for subsequent extrapolation.

[0088] Based on this, the preset digital twin model is driven by the preset collaborative inference strategy to explore multiple preset drying parameter adjustment sequences in parallel in the future time period, starting from the current drying parameters. This is used to simulate the future evolution of agricultural product quality and process under different control strategies, thereby generating multiple initial inference paths that include the predicted agricultural product quality status and the corresponding drying control parameter sequences.

[0089] By upgrading decision support from passive response to proactive planning, and by building a future scenario library covering different control logics and optimization directions through the parallel computing capabilities of a pre-set digital twin model, an evaluation and selection basis is provided for subsequent real-time multi-objective optimization decisions.

[0090] Furthermore, step S21 may also include steps S211 to S213:

[0091] Step S211: Based on the model characteristics of the current drying parameters and the preset digital twin model, the integrated decision generator is invoked to output multiple preset drying parameter adjustment sequences in parallel.

[0092] It should be noted that the current drying parameters refer to a set of control settings that the drying equipment is currently executing, including but not limited to key process parameters such as temperature, humidity, and wind speed. The integrated decision generator is a software module that integrates multiple strategy generation algorithms, such as predictive control algorithms based on preset digital twin models, heuristic rule bases, and deep reinforcement learning policy networks. It can generate diverse strategy suggestions by calling different algorithms in parallel based on the input state and model knowledge.

[0093] The current drying parameters, along with model features extracted from a preset digital twin model, such as the current drying stage and model sensitivity information, are input into an integrated decision generator. Each algorithm module within this generator then independently performs calculations based on the input data, outputting a series of specific drying parameter adjustment schemes for multiple future time steps. Each drying parameter adjustment scheme is a parameter change sequence arranged in chronological order, i.e., a preset drying parameter adjustment sequence. Because it operates in parallel, the integrated decision generator can simultaneously output multiple preset drying parameter adjustment sequences that differ in adjustment magnitude, change trend, or optimization focus. This significantly improves the initial efficiency and diversity of strategy exploration. At the very beginning of the decision loop, it can deploy a set of adjustment sequences covering different control approaches based on the current state, effectively avoiding the problems of single strategies or slow generation in traditional methods.

[0094] Step S212: After path enhancement of multiple preset drying parameter adjustment sequences, based on preset perturbation rules and historical drying parameter adjustment sequences, each path-enhanced preset drying parameter adjustment sequence is mutated and recombined to obtain optimized drying parameter adjustment sequences covering different optimization directions.

[0095] It should be noted that the preset disturbance rules define the maximum allowable adjustment range, adjustment frequency, and linkage constraints between different parameters for various drying parameters, such as temperature and humidity, within a single time step.

[0096] First, a lightweight prospective evaluation network is used to simulate and score each preset drying parameter adjustment sequence. Based on the score, small deterministic optimization adjustments are made to the key time step parameters in each preset drying parameter adjustment sequence. For example, the adjustment step size of the drying parameter is automatically increased for periods when the quality of agricultural products changes slowly in the score prediction to enhance the exploratory nature. This generates a path-enhanced preset drying parameter adjustment sequence for each preset drying parameter adjustment sequence, i.e., a path-enhanced preset drying parameter adjustment sequence. Next, according to preset perturbation rules and historical drying parameter adjustment sequences, each path-enhanced preset drying parameter adjustment sequence is mutated and recombined. Specifically, the preset perturbation rules define the random perturbation amplitude and probability of each parameter within the legal range. Based on this, the random time point parameter values ​​of the path-enhanced preset drying parameter adjustment sequence are superimposed and perturbed to generate a mutated sequence. At the same time, historical high-quality segments that are complementary to the current mutated sequence are matched from the historical drying parameter adjustment sequences. Historical experience is injected through segment replacement operations to generate an optimized drying parameter adjustment sequence.

[0097] This process executes each preset drying parameter adjustment sequence independently, ensuring that each preset drying parameter adjustment sequence input can produce an optimized drying parameter adjustment sequence that integrates self-enhancement, random exploration, and historical experience. The generated optimized drying parameter adjustment sequence not only retains the diversity of preset drying parameter adjustment sequences, but also significantly improves the targeting and effectiveness of each preset drying parameter adjustment sequence.

[0098] Among them, the historical drying parameter adjustment sequence is a complete sequence of effective drying parameters that have been recorded and verified in the past drying process and change over time; the historical high-quality segment is a continuous subsequence extracted from the historical drying parameter adjustment sequence, which shows outstanding performance in terms of quality improvement, energy efficiency and other aspects.

[0099] Step S213: Input each optimized drying parameter adjustment sequence into the preset digital twin model and perform extrapolation over a future time period to obtain the initial extrapolation path corresponding to each optimized drying parameter adjustment sequence.

[0100] It should be noted that the computational sub-models included in the preset digital twin model are the drying kinetics model and the quality evolution model. The drying kinetics model is used to calculate the transfer and transformation process of physical quantities such as moisture and heat, while the quality evolution model calculates the real-time changes of quality indicators such as color and nutritional components based on the results of the kinetic process.

[0101] Each optimized drying parameter adjustment sequence is treated as an independent control command input and passed to the simulation instance replicated from the preset digital twin model. After receiving the corresponding optimized drying parameter adjustment sequence, each simulation instance drives its internal drying kinetics model and quality evolution model to perform step-by-step calculations according to the drying control parameters that should be applied at each future simulation time step in the future time period as specified by the corresponding optimized drying parameter adjustment sequence.

[0102] Specifically, the current time is set as the initial time step of the simulation (t=0), and the current state of the preset digital twin model, i.e., the state of agricultural products represented by quality metabolic flow data, is used as the initial state vector. Simultaneously, an optimized drying parameter adjustment sequence is input, which is a predetermined control parameter, such as temperature, for the next N time steps (from t=0 to t=N). ,humidity Wind speed The instruction set, for each future time step t, uses the state vector calculated in the previous time step. and the control parameter command set at the current time step ( , , Input the drying kinetics model, solve the heat and mass transfer physical equations, and calculate the physical state of the agricultural product at the end of time step t under the current control parameters. This mainly includes moisture distribution, temperature distribution, etc., and is denoted as the physical state vector. Next, the physical state vector The quality state vector of the previous time step And the current control parameter input quality evolution model, which is based on mechanisms such as chemical reaction kinetics and enzyme activity mechanics or data-driven correlations, to calculate the various quality indicators, such as color value, of agricultural products at the end of time step t under the current physical conditions and control parameters. Vitamin content Predicted rehydration rate The numerical value of the quality state vector is denoted as the quality state vector. Then, the physical state With quality status Merge and update to the complete state vector of the current time step. And at the same time, record the control parameter instructions executed in this time step ( , , ).

[0103] Repeat the above steps iteratively to calculate time steps t=1,2,...,N. After N time steps are calculated, record the N state vectors in chronological order throughout the entire deduction process. ,..., The combination of the optimized drying parameter adjustment sequence and the corresponding N control parameter instructions forms a complete initial deduction path. Each optimized drying parameter adjustment sequence will independently trigger an iterative calculation as described above, thereby outputting a complete trajectory. This trajectory consists of two parts: first, the predicted agricultural product quality state sequence (i.e., N state vectors) for each future time step derived from the deduction; and second, the drying control parameter sequence (i.e., N control parameter instructions) for each future time step on which the deduction is based. The combination of these two parts constitutes the initial deduction path corresponding to the optimized drying parameter adjustment sequence.

[0104] Step S22: Through the meta-decision maker in the preset digital twin model, multiple initial deduction paths are simulated and verified for constraints. After obtaining candidate deduction paths that meet the preset feasibility conditions, several candidate deduction paths are constructed into a decision space.

[0105] It should be noted that the meta-decision maker is a high-level coordination and adjudication module built into the preset digital twin model. Its core function is to systematically evaluate all initial deduction paths based on a predefined set of rules that integrates physical feasibility, process safety and goal orientation.

[0106] The simulation verification in this step refers to the meta-decision-maker calling upon or reviewing the intermediate data generated during the derivation process of each initial derivation path. Using the verification sub-models within the pre-defined digital twin model it accesses, such as stress models and local overheating monitoring models, it performs secondary verification on potential contradictions or numerical instabilities in the initial derivation paths, such as moisture evaporation rates far exceeding heating limits. This ensures that the evolution process of each initial derivation path is self-consistent and feasible at the physical principle level. The constraint verification involves comparing the predicted agricultural product quality state sequence and drying control parameter sequence of each initial derivation path with pre-defined hard constraints, such as maximum allowable temperature, maximum energy consumption limit, and minimum final moisture content, and soft optimization objectives, such as the target range of quality indicators. Paths that fully meet all hard constraints and satisfy the soft optimization objectives are selected. Paths that pass the verification and validation are marked as candidate derivation paths.

[0107] Next, the meta-decision maker structures and encodes all candidate deduction paths. For example, each candidate deduction path is represented as a point or vector in a high-dimensional space. The set of all these points or vectors mathematically defines a multi-dimensional, non-linear region, which is constructed as the decision space for subsequent optimization. This step cleans, verifies, and organizes a large number of initial deduction paths that may contain invalid or high-risk options, thereby outputting a candidate deduction path that can be directly used for multi-objective trade-offs and comparisons, which greatly improves the efficiency of the subsequent optimization process.

[0108] In one feasible implementation, refer to Figure 4 As shown, step S30 may specifically include steps S31 to S33:

[0109] Step S31: Based on the preset multi-objective evaluation function, evaluate each candidate deduction path in the decision space to obtain the utility value of each candidate deduction path.

[0110] First, from each candidate derivation path, extract the agricultural product quality state sequence representing quality evolution and the drying control parameter sequence representing drying actions. Input these two types of data into a preset multi-objective evaluation function, which takes the form of: ,in, The sub-function processes the sequence of agricultural product quality status, calculates its time integral or weighted sum of the final value, and outputs a normalized quality score. The sub-function processes the energy consumption data and calculates the total energy consumption of the corresponding candidate deduction path by combining the heating power, fan frequency and other parameters in the control parameter sequence with time integration. The reciprocal or negative value of the total energy consumption is taken to output a normalized energy efficiency score. The sub-function processes the total time consumption, taking the reciprocal of the preset total time for the corresponding candidate derivation path to output a normalized efficiency score; and , , These are preset or dynamically adjusted weighting coefficients for quality, energy consumption, and total duration. The preset multi-objective evaluation function sums the outputs of the three sub-functions according to their respective weights, and finally calculates a utility value representing the corresponding candidate inference path.

[0111] For each candidate deduction path in the decision space, the complete data extraction and function calculation process described above is executed independently once, thereby assigning a corresponding utility value to each candidate deduction path. In this step, each candidate deduction path, which contains complex multidimensional information, namely quality, energy consumption, and total duration, is transformed into a directly comparable utility value, thus establishing a unified comparison benchmark for the subsequent decision-making process that requires comparison, ranking, and negotiation.

[0112] Step S32: After each agent sorts the utility values ​​and generates multiple sets of path preference rankings, multiple rounds of iterative weighted voting are performed on the multiple sets of path preference rankings to update the consensus score of each agent for each candidate deduction path until the preset convergence condition is met. The candidate deduction path corresponding to the consensus score higher than the preset consensus score is determined as the target deduction path.

[0113] The pre-defined digital twin model contains multiple agents, each representing an independent computing and decision-making unit with different optimization objectives, such as prioritizing quality, energy consumption, or total time. Each agent defines a dedicated evaluation function that is strongly correlated with its objective.

[0114] First, each agent encapsulates a dedicated, highly specialized mathematical evaluation function designed specifically for its chosen single optimization objective. For example, an agent prioritizing quality receives the predicted agricultural product quality state sequence from the candidate inference path as input to its evaluation function. The core of this function is a mathematical operator focused on quality indicators, calculating the final quality, average quality integral, or retention rate of specific nutrients from the predicted agricultural product quality state sequence, ultimately outputting a quality-specific score. An agent prioritizing energy efficiency receives the drying control parameter sequence from the candidate inference path as input to its evaluation function, calculates the total energy consumption of the drying control parameter sequence, and outputs an energy efficiency-specific score. An agent prioritizing efficiency directly outputs an efficiency-specific score based on the preset total duration of the candidate inference path. During evaluation, each agent, based on its specific objective, extracts the specific data sequence it is interested in from the same candidate inference path, then calls the corresponding dedicated evaluation function to calculate a specific score that reflects only the performance of the candidate inference path towards its single objective. Next, each agent sorts all candidate paths from highest to lowest score based on the specific scores it has calculated for all candidate paths, thus generating a set of path preference rankings that represent which paths are better if only the objective is considered. Multiple agents execute this process in parallel, thus generating multiple sets of path preference rankings.

[0115] Next, a multi-round iterative weighted voting process is initiated. In each round of voting, each agent votes on all candidate deduction paths according to its current path preference, based on its currently assigned dynamic voting weight (the initial weight can be equal or set according to the importance of the target). For example, the candidate deduction path ranked first receives the highest score, and the one ranked last receives the lowest score, and so on. The weighted voting scores of all agents for all candidate deduction paths are aggregated, and the consensus score of each candidate deduction path is calculated and updated. After each round of voting, the voting weights of each agent in the next round are dynamically adjusted based on the difference between the voting results of each agent in this round and the overall consensus. For example, the agent with the greater difference from the current consensus may have its voting weight appropriately reduced in the next round to promote consensus formation, and then a new round of voting is conducted based on the new weights. This iterative process continues until a preset convergence condition is met, such as when the change in consensus score over multiple consecutive rounds is less than a very small threshold, or when the maximum number of iterations is reached. The iteration then terminates, and the candidate deduction path with the highest final consensus score is determined as the target deduction path. In this way, different, even conflicting, target demands are integrated into a group decision, thereby dynamically selecting an optimal solution that can be accepted by all agents without the need for manual pre-setting of fixed compromise weights.

[0116] It should be noted that the preset consensus score in this embodiment refers to the consensus score that is ranked second highest after the preset convergence condition is met.

[0117] Step S33: Extract the target drying parameter adjustment sequence corresponding to the current moment from the target simulation path, and determine the drying parameters in the target drying parameter adjustment sequence as the optimized drying control parameters.

[0118] The determined target projection path is analyzed. In terms of data structure, this path is a list arranged sequentially according to future time steps. Each element in the list contains the predicted quality status of agricultural products and the planned drying control parameters for that time step. From the data structure of this target projection path, a subsequence completely recording the planned drying control parameters for each future time step is extracted—the target drying parameter adjustment sequence. The first element of this target drying parameter adjustment sequence is located; the parameters contained in this element are the drying parameters planned for the current moment, i.e., the next execution time step, such as the set values ​​for target temperature, target humidity, and target wind speed. These drying parameters are output from the decision logic and determined as the optimized drying control parameters to be issued to drying equipment, such as heaters, humidifiers, and variable frequency fans. This achieves the continuity of optimization decisions in the time dimension and the immediacy of control execution.

[0119] In one feasible implementation, refer to Figure 5 As shown, step S40 may specifically include steps S41 to S43:

[0120] Step S41: Calculate the multi-dimensional state deviation data between the new product state information and the deduced product state information represented by the target deduction path. Based on the preset reliability evaluation rules, perform weighted processing on the multi-dimensional state deviation data to generate incremental quality metabolic flow data.

[0121] It should be noted that the product status information represented by the target simulation path is the virtual status data predicted by the preset digital twin model at the same execution step in the target simulation path.

[0122] First, the status information of two new products and the status information of the inferred product are precisely aligned in time. The status information of the new products and the status information of the inferred product are compared dimensionally and spatially, and their absolute difference or relative error is calculated to obtain a multi-dimensional status deviation data in terms of dimensions.

[0123] Next, weighted processing is performed according to pre-defined reliability assessment rules. These rules are a set of predefined logical rules based on sensor characteristics, environmental conditions, and historical error statistics, used to evaluate the reliability of the current deviation data for each dimension. For example, it is stipulated that in a high-humidity environment, the confidence of the optical sensor decreases, and the weight of its corresponding dimension deviation should be lowered; when the short-term fluctuation of a certain dimension's data exceeds its historical normal range, its current confidence weight is also reduced accordingly. According to the pre-defined reliability assessment rules, a dynamic confidence weight is assigned to each data point in the multi-dimensional state deviation data. Then, all deviation data are weighted averaged or weighted fused to finally generate an incremental quality metabolic flow data that has been noise-suppressed and reliability-enhanced, capable of representing the state correction vector. This provides reliable input for subsequent updates to the pre-defined digital twin model and quality metabolic flow data, effectively preventing erroneous or accidental deviation data from contaminating the decision system and ensuring the correctness and stability of the self-learning evolution direction.

[0124] Step S42: Merge the incremental quality metabolic flow data and the quality metabolic flow data to generate updated quality metabolic flow data.

[0125] The recursive update algorithm based on the forgetting factor aligns the latest state vector of the quality metabolic flow data with the state vector of the incremental quality metabolic flow data. The algorithm assigns a high-weight base coefficient (e.g., 0.9) to the latest state vector of the quality metabolic flow data and a low-weight update coefficient (e.g., 0.1) to the state vector of the incremental quality metabolic flow data. These two coefficients together constitute the fusion weight, and their sum is always 1.

[0126] Next, a weighted summation is performed: the updated quality metabolic flow data vector value at the current time step = (latest state vector of quality metabolic flow data × high-weight base coefficient) + (state vector of incremental quality metabolic flow data × low-weight update coefficient). This integrates the incremental quality metabolic flow data reflecting the latest state changes into the overall quality metabolic flow data without being overly affected by single-observation noise. This ensures the continuity and stability of the state representation over time, while enabling it to continuously track and closely reflect the true dynamic changes of agricultural products.

[0127] Among them, the recursive update algorithm based on the forgetting factor is a calculation method that uses a preset weighting coefficient to linearly combine the latest state vector with the current state vector to generate the updated state vector at the current time. It can align the two because, in the design of the decision system, the incremental quality metabolic flow data is defined as the correction amount of the state vector corresponding to the latest time in the quality metabolic flow data. The two naturally correspond at the time point, so algebraic operations can be performed directly.

[0128] Step S43: Based on incremental quality metabolic flow data, the preset digital twin model is corrected through incremental learning to obtain an updated preset digital twin model.

[0129] Incremental quality metabolic flow data is used as the core training sample. This incremental quality metabolic flow data contains the prediction bias information of the preset digital twin model in the previous decision cycle.

[0130] The correction process is executed through an online incremental learning algorithm, the core of which is directed backpropagation and parameter fine-tuning. Specifically, the online incremental learning algorithm first inputs the quality metabolic stream data that generated the incremental quality metabolic stream data into the preset digital twin model to be corrected. A forward computation is performed on the preset digital twin model to obtain the predicted output at the current time step. Next, the online incremental learning algorithm uses the incremental quality metabolic stream data as the target correction direction, calculates the difference between the predicted output and the target direction, and calculates a loss function. Then, this loss function is used through a backpropagation algorithm to selectively update only the gradient-sensitive parameters in the preset digital twin model calculated with the current batch of data—usually the parameters of the last few layers or specific sub-modules—with small increments, rather than retraining the entire preset digital twin model with all quality metabolic stream data.

[0131] After the parameter fine-tuning described above, the internal mapping relationship of the preset digital twin model is slightly and directionally corrected, resulting in an updated preset digital twin model whose predictive behavior is closer to the latest observed reality of agricultural products. This allows the preset digital twin model to continuously and stably optimize its predictive accuracy based on the confidence-weighted real deviation fed back after each decision execution. This ensures that the preset digital twin model can dynamically track and adapt to time-varying factors such as changes in agricultural product characteristics and equipment performance drift, providing an increasingly reliable simulation basis for all subsequent decision cycles.

[0132] This application also provides a drying scheme decision system for improving the quality of agricultural products, referring to... Figure 6 As shown, the decision-making system for improving the drying process of agricultural products includes:

[0133] The data generation module 10 is used to collect product status information and environmental parameters of the target agricultural products in real time during the drying process, and to fuse the product status information and environmental parameters to generate quality metabolic flow data.

[0134] The data extrapolation module 20 is used to input quality metabolic flow data into a preset digital twin model, perform multi-step forward extrapolation, and obtain the decision space;

[0135] The data optimization module 30 is used to perform optimization in the decision space and generate optimized drying control parameters for the current moment.

[0136] The control update module 40 is used to control the drying equipment to perform drying operations on the target agricultural products according to the optimized drying control parameters, collect the new product status information of the target agricultural products after the drying operation, and update the quality metabolic flow data and the preset digital twin model according to the new product status information.

[0137] Optionally, the data generation module 10 is also used for:

[0138] Based on a pre-defined quality association knowledge graph, spatiotemporal alignment and association analysis of product status information and environmental parameters are performed to obtain a multidimensional physical property matrix.

[0139] By performing weighted fusion and dimensionality reduction on the multidimensional property matrix, quality metabolic flow data that characterizes the quality migration and transformation state of the target agricultural product during the drying process is generated.

[0140] Optionally, the data generation module 10 is also used for:

[0141] Based on spatiotemporal mapping rules, timestamp alignment and spatial coordinate unification are performed on product status information and environmental parameters to obtain multi-source data streams;

[0142] Based on product association rules, graph structure association reasoning is performed on multi-source data streams to obtain an association feature set;

[0143] The associated feature set is embedded into a preset graph embedding model for processing to obtain a multidimensional property matrix.

[0144] Optionally, the data extrapolation module 20 is also used for:

[0145] After synchronizing the state of the preset digital twin model based on the quality metabolic flow data, the preset digital twin model is driven to explore multiple preset drying parameter adjustment sequences in parallel in the future time period, starting from the current drying parameters, and generate multiple initial inference paths according to the preset collaborative inference strategy.

[0146] By using the meta-decision maker in the preset digital twin model, multiple initial deduction paths are simulated and verified for constraints. After obtaining candidate deduction paths that meet the preset feasibility conditions, several candidate deduction paths are constructed into a decision space.

[0147] Optionally, the data extrapolation module 20 is also used for:

[0148] Based on the current drying parameters and the model characteristics of the preset digital twin model, the integrated decision generator is invoked to output multiple preset drying parameter adjustment sequences in parallel.

[0149] After path enhancement of multiple preset drying parameter adjustment sequences, based on preset perturbation rules and historical drying parameter adjustment sequences, each path-enhanced preset drying parameter adjustment sequence is mutated and recombined to obtain optimized drying parameter adjustment sequences covering different optimization directions.

[0150] Each optimized drying parameter adjustment sequence is input into a preset digital twin model and extrapolated over a future time period to obtain an initial extrapolation path corresponding to each optimized drying parameter adjustment sequence.

[0151] Optionally, the data optimization module 30 is also used for:

[0152] Based on a preset multi-objective evaluation function, each candidate deduction path in the decision space is evaluated to obtain the utility value of each candidate deduction path;

[0153] After each agent sorts the utility values ​​and generates multiple sets of path preference rankings, multiple rounds of iterative weighted voting are performed on the multiple sets of path preference rankings to update the consensus score of each agent for each candidate deduction path until the preset convergence condition is met. The candidate deduction path corresponding to the consensus score higher than the preset consensus score is determined as the target deduction path.

[0154] Extract the target drying parameter adjustment sequence corresponding to the current moment from the target simulation path, and determine the drying parameters in the target drying parameter adjustment sequence as the optimized drying control parameters.

[0155] Optionally, the control update module 40 is also used for:

[0156] Calculate the multi-dimensional state deviation data between the new product state information and the deduced product state information represented by the target deduction path. Based on the pre-set reliability evaluation rules, the multi-dimensional state deviation data is weighted and processed to generate incremental quality metabolic flow data.

[0157] Integrate incremental quality metabolic flow data with quality metabolic flow data to generate updated quality metabolic flow data;

[0158] Based on incremental quality metabolic flow data, the preset digital twin model is corrected through incremental learning to obtain an updated preset digital twin model.

[0159] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method of decision making for a drying protocol to improve the quality of an agricultural product, characterized in that, include: Real-time data collection of product status information and environmental parameters of target agricultural products during the drying process, and fusion of the product status information and environmental parameters to generate quality metabolic flow data; The quality metabolic flow data is input into a preset digital twin model, and a multi-step forward inference is performed to obtain the decision space; The optimal drying control parameters for the current moment are generated by performing optimization in the decision space. The drying equipment is controlled to perform drying operations on the target agricultural product according to the optimized drying control parameters. The new product status information of the target agricultural product after the drying operation is collected, and the quality metabolic flow data and the preset digital twin model are updated according to the new product status information. The step of fusing the product status information and the environmental parameters to generate quality metabolic flow data includes: Based on a preset quality association knowledge graph, spatiotemporal alignment and association analysis are performed on the product status information and the environmental parameters to obtain a multidimensional physical property matrix. The multidimensional property matrix is ​​weighted, fused, and dimensionality reduced to generate quality metabolic flow data that characterizes the quality migration and transformation state of the target agricultural product during the drying process.

2. The method of claim 1, wherein the method is used for improving the quality of agricultural products. The preset quality association knowledge graph includes spatiotemporal mapping rules and product association rules. The step of performing spatiotemporal alignment and association analysis on the product status information and the environmental parameters based on the preset quality association knowledge graph to obtain a multidimensional property matrix includes: Based on the spatiotemporal mapping rules, the product status information and the environmental parameters are timestamp aligned and spatial coordinates unified to obtain a multi-source data stream. Based on the product association rules, graph structure association reasoning is performed on the multi-source data stream to obtain an association feature set; The associated feature set is embedded into a preset graph embedding model for processing to obtain the multidimensional property matrix.

3. The method of claim 1, wherein the method is used for improving the quality of agricultural products. The step of inputting the quality metabolic flow data into a preset digital twin model and performing multi-step forward inference to obtain the decision space includes: After synchronizing the state of the preset digital twin model based on the quality metabolic flow data, the preset digital twin model is driven to explore multiple preset drying parameter adjustment sequences in parallel in the future time period, starting from the current drying parameters, and generate multiple initial inference paths according to the preset collaborative inference strategy. The meta-decision maker in the preset digital twin model is used to simulate and verify the multiple initial deduction paths and check the constraints. After obtaining candidate deduction paths that meet the preset feasibility conditions, the multiple candidate deduction paths are constructed into the decision space.

4. The method of claim 3, wherein the method is used for improving the quality of agricultural products. The step of starting with the current drying parameters, exploring multiple preset drying parameter adjustment sequences in parallel, and generating multiple initial deduction paths includes: Based on the current drying parameters and the model characteristics of the preset digital twin model, the integrated decision generator is invoked to output the multiple preset drying parameter adjustment sequences in parallel. After path enhancement of the various preset drying parameter adjustment sequences, based on preset perturbation rules and historical drying parameter adjustment sequences, each path-enhanced preset drying parameter adjustment sequence is mutated and recombined to obtain optimized drying parameter adjustment sequences covering different optimization directions. Each optimized drying parameter adjustment sequence is input into the preset digital twin model, and extrapolation is performed over the future time period to obtain the initial extrapolation path corresponding to each optimized drying parameter adjustment sequence.

5. The method of claim 3, wherein the method is used for improving the quality of agricultural products. The decision space includes multiple agents, and the step of optimizing the drying control parameters in the decision space to generate the optimal drying control parameters at the current moment includes: Based on a preset multi-objective evaluation function, each of the candidate deduction paths in the decision space is evaluated to obtain the utility value of each candidate deduction path; Each of the aforementioned intelligent agents sorts the utility values ​​to generate multiple sets of path preference rankings. Then, multiple rounds of iterative weighted voting are performed on the multiple sets of path preference rankings to update the consensus score of each intelligent agent for each candidate deduction path until a preset convergence condition is met. The candidate deduction path corresponding to the consensus score higher than the preset consensus score is determined as the target deduction path. Extract the target drying parameter adjustment sequence corresponding to the current moment from the target derivation path, and determine the drying parameters in the target drying parameter adjustment sequence as the optimized drying control parameters.

6. The method of claim 5, wherein the method is used for improving the quality of agricultural products. The step of updating the quality metabolic flow data and the preset digital twin model based on the new product status information includes: Calculate the multi-dimensional state deviation data between the new product state information and the deduced product state information represented by the target deduction path, and perform weighted processing on the multi-dimensional state deviation data according to the preset reliability evaluation rules to generate incremental quality metabolic flow data. The incremental quality metabolic flow data and the quality metabolic flow data are merged to generate updated quality metabolic flow data; Based on the incremental quality metabolic flow data, the preset digital twin model is corrected through incremental learning to obtain an updated preset digital twin model.

7. A drying protocol decision system for improving the quality of an agricultural product, characterized in that, include: The data generation module is used to collect product status information and environmental parameters of the target agricultural product in real time during the drying process, and to fuse the product status information and environmental parameters to generate quality metabolic flow data. The data extrapolation module is used to input the quality metabolic flow data into a preset digital twin model, perform multi-step forward extrapolation, and obtain the decision space; The data optimization module is used to perform optimization in the decision space and generate optimized drying control parameters for the current moment; The control update module is used to control the drying equipment to perform drying operations on the target agricultural product according to the optimized drying control parameters, collect the new product status information of the target agricultural product after the drying operation, and update the quality metabolic flow data and the preset digital twin model according to the new product status information. The data generation module is also used for: Based on a preset quality association knowledge graph, spatiotemporal alignment and association analysis are performed on the product status information and the environmental parameters to obtain a multidimensional physical property matrix. The multi-dimensional physical property matrix is weighted, fused and reduced in dimension to generate the quality metabolic flow data representing the quality migration and transformation state of the target agricultural product in the drying process.

Images