Deep learning-based tea production path optimization method

Abstract

The invention discloses a tea production path optimization method based on deep learning. The method comprises the following steps: S1, collecting tea production multi-source data through a distributed sensor network; s2, performing data preprocessing on the acquired multi-source data to form a standardized data set; s3, based on the standardized data set, generating a high-order feature vector by using a space-time self-attention neural network; s4, optimizing network parameters by adopting a differential evolution algorithm; s5, constructing a dynamic multi-target production path optimization function, solving by adopting a distributed non-dominated sorting genetic algorithm III, and screening an optimal production path optimization scheme based on an attention mechanism; s6, refining and implementing an optimal production path optimization scheme; and S7, continuously optimizing the space-time self-attention neural network based on federal learning, and adaptively adjusting the tea production path optimization scheme. According to the method, self-adaptive optimization of the production path can be realized in a multi-process and complex scene, and the tea production intelligence and management level are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural product processing and industrial data analysis, and in particular to a tea production path optimization method based on deep learning. Background Art

[0002] Currently, the tea production industry is in the process of transformation to intelligent and lean management. The traditional tea production process includes multiple links such as planting, picking, transportation and processing, involving numerous raw materials, equipment resources and quality requirements. In the actual production process, there are complex spatiotemporal relationships between the various links, and they are highly sensitive to multi-source heterogeneous data such as environmental conditions, equipment status and personnel operations. However, existing path optimization methods mostly rely on manual experience scheduling and fixed process flows. It is difficult to integrate multi-source data collected by sensors in real time, and it is also impossible to characterize the dynamic changes in the production process, resulting in insufficient decision-making flexibility and global optimality. With the expansion of production scale and the increase in high-quality demand, the industry urgently needs a new generation of tea production path optimization technology that can integrate data-driven, intelligent prediction and real-time optimization.

[0003] In the existing technology, some studies have applied machine learning and deep neural networks to data analysis and yield prediction in tea production, but there are still many shortcomings. When faced with the spatiotemporal heterogeneity of production links and the complexity of process flows, traditional deep models find it difficult to deeply explore high-order correlations between multi-dimensional features, and their generalization capabilities for bottleneck identification, resource balancing, and quality prediction are limited. Mainstream model training methods mostly rely on local search algorithms such as gradient descent, which are prone to falling into local optimality and difficult to obtain global optimal solutions in high-dimensional and non-convex spaces. At the same time, they lack compatibility with loss functions and priors. In addition, multi-objective optimization often uses manually set static weights, which makes it difficult to dynamically adjust the trade-off strategy according to actual scenarios and characteristics, affecting the intelligence and adaptability of decision-making in complex production environments.

[0004] During the implementation of production plans, traditional methods rely heavily on manual monitoring or simple data feedback, making it difficult to achieve virtual-to-real mapping, error correction, and automatic adjustment of process parameters with the help of digital twins and multimodal data. This results in untimely responses to production anomalies, impacting product quality. Furthermore, the generalization and adaptability of existing models are limited, making it difficult to effectively integrate data resources from multiple tea gardens or factories. Model performance is prone to degradation when responding to environmental changes and concept drift. Furthermore, each production entity independently trains models, which poses data security risks and insufficient collaborative optimization capabilities, limiting the industry's overall continued improvement.

[0005] Therefore, how to provide a tea production path optimization method based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a tea production path optimization method based on deep learning. The present invention makes full use of the spatiotemporal self-attention neural network and combines it with multi-objective decision-making for modeling, thereby realizing intelligent modeling, global optimization and closed-loop adaptive adjustment of the entire tea production process. It has the advantages of strong data fusion capability, good path optimization effect and high adaptability.

[0007] A tea production path optimization method based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1, collect multi-source data of tea production process through distributed sensor network;

[0009] S2. Preprocess the collected multi-source data to form a standardized data set;

[0010] S3, based on the standardized dataset, uses the spatiotemporal self-attention neural network to generate high-order feature vectors;

[0011] S4. Use differential evolution algorithm to optimize the parameters of spatiotemporal self-attention neural network and minimize the total loss function through evolutionary operation;

[0012] S5. Model the tea production path based on high-order eigenvectors, construct a dynamic multi-objective production path optimization function, use a distributed non-dominated sorting genetic algorithm III to solve the Pareto optimal solution set, and screen the optimal production path optimization solution based on the attention mechanism;

[0013] S6. Refine and implement the optimal production path optimization plan, and conduct virtual and real simulation comparison through digital twins;

[0014] S7. Update the spatiotemporal self-attention neural network based on the federated learning architecture and adaptively adjust the tea production path optimization plan.

[0015] Optionally, step S1 specifically includes:

[0016] A distributed sensor network is set up in the tea production process to collect multi-source data. The tea production process includes tea planting, picking, transportation and processing. The multi-source data includes ambient temperature, humidity, light intensity, soil nutrients, pH value, equipment operating status, working hours, and logistics trajectory.

[0017] Optionally, step S2 specifically includes:

[0018] S21. Perform data cleaning on the collected multi-source data, wherein the data cleaning includes filling missing values, removing outliers, and unifying the format to obtain a cleaned data set;

[0019] S22. Standardize the cleaned dataset. The standardization process includes performing min-max normalization on the numerical data in the dataset to scale them to the interval [0, 1], and performing one-hot encoding on the categorical data in the dataset to convert each categorical category into a unique binary vector.

[0020] S23. Integrate the standardized multi-source data into a multidimensional feature data set.

[0021] Optionally, step S3 specifically includes:

[0022] S31, feeding the multidimensional feature dataset into a spatiotemporal self-attention neural network, wherein the spatiotemporal self-attention neural network includes an input layer, a local spatiotemporal perception embedding layer, a topological prior integration layer, a hierarchical causal gated self-attention layer, a dynamic bottleneck detection auxiliary branch, and an output layer;

[0023] S32, through the input layer, reorganize the multidimensional feature dataset according to production batch, time step and feature type, and convert it into an input tensor Where b represents the number of production batches, t represents the continuous time step, and f represents the feature vector dimension;

[0024] S33. Extracting spatial features and residual temporal features of the input tensor through a local spatiotemporal perception embedding layer. The local spatiotemporal perception embedding layer uses a parallel structure of two convolution kernels of different sizes to extract the spatial features of the input tensor and introduces a local autoregressive residual unit to extract the residual temporal features of the input tensor.

[0025] The convolution kernel sizes include 1×3 and 1×7. The 1×3 convolution kernel is responsible for capturing the detailed changes between a certain link and its adjacent links, while the 1×7 convolution kernel extracts the full-process collaboration and trend features across multiple links.

[0026] The local autoregressive residual unit is used to calculate the feature x of each production batch at each time step. t Make a prediction, and the predicted value is where a k is the learnable autoregressive coefficient, K is the backtracking step size, and the residual between the actual observation value and the predicted value is calculated As residual time series features;

[0027] S34, input the spatial features and residual time series features into the topological prior integration layer, and introduce the topological prior matrix T∈{0,1} f×f Constrain the attention weights between features;

[0028] The topological prior matrix T is automatically generated through the production process knowledge graph. If there is information flow interaction in the production links corresponding to features i and j, T i,j=1;

[0029] The attention weight α ij Adjust according to the following formula:

[0030]

[0031] Among them, q i is the embedding vector of the i-th feature, k j is the embedding vector of the jth feature, T i,j =1 means that features i and j are related in the process flow, T i,j =0 means no association;

[0032] S35. Input the feature representation output by the topological prior integration layer into the hierarchical causal gated self-attention layer. The hierarchical causal gated self-attention layer has three sub-layers: short-term dependency, medium-term dependency, and long-term dependency. Each sub-layer performs self-attention calculation on the input features through causal masks of different time windows. The self-attention outputs of each sub-layer are dynamically weighted by the gating function to generate gated attention features of the corresponding time scale.

[0033] S36. Classify and calculate the gated attention features through the dynamic bottleneck detection auxiliary branch to obtain a bottleneck score, and fuse the bottleneck score with the gated attention features using a residual connection method to generate an enhanced feature;

[0034] S37. The enhanced features are input into the output layer, and the features related to the optimization target are extracted through the multi-task decoding module. The features are then fused and compressed through the adaptive distillation module to generate a high-order feature vector.

[0035] Optionally, step S4 specifically includes:

[0036] The differential evolution algorithm is used to train and optimize the parameters of the spatiotemporal self-attention neural network. By performing mutation, crossover, and selection iterations on the parameter population, the total loss function is minimized. The total loss function includes production link state loss, bottleneck classification loss, feature decoupling loss, and temporal consistency loss.

[0037] The production link status loss The cross entropy between the model's output of multiple categories of states in the production process and the corresponding true category labels;

[0038] The bottleneck classification loss The cross entropy between the output probability distribution of the model and the true bottleneck state label when performing bottleneck detection tasks;

[0039] The characteristic decoupling loss To minimize the mutual information between the dimensions of the high-order eigenvectors, specifically Among them, MI(zi ,z j ) represents the mutual information between the i-th dimension and the j-th dimension high-order features;

[0040] The timing consistency loss It means constraining the changes between high-order eigenvectors of adjacent time steps to remain smooth, specifically Where z represents the high-order eigenvector, t represents the time step, and ‖·‖ represents the Euclidean norm;

[0041] The total loss function formula is:

[0042]

[0043] Among them, β,γ are adaptive weight coefficients;

[0044] Optionally, step S5 specifically includes:

[0045] S51. Model the entire tea production process based on the high-order eigenvectors, and define the decision variables in the tea production path optimization plan as X. The decision includes process sequence, resource allocation, equipment scheduling, and logistics path;

[0046] S52. Construct a dynamic multi-objective production path optimization function. The dynamic multi-objective production path optimization function comprehensively considers the overall production time, production cost, resource utilization balance, and tea quality score, and optimizes each indicator as an independent target simultaneously. Specifically,

[0047]

[0048] Where f1(X) represents the overall production time, and f2(X) represents the production cost function integrating the tea quality regulation mechanism, specifically: α is the adjustment parameter, is the minimum quality reference value, and f3(X) represents the resource utilization balance index function, specifically: σ represents the standard deviation of the waiting time, f4(X) represents the tea quality prediction function based on the high-order eigenvector, and T represents the transposed sign;

[0049] S53. Setting enhanced constraint conditions, wherein the enhanced constraint conditions include quality constraints and elastic resource constraints;

[0050] The quality constraint is: the quality prediction value of the tea production plan must be higher than the sum of the set threshold and the quality compensation amount calculated based on the high-order eigenvector;

[0051] The elastic resource constraint is: under the premise of ensuring production continuity, the dynamic elastic adjustment amount does not exceed the upper limit of the corresponding resource;

[0052] S54, using the distributed non-dominated sorting genetic algorithm III to perform global optimization on the dynamic multi-objective production path optimization function under enhanced constraints and obtain the Pareto optimal solution set;

[0053] S55. Screening the optimal production path optimization solution from the Pareto optimal solution set based on the attention mechanism X * ,

[0054]

[0055] Among them, the Pareto set represents the Pareto optimal solution set obtained, f′ i (X) represents the evaluation function value of the decision variable X on the i-th optimization objective, λ i Represents the adaptive weight coefficient corresponding to the i-th target, which is calculated by performing softmax normalization on the high-order feature vector.

[0056] Optionally, step S6 specifically includes:

[0057] S61, optimize the optimal production path plan X * It is broken down into process-level execution instruction sets, and blockchain evidence storage technology is used to perform distributed encrypted recording of each process execution status and operation process data;

[0058] S62. During the data collection process, new features are added: equipment vibration spectrum, tea phenotype images, and operator biometrics acquired in real time;

[0059] S63, based on real-time collected data and historical optimization schemes, build a physical-digital mapping digital twin of tea production, and the simulation output is Y sim =G(X * ,z), where G is the twin simulation model, z is the high-order eigenvector, and the actual observation result Y is quantified real and virtual twin result Y sim The mapping error Δ = ‖Y real -Y sim ‖;

[0060] S64, when the error Δ exceeds the set error threshold, the process correction process is automatically triggered to generate a process correction vector δ=sign(Y real -Y sim ) and adjust the equipment parameters in real time based on the correction amount.

[0061] Optionally, step S7 specifically includes:

[0062] S71, a cross-regional, multi-agent collaborative federal learning model update mechanism is constructed, different tea garden scenes are regarded as distributed clients, each tea garden independently completes the incremental training of the model parameters under the local private data, and the model parameters of each place are aggregated by the cloud platform through the weighted average method;

[0063] S72, for model iteration, an incremental knowledge distillation strategy is adopted, the previous version model is regarded as the teacher, the current model is regarded as the student, and the historical knowledge is transferred and reserved by minimizing the distillation loss;

[0064] S73, a concept drift detection mechanism based on high-order feature vectors is set, when the variation amplitude of the feature vectors at consecutive time steps is greater than a threshold value, the retraining and adaptive updating process of the local model is automatically triggered;

[0065] S74, a reinforcement learning mechanism is introduced, and production scheduling and parameter adjustment strategies are generated according to the current high-order feature vector and the production environment state.

[0066] The beneficial effects of the present application are:

[0067] Firstly, the present application effectively improves the modeling ability of complex space-time correlation and high-order features in the tea production process by introducing a space-time self-attention neural network. Secondly, combined with the differential evolution algorithm and multi-objective optimization modeling, global optimization can be achieved among production time, cost, resource balance and quality, overcoming the shortcomings of traditional methods such as easy to fall into local optimum and static weight setting, significantly improving the scientificity and adaptability of the optimization scheme. In addition, through digital twinning and real-time multi-modal data feedback, the production process can dynamically compare the virtual-actual difference and adaptively adjust the process parameters, improving the abnormal response speed and the level of closed-loop intelligent optimization of production, and based on the federal learning mechanism, realizing the data collaboration and model continuous evolution across tea gardens and factories, ensuring data security and model generalization ability. In summary, the present application can realize intelligent modeling, global optimization and dynamic adaptive adjustment of tea production path, improve production efficiency, resource utilization and product quality, and promote the intelligent upgrading of tea industry. BRIEF DESCRIPTION OF DRAWINGS

[0068] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, used to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0069] Fig. 1 is a general flow chart of a tea production path optimization method based on deep learning proposed by the present application;

[0070] Fig. 2 is a structure framework diagram of a space-time self-attention neural network in the present application;

[0071] Fig. 3 It is a multi-objective optimization flow chart of the tea production path in the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figs. 1-3 , a tea production path optimization method based on deep learning, comprising:

[0074] S1, collect multi-source data of tea production process through distributed sensor network;

[0075] S2. Preprocess the collected multi-source data to form a standardized data set;

[0076] S3, based on the standardized dataset, uses the spatiotemporal self-attention neural network to generate high-order feature vectors;

[0077] S4. Use differential evolution algorithm to optimize the parameters of spatiotemporal self-attention neural network and minimize the total loss function through evolutionary operation;

[0078] S5. Model the tea production path based on high-order eigenvectors, construct a dynamic multi-objective production path optimization function, use a distributed non-dominated sorting genetic algorithm III to solve the Pareto optimal solution set, and screen the optimal production path optimization solution based on the attention mechanism;

[0079] S6. Refine and implement the optimal production path optimization plan, and conduct virtual and real simulation comparison through digital twins;

[0080] S7. Update the spatiotemporal self-attention neural network based on the federated learning architecture and adaptively adjust the tea production path optimization plan.

[0081] In this embodiment, step S1 specifically includes:

[0082] A distributed sensor network is set up in the tea production process to collect multi-source data. The tea production process includes tea planting, picking, transportation and processing. The multi-source data includes ambient temperature, humidity, light intensity, soil nutrients, pH value, equipment operating status, working hours, and logistics trajectory.

[0083] In this embodiment, step S2 specifically includes:

[0084] S21. Perform data cleaning on the collected multi-source data, wherein the data cleaning includes filling missing values, removing outliers, and unifying the format to obtain a cleaned data set;

[0085] S22. Standardize the cleaned dataset. The standardization process includes performing min-max normalization on the numerical data in the dataset to scale them to the interval [0, 1], and performing one-hot encoding on the categorical data in the dataset to convert each categorical category into a unique binary vector.

[0086] S23. Integrate the standardized multi-source data into a multidimensional feature data set.

[0087] In this embodiment, step S3 specifically includes:

[0088] S31, feeding the multidimensional feature dataset into a spatiotemporal self-attention neural network, wherein the spatiotemporal self-attention neural network includes an input layer, a local spatiotemporal perception embedding layer, a topological prior integration layer, a hierarchical causal gated self-attention layer, a dynamic bottleneck detection auxiliary branch, and an output layer;

[0089] S32, through the input layer, reorganize the multidimensional feature dataset according to production batch, time step and feature type, and convert it into an input tensor Where b represents the number of production batches, t represents the continuous time step, and f represents the feature vector dimension;

[0090] S33. Extracting spatial features and residual temporal features of the input tensor through a local spatiotemporal perception embedding layer. The local spatiotemporal perception embedding layer uses a parallel structure of two convolution kernels of different sizes to extract the spatial features of the input tensor and introduces a local autoregressive residual unit to extract the residual temporal features of the input tensor.

[0091] The convolution kernel sizes include 1×3 and 1×7. The 1×3 convolution kernel is responsible for capturing the detailed changes between a certain link and its adjacent links, while the 1×7 convolution kernel extracts the full-process collaboration and trend features across multiple links.

[0092] The local autoregressive residual unit is used to calculate the feature x of each production batch at each time step. t Make a prediction, and the predicted value is where a k is the learnable autoregressive coefficient, K is the backtracking step size, and the residual between the actual observation value and the predicted value is calculated As residual time series features;

[0093] S34, input the spatial features and residual time series features into the topological prior integration layer, and introduce the topological prior matrix T∈{0,1} f×f Constrain the attention weights between features;

[0094] The topological prior matrix T is automatically generated through the production process knowledge graph. If there is information flow interaction in the production links corresponding to features i and j, T i,j =1;

[0095] The attention weight α ij Adjust according to the following formula:

[0096]

[0097] Among them, q i is the embedding vector of the i-th feature, k j is the embedding vector of the jth feature, T i,j =1 means that features i and j are related in the process flow, T i,j =0 means no association;

[0098] S35. Input the feature representation output by the topological prior integration layer into the hierarchical causal gated self-attention layer. The hierarchical causal gated self-attention layer has three sub-layers: short-term dependency, medium-term dependency, and long-term dependency. Each sub-layer performs self-attention calculation on the input features through causal masks of different time windows. The self-attention outputs of each sub-layer are dynamically weighted by the gating function to generate gated attention features of the corresponding time scale.

[0099] S36. Classify and calculate the gated attention features through the dynamic bottleneck detection auxiliary branch to obtain a bottleneck score, and fuse the bottleneck score with the gated attention features using a residual connection method to generate an enhanced feature;

[0100] S37. The enhanced features are input into the output layer, and the features related to the optimization target are extracted through the multi-task decoding module. The features are then fused and compressed through the adaptive distillation module to generate a high-order feature vector.

[0101] In this embodiment, step S4 specifically includes:

[0102] The differential evolution algorithm is used to train and optimize the parameters of the spatiotemporal self-attention neural network. By performing mutation, crossover, and selection iterations on the parameter population, the total loss function is minimized. The total loss function includes production link state loss, bottleneck classification loss, feature decoupling loss, and temporal consistency loss.

[0103] The production link status loss The cross entropy between the model's output of multiple categories of states in the production process and the corresponding true category labels;

[0104] The bottleneck classification loss The cross entropy between the output probability distribution of the model and the true bottleneck state label when performing bottleneck detection tasks;

[0105] The characteristic decoupling loss To minimize the mutual information between the dimensions of the high-order eigenvectors, specifically Where MI(z i ,z j ) represents the mutual information between the i-th dimension and the j-th dimension high-order features;

[0106] The timing consistency loss It means constraining the changes between high-order eigenvectors of adjacent time steps to remain smooth, specifically Where z represents the high-order eigenvector, t represents the time step, and ‖·‖ represents the Euclidean norm;

[0107] The total loss function formula is:

[0108]

[0109] Among them, β,γ are adaptive weight coefficients;

[0110] In this embodiment, step S5 specifically includes:

[0111] S51. Modeling the entire tea production process based on high-order eigenvectors, defining the decision variables in the tea production path optimization plan as X, wherein the decision includes process sequence, resource allocation, equipment scheduling, and logistics path;

[0112] S52. Construct a dynamic multi-objective production path optimization function. The dynamic multi-objective production path optimization function comprehensively considers the overall production time, production cost, resource utilization balance, and tea quality score, and optimizes each indicator as an independent target simultaneously. Specifically,

[0113]

[0114] Where f1(X) represents the overall production time, and f2(X) represents the production cost function integrating the tea quality regulation mechanism, specifically: α is the adjustment parameter, is the minimum quality reference value, and f3(X) represents the resource utilization balance index function, specifically: σ represents the standard deviation of the waiting time, f4(X) represents the tea quality prediction function based on the high-order eigenvector, and T represents the transposed sign;

[0115] S53. Setting enhanced constraint conditions, wherein the enhanced constraint conditions include quality constraints and elastic resource constraints;

[0116] The quality constraint is: the quality prediction value of the tea production plan must be higher than the sum of the set threshold and the quality compensation amount calculated based on the high-order eigenvector;

[0117] The elastic resource constraint is: under the premise of ensuring production continuity, the dynamic elastic adjustment amount does not exceed the upper limit of the corresponding resource;

[0118] S54, using the distributed non-dominated sorting genetic algorithm III to perform global optimization on the dynamic multi-objective production path optimization function under enhanced constraints and obtain the Pareto optimal solution set;

[0119] S55. Screening the optimal production path optimization solution from the Pareto optimal solution set based on the attention mechanism X * ,

[0120]

[0121] Among them, the Pareto set represents the Pareto optimal solution set obtained, f′ i (X) represents the evaluation function value of the decision variable X on the i-th optimization objective, λ i Represents the adaptive weight coefficient corresponding to the i-th target, which is calculated by performing softmax normalization on the high-order feature vector.

[0122] In this embodiment, step S6 specifically includes:

[0123] S61, optimize the optimal production path plan X * It is broken down into process-level execution instruction sets, and blockchain evidence storage technology is used to perform distributed encrypted recording of each process execution status and operation process data;

[0124] S62. During the data collection process, new features are added: equipment vibration spectrum, tea phenotype images, and operator biometrics acquired in real time;

[0125] S63, based on real-time collected data and historical optimization schemes, build a physical-digital mapping digital twin of tea production, and the simulation output is Y sim =G(X * ,z), where G is the twin simulation model, z is the high-order eigenvector, and the actual observation result Y is quantified real and virtual twin result Y sim The mapping error Δ = ‖Y real -Y sim ‖;

[0126] S64, when the error Δ exceeds the set error threshold, the process correction process is automatically triggered to generate a process correction vector δ=sign(Y real -Y sim ) and adjust the equipment parameters in real time based on the correction amount.

[0127] In this embodiment, step S7 specifically includes:

[0128] S71. Build a cross-regional, multi-agent collaborative federated learning model update mechanism. Treat different tea garden scenarios as distributed clients. Each tea garden independently completes incremental training of model parameters using local private data. The model parameters from various locations are aggregated using a weighted average method through the cloud platform.

[0129] S72. For model iteration, an incremental knowledge distillation strategy is adopted, using the previous version of the model as the teacher and the current model as the student, migrating and retaining historical knowledge by minimizing the distillation loss;

[0130] S73. Setting a concept drift detection mechanism based on high-order feature vectors. When the amplitude of the feature vector change in consecutive time steps is greater than a threshold, the retraining and adaptive update process of the local model is automatically triggered.

[0131] S74. Introduce a reinforcement learning mechanism to generate production scheduling and parameter adjustment strategies based on the current high-order feature vector and production environment status.

[0132] Example 1:

[0133] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a large-scale intelligent white tea production base. The base has 800 acres of tea gardens and is equipped with a modern production line with an annual output of 350 tons, covering the entire process of planting, picking, transportation, processing, etc. Traditional management relies on manual experience and static scheduling, resulting in poor process connection, long production cycle, uneven resource utilization, high cost and large fluctuations in product quality. In order to solve these problems, the base adopted the tea production path optimization method of the present invention to optimize production management.

[0134] During application, the solution of the present invention first collects multidimensional data such as temperature, humidity, light, soil pH, equipment operating status, personnel trajectory, equipment vibration spectrum, and tea phenotypic images in real time through a distributed sensor network. After preprocessing, the data is input into a spatiotemporal self-attention neural network for high-order feature extraction and key bottleneck identification. The network parameters are globally optimized using a differential evolution algorithm, and the network's generalization capability is improved by combining cross-entropy loss, bottleneck classification loss, feature decoupling loss, and temporal consistency loss. After obtaining high-order feature vectors, a dynamic multi-objective production path optimization method is used, with four core indicators of production time, production cost, resource balance, and tea quality as simultaneous optimization targets. A distributed non-dominated sorting genetic algorithm III is used to obtain the Pareto optimal solution of the production path under optimization constraints, and the optimal production path optimization solution is dynamically selected based on the real-time feature vectors. Subsequently, a virtual-to-real comparison is performed between the actual production and simulation results through a digital twin, and equipment parameters and process flows are adjusted in a timely manner based on the errors between the two, achieving closed-loop adaptive optimization of the production process. At the same time, two affiliated tea plantations at the base participated in collaborative model training using a federated learning mechanism, improving generalization capabilities while protecting data privacy. During the three-month pilot, 536 production batches were scheduled. A comparison of the specific implementation results is shown in Table 1.

[0135] Table 1 Comparison of the results before and after the implementation of the tea production path optimization plan

[0136]

[0137] Compared with the historical data of the same period, after using the method of the present application, the average production cycle of a single batch is shortened from 4.25 days to 2.93 days, the average production cost of each batch is reduced from 8342 yuan to 7363 yuan, with a decrease of 11.7%, effectively reducing the enterprise operation expenditure and enhancing the market competitiveness. The average score of tea quality is improved from 8.46 to 8.74, and the high-quality good product rate is improved from 78% to 89%, which shows that the enterprise can produce more products with higher added value. The resource utilization range is reduced from 2.40 to 1.38, and the standard deviation of the waiting time in the production link is reduced from 17.4 minutes to 8.2 minutes, which shows that the resource allocation in the production process is more balanced, and the connection of each process is more smooth, effectively reducing the idle and congestion of local resources. The frequency of major process bottlenecks is reduced from 8 times per month to 2 times, and the response time is shortened from 33 minutes to 8 minutes, which significantly improves the flexibility and emergency handling capacity of the production process. The federal model update cycle is shortened from 23 days to 14 days, the new scene production adaptation period is shortened from 17 days to 8 days, and the prediction accuracy of the scheduling model is improved from 86.2% to 94.4%, which fully shows that the present application not only improves the generalization ability and adaptation speed of the model to new scenes, but also greatly enhances the guiding effect and application value of path optimization to actual production. In summary, the present application can effectively improve the intelligent level of tea production management, and has achieved good results in improving production efficiency, reducing production cost, ensuring product quality and optimizing resource scheduling.

[0138] The above is only the preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A tea production path optimization method based on deep learning, characterized in that: The steps include: S1, collect multi-source data of tea production process through distributed sensor network; S2. Preprocess the collected multi-source data to form a standardized data set; S3, based on the standardized dataset, uses the spatiotemporal self-attention neural network to generate high-order feature vectors; S4. Use differential evolution algorithm to optimize the parameters of spatiotemporal self-attention neural network and minimize the total loss function through evolutionary operation; S5. Model the tea production path based on high-order eigenvectors, construct a dynamic multi-objective production path optimization function, use a distributed non-dominated sorting genetic algorithm III to solve the Pareto optimal solution set, and screen the optimal production path optimization solution based on the attention mechanism; S6. Refine and implement the optimal production path optimization plan, and conduct virtual and real simulation comparison through digital twins; S7. Update the spatiotemporal self-attention neural network based on the federated learning architecture and adaptively adjust the tea production path optimization plan.

2. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S1 specifically includes: setting up a distributed sensor network in the tea production process to collect multi-source data, the tea production process includes tea planting, picking, transportation and processing, and the multi-source data includes ambient temperature, humidity, light intensity, soil nutrients, pH value, equipment operating status, working hours, and logistics trajectory.

3. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S2 specifically includes: S21. Perform data cleaning on the collected multi-source data, wherein the data cleaning includes filling missing values, removing outliers, and unifying the format to obtain a cleaned data set; S22. Standardize the cleaned dataset. The standardization process includes performing min-max normalization on the numerical data in the dataset to scale them to the interval [0, 1], and performing one-hot encoding on the categorical data in the dataset to convert each categorical category into a unique binary vector. S23. Integrate the standardized multi-source data into a multidimensional feature data set.

4. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S3 specifically includes: S31, feeding the multidimensional feature dataset into a spatiotemporal self-attention neural network, wherein the spatiotemporal self-attention neural network includes an input layer, a local spatiotemporal perception embedding layer, a topological prior integration layer, a hierarchical causal gated self-attention layer, a dynamic bottleneck detection auxiliary branch, and an output layer; S32, through the input layer, reorganize the multidimensional feature dataset according to production batch, time step and feature type, and convert it into an input tensor Where b represents the number of production batches, t represents the continuous time step, and f represents the feature vector dimension; S33. Extracting spatial features and residual temporal features of the input tensor through a local spatiotemporal perception embedding layer. The local spatiotemporal perception embedding layer uses a parallel structure of two convolution kernels of different sizes to extract the spatial features of the input tensor and introduces a local autoregressive residual unit to extract the residual temporal features of the input tensor. The convolution kernel sizes include 1×3 convolution kernel and 1×7 convolution kernel; The local autoregressive residual unit is used to calculate the feature x of each production batch at each time step. t Make a prediction, and the predicted value is where a k is the learnable autoregressive coefficient, K is the backtracking step size, and the residual between the actual observation value and the predicted value is calculated As residual time series features; S34, input the spatial features and residual time series features into the topological prior integration layer, and introduce the topological prior matrix T∈{0,1} f×f Constrain the attention weights between features; The topological prior matrix T is automatically generated through the production process knowledge graph. If there is information flow interaction in the production links corresponding to features i and j, T i,j =1; The attention weight α ij Adjust according to the following formula: Among them, q i is the embedding vector of the i-th feature, k j is the embedding vector of the jth feature, T i,j =1 means that features i and j are related in the process flow, T i,j =0 means no association; S35. Input the feature representation output by the topological prior integration layer into the hierarchical causal gated self-attention layer. The hierarchical causal gated self-attention layer has three sub-layers: short-term dependency, medium-term dependency, and long-term dependency. Each sub-layer performs self-attention calculation on the input features through causal masks of different time windows. The self-attention outputs of each sub-layer are dynamically weighted by the gating function to generate gated attention features of the corresponding time scale. S36. Classify and calculate the gated attention features through the dynamic bottleneck detection auxiliary branch to obtain a bottleneck score, and fuse the bottleneck score with the gated attention features using a residual connection method to generate an enhanced feature; S37. The enhanced features are input into the output layer, and the features related to the optimization target are extracted through the multi-task decoding module. The features are then fused and compressed through the adaptive distillation module to generate a high-order feature vector.

5. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S4 specifically includes: The differential evolution algorithm is used to train and optimize the parameters of the spatiotemporal self-attention neural network. By performing mutation, crossover, and selection iterations on the parameter population, the total loss function is minimized. The total loss function includes production link state loss, bottleneck classification loss, feature decoupling loss, and temporal consistency loss. The production link status loss The cross entropy between the model's output of multiple categories of states in the production process and the corresponding true category labels; The bottleneck classification loss The cross entropy between the output probability distribution of the model and the true bottleneck state label when performing bottleneck detection tasks; The characteristic decoupling loss To minimize the mutual information between the dimensions of the high-order eigenvectors, specifically Among them, MI(z i ,z j ) represents the mutual information between the i-th dimension and the j-th dimension high-order features; The timing consistency loss It means constraining the changes between high-order eigenvectors of adjacent time steps to remain smooth, specifically Where z represents the high-order eigenvector, t represents the time step, and ‖·‖ represents the Euclidean norm; The total loss function formula is: Among them, β and γ are adaptive weight coefficients.

6. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S5 specifically includes: S51. Model the entire tea production process based on the high-order eigenvectors, and define the decision variables in the tea production path optimization plan as X. The decision includes process sequence, resource allocation, equipment scheduling, and logistics path; S52. Construct a dynamic multi-objective production path optimization function. The dynamic multi-objective production path optimization function comprehensively considers the overall production time, production cost, resource utilization balance, and tea quality score, and optimizes each indicator as an independent target simultaneously. Specifically, Where f1(X) represents the overall production time, and f2(X) represents the production cost function integrating the tea quality regulation mechanism, specifically: α is the adjustment parameter, is the minimum quality reference value, and f3(X) represents the resource utilization balance index function, specifically: σ represents the standard deviation of the waiting time, f4(X) represents the tea quality prediction function based on the high-order eigenvector, and T represents the transposed sign; S53. Setting enhanced constraint conditions, wherein the enhanced constraint conditions include quality constraints and elastic resource constraints; The quality constraint is: the quality prediction value of the tea production plan must be higher than the sum of the set threshold and the quality compensation amount calculated based on the high-order eigenvector; The elastic resource constraint is: under the premise of ensuring production continuity, the dynamic elastic adjustment amount does not exceed the upper limit of the corresponding resource; S54, using the distributed non-dominated sorting genetic algorithm III to perform global optimization on the dynamic multi-objective production path optimization function under enhanced constraints and obtain the Pareto optimal solution set; S55. Screening the optimal production path optimization solution from the Pareto optimal solution set based on the attention mechanism X * , Among them, the Pareto set represents the Pareto optimal solution set obtained, f′ i (X) represents the evaluation function value of the decision variable X on the i-th optimization objective, λ i Represents the adaptive weight coefficient corresponding to the i-th target, which is calculated by performing softmax normalization on the high-order feature vector.

7. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S6 specifically includes: S61, optimize the optimal production path plan X * It is broken down into process-level execution instruction sets, and blockchain evidence storage technology is used to perform distributed encrypted recording of each process execution status and operation process data; S62. During the data collection process, new features are added: equipment vibration spectrum, tea phenotype images, and operator biometrics acquired in real time; S63, based on real-time collected data and historical optimization schemes, build a physical-digital mapping digital twin of tea production, and the simulation output is Y sim =G(X * ,z), where G is the twin simulation model, z is the high-order eigenvector, and the actual observation result Y is quantified real and virtual twin result Y sim The mapping error Δ = ‖Y real -Y sim ‖; S64, when the error Δ exceeds the set error threshold, the process correction process is automatically triggered to generate a process correction vector δ=sign(Y real -Y sim ) and adjust the equipment parameters in real time based on the correction amount.

8. The tea production path optimization method based on deep learning according to claim 1, characterized in that: The step S7 specifically includes: S71. Build a cross-regional, multi-agent collaborative federated learning model update mechanism. Treat different tea garden scenarios as distributed clients. Each tea garden independently completes incremental training of model parameters using local private data. The model parameters from various locations are aggregated using a weighted average method through the cloud platform. S72. For model iteration, an incremental knowledge distillation strategy is adopted, using the previous version of the model as the teacher and the current model as the student, migrating and retaining historical knowledge by minimizing the distillation loss; S73. Setting a concept drift detection mechanism based on high-order feature vectors. When the amplitude of the feature vector change in consecutive time steps is greater than a threshold, the retraining and adaptive update process of the local model is automatically triggered. S74. Introduce a reinforcement learning mechanism to generate production scheduling and parameter adjustment strategies based on the current high-order feature vector and production environment status.

Images