A tea processing auxiliary method and system based on intelligent decision-making

Abstract

The present invention discloses a tea processing assistance method and system based on intelligent decision-making, comprising: performing tea processing monitoring to obtain tea processing monitoring data, pre-processing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information; analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing multi-modal fusion features to obtain multi-modal tea processing fusion features; constructing a tea processing quality perception model, inputting the multi-modal tea processing fusion features into the trained tea processing quality perception model to perceive the processing quality of tea in a current processing batch to obtain tea processing quality perception information; judging whether the quality of the tea in the current processing batch meets expectations based on the tea processing quality perception information, and if not, optimizing the current processing plan to provide tea processing assistance. The method enhances the perception capability of the tea processing process and improves the final tea processing quality.

Description

Technical Field

[0001] The present invention relates to the technical field of tea processing, and in particular to a tea processing auxiliary method and system based on intelligent decision-making. Background Art

[0002] The traditional tea processing process relies on manual experience to control key parameters such as withering temperature, rolling pressure, and fermentation humidity. It has inherent defects such as low control accuracy, poor quality stability, and high energy consumption. Existing automated equipment mostly uses preset process curves or single sensor feedback control, which is difficult to adapt to the differences in the physical and chemical properties of fresh leaf raw materials and dynamic environmental disturbances, resulting in insufficient conversion of tea ingredients and substandard morphological characteristics during processing. Although some studies in recent years have attempted to introduce machine vision to detect changes in leaf color or analyze moisture content through near-infrared spectroscopy, the sensor data are processed in isolation, lacking multimodal feature fusion and dynamic correlation modeling between process parameters, and it is impossible to build a full-link decision-making system covering "raw material characteristics-processing status-finished product quality".

[0003] Furthermore, existing technologies lack the ability to deeply integrate and intelligently analyze multi-source data (such as spectra, images, and environmental sensors). Most systems rely solely on feedback from a single sensor for local adjustments, making it difficult to globally optimize the collaborative efficiency of the entire processing chain. For example, during the oolong tea fermentation process, while traditional equipment can maintain constant temperature and humidity, it cannot dynamically adjust the shaking frequency based on the degree of red edge formation and aroma release characteristics of the leaves, which can easily lead to insufficient or excessive fermentation. These limitations continue to plague tea processing, resulting in bottlenecks such as large quality fluctuations, high energy consumption, and low standardization. Intelligent decision-making technologies are urgently needed to achieve adaptive optimization of process parameters and precise control of the entire process. Summary of the Invention

[0004] The present invention overcomes the defects of the prior art and provides a tea processing auxiliary method and system based on intelligent decision-making, the important purpose of which is to enhance the perception ability of the tea processing process and improve the final tea processing quality.

[0005] To achieve the above objectives, the present invention provides a tea processing auxiliary method based on intelligent decision-making in a first aspect, comprising:

[0006] Performing tea processing monitoring on target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information;

[0007] Analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing a multimodal fusion feature to obtain a multimodal tea processing fusion feature;

[0008] Constructing a tea processing quality perception model, inputting the multimodal tea processing fusion features into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtaining tea processing quality perception information;

[0009] Based on the tea processing quality perception information, it is determined whether the quality of the tea in the current processing batch meets expectations. If not, the current processing plan is optimized to assist in tea processing.

[0010] In this solution, the tea processing monitoring of the target processing equipment is performed to obtain tea processing monitoring data, and the obtained tea processing monitoring data is preprocessed to obtain cross-modal tea processing monitoring information, specifically including:

[0011] When the target processing equipment is processing tea, the installed sensor array is used to monitor the target processing equipment for tea processing to obtain tea processing monitoring data, wherein the tea processing monitoring data includes processing environment monitoring data, processing equipment operation data, and processed tea monitoring data;

[0012] Performing wavelet packet decomposition on the tea processing monitoring data using a preset wavelet basis function to obtain different sub-bands and calculating the energy entropy of each sub-band, and performing soft threshold filtering based on the calculated energy entropy to identify effective characteristic frequency bands;

[0013] After filtering, the tea processing monitoring data was time-series aligned, noise was removed using median filtering, and motion blur was eliminated through Gaussian smoothing. A sliding window algorithm was introduced to slide based on a preset window and calculate the mean and standard deviation of the data within the window. A linear interpolation algorithm was then used for data cleaning.

[0014] Maximum-minimum normalization is performed on the tea processing monitoring data after cleaning, and the continuous data is mapped to the interval [0,1]. The discrete data is standardized through binning to generate cross-modal tea processing monitoring information.

[0015] In this solution, the cross-modal correlation relationship of each monitoring data is analyzed based on the cross-modal tea processing monitoring information, and a multi-modal fusion feature is constructed to obtain a multi-modal tea processing fusion feature, which specifically includes:

[0016] Obtaining cross-modal tea processing monitoring information, obtaining processed tea monitoring data through the cross-modal tea processing monitoring information and extracting processed tea image monitoring data, and importing the processed tea image monitoring data into the HSV color space to obtain a histogram distribution;

[0017] The histogram distribution is used to calculate the local area hue mean and saturation variance, and then a weighted fusion is performed to form the browning degree index. Subsequently, a morphological algorithm is used to extract the contour of the maximum connected area of ​​the processed tea image. The ratio between the contour perimeter and the tea area is calculated as the curling index, and the first feature information is generated by combining it with the texture characteristics of the processed tea.

[0018] Extracting near-infrared spectral data from processed tea monitoring data and calculating the first-order derivative to generate a near-infrared derivative spectrum, inputting the near-infrared derivative spectrum into a pre-trained least squares regression model to invert the chemical composition of the tea leaves and generate a spectral domain feature vector of the processed tea leaves, thereby obtaining second feature information;

[0019] A multi-head attention mechanism is introduced to use the first feature information as the query vector, the second feature information as the key vector and the value vector, and the scaled click attention score is calculated to perform feature fusion and generate fused feature information;

[0020] Cross-modal tea processing monitoring information is used to extract processing environment monitoring data and processing equipment operation data, which are combined with the fusion feature information and input into the BiGRU network for multimodal feature fusion to obtain multimodal tea processing fusion features.

[0021] In this solution, the cross-modal tea processing monitoring information is used to extract processing environment monitoring data and processing equipment operation data, which are combined with the fusion feature information and input into the BiGRU network for multimodal feature fusion, specifically including:

[0022] In the time dimension, the data input into the BiGRU network is converted into several time series feature vectors, where each time series feature vector contains the tea texture details, spectral chemical composition, equipment operating parameters, and processing environment parameters at the current moment;

[0023] The forward propagation layer captures the immediate changes of the blade state caused by the adjustment of the processing parameters through the update gate, and the backward propagation layer uses the reset gate to filter the lagged changes of the blade state caused by the adjustment of the processing parameters to generate temporal causal features;

[0024] In the spatial dimension, the data input to the BiGRU network is spatially convolved using a convolutional layer to obtain a convolutional feature map, and the convolutional feature map is globally averaged pooled in the channel dimension to generate a spatial significance weight vector;

[0025] Obtain the temporal hidden state of the BiGRU, perform matrix multiplication with the spatial significance weight vector, calculate the importance score of each channel in the temporal context, perform channel weighting on the original convolution feature map, and obtain the spatial response enhancement feature;

[0026] The temporal causal features and spatial response enhancement features are input into the fully connected layer for feature vector fusion to generate multimodal tea processing fusion features when the target processing equipment is processing tea.

[0027] In this solution, the tea processing quality perception model is constructed, and the multimodal tea processing fusion features are input into the trained tea processing quality perception model to perform processing quality perception of the current processing batch of tea, specifically including:

[0028] Acquire historical processing instances of different tea processing qualities based on historical data retrieval, extract historical tea processing features when tea is processed in each historical processing instance according to the acquired historical processing instances, and obtain historical tea processing feature information;

[0029] Extracting the tea processing flow corresponding to each historical processing instance, performing feature partitioning on the historical tea processing feature information, and generating tea processing feature subsets of several different historical processing instances, each tea processing feature subset corresponding to a different tea processing flow;

[0030] The historical tea processing feature information after feature partitioning is associated with the corresponding historical tea processing quality, and the tea processing feature is used as the first node, the tea processing quality is used as the second node, and the tea processing process is defined as the node label of the first node to establish a topological structure diagram;

[0031] A tea processing quality perception model was constructed based on a long short-term memory network and a graph neural network. The historical tea processing feature information after feature partitioning was input into the long short-term memory network to extract multi-scale temporal features of different processing qualities.

[0032] The adjacency matrix is ​​obtained through the established topological structure graph, and the nodes are aggregated and updated based on the message passing mechanism to obtain the updated node representation. The updated node representation of the entire graph is globally pooled to obtain spatial features, and the multi-scale temporal features are fused with the spatial features to generate spatiotemporal fusion features.

[0033] The spatiotemporal fusion features are input into the fully connected layer to perform tea processing quality perception. The tea processing quality perception results are verified and parameters are adjusted using a preset validation set. After iterative training, a tea processing quality perception model that meets the expectations is obtained.

[0034] Acquire multimodal tea processing fusion features, input the multimodal tea processing fusion features into a trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtain tea processing quality perception information.

[0035] In this solution, the method of judging whether the quality of the tea leaves of the current processing batch meets the expectation based on the tea processing quality perception information, and if not, optimizing the current processing plan to assist in tea processing, specifically includes:

[0036] Acquiring tea processing quality perception information, and generating a tea quality change perception curve of a current processing batch based on the tea processing quality perception information, wherein the tea quality change perception curve represents the tea processing quality of the current processing batch of tea at different time stamps;

[0037] Obtaining the tea quality requirements for the current processing batch, calculating the deviation between the tea processing quality and the tea quality requirements corresponding to different time stamps based on the tea quality change perception curve, and comparing the calculated deviation with a preset tolerance range;

[0038] If it is greater than the preset tolerance range, it means that the current tea processing control parameters are not suitable for the tea quality requirements of the current processing batch, then the preset processing plan of the current processing batch is obtained to optimize the processing plan;

[0039] Retrieving historical processing cases that meet the current tea quality requirements and processed tea varieties from a preset historical processing database, and extracting historical processing control plans corresponding to each historical processing case;

[0040] A multi-objective gray wolf optimization algorithm is introduced. The population is initialized according to the historical processing control scheme corresponding to each historical processing case. The objective function is preset and the constraints are set. The objective function value of each individual in the initial gray wolf population is calculated through the objective function, and all individuals in the population are divided into different frontier levels through non-dominated sorting.

[0041] Congestion calculation is performed on each frontier layer to obtain the congestion distance. The individual with the highest congestion in each layer is selected as the leader wolf, and the remaining individuals are used as follower wolves. The direction vector of the leader wolf in the decision space is calculated to update its position. The optimal solution set is output after repeated iterations until the stopping condition is met.

[0042] Based on the output optimal solution set, several candidate processing optimization schemes are generated, and the feasibility of each candidate processing optimization scheme is screened. Based on the feasibility screening results, the optimal processing optimization scheme is selected to assist tea processing.

[0043] A second aspect of the present invention provides a tea processing assistance system based on intelligent decision-making, the system comprising: a memory and a processor, wherein the memory contains a tea processing assistance method program based on intelligent decision-making, and when the tea processing assistance method program based on intelligent decision-making is executed by the processor, the following steps are implemented:

[0044] Performing tea processing monitoring on target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information;

[0045] Analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing a multimodal fusion feature to obtain a multimodal tea processing fusion feature;

[0046] Constructing a tea processing quality perception model, inputting the multimodal tea processing fusion features into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtaining tea processing quality perception information;

[0047] Based on the tea processing quality perception information, it is determined whether the quality of the tea in the current processing batch meets expectations. If not, the current processing plan is optimized to assist in tea processing.

[0048] The present invention discloses a tea processing assistance method and system based on intelligent decision-making, comprising: performing tea processing monitoring to obtain tea processing monitoring data, pre-processing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information; analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing multi-modal fusion features to obtain multi-modal tea processing fusion features; constructing a tea processing quality perception model, inputting the multi-modal tea processing fusion features into the trained tea processing quality perception model to perceive the processing quality of tea in a current processing batch to obtain tea processing quality perception information; judging whether the quality of the tea in the current processing batch meets expectations based on the tea processing quality perception information, and if not, optimizing the current processing plan to provide tea processing assistance. The method enhances the perception capability of the tea processing process and improves the final tea processing quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0050] Figure 1 A flow chart of a tea processing auxiliary method based on intelligent decision-making provided by one embodiment of the present invention;

[0051] Figure 2 A flow chart of an intelligent decision-making and auxiliary method for tea processing provided by one embodiment of the present invention;

[0052] Figure 3A block diagram of a tea processing auxiliary system based on intelligent decision-making provided by one embodiment of the present invention;

[0053] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0054] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0056] Figure 1 A flow chart of a tea processing auxiliary method based on intelligent decision-making provided by one embodiment of the present invention;

[0057] like Figure 1 As shown, the present invention provides a flow chart of a tea processing auxiliary method based on intelligent decision-making, including:

[0058] S102, performing tea processing monitoring on the target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information;

[0059] S104, analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing a multi-modal fusion feature to obtain a multi-modal tea processing fusion feature;

[0060] S106, constructing a tea processing quality perception model, inputting the multimodal tea processing fusion features into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtaining tea processing quality perception information;

[0061] S108: Based on the tea processing quality perception information, determine whether the quality of the tea leaves in the current processing batch meets expectations. If not, optimize the current processing plan to assist in tea processing.

[0062] Furthermore, in a preferred embodiment of the present invention, the tea processing monitoring of the target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information specifically includes:

[0063] When the target processing equipment is processing tea, the installed sensor array is used to monitor the target processing equipment for tea processing to obtain tea processing monitoring data, wherein the tea processing monitoring data includes processing environment monitoring data, processing equipment operation data, and processed tea monitoring data;

[0064] Performing wavelet packet decomposition on the tea processing monitoring data using a preset wavelet basis function to obtain different sub-bands and calculating the energy entropy of each sub-band, and performing soft threshold filtering based on the calculated energy entropy to identify effective characteristic frequency bands;

[0065] After filtering, the tea processing monitoring data was time-series aligned, noise was removed using median filtering, and motion blur was eliminated through Gaussian smoothing. A sliding window algorithm was introduced to slide based on a preset window and calculate the mean and standard deviation of the data within the window. A linear interpolation algorithm was then used for data cleaning.

[0066] Maximum-minimum normalization is performed on the tea processing monitoring data after cleaning, and the continuous data is mapped to the interval [0,1]. The discrete data is standardized through binning to generate cross-modal tea processing monitoring information.

[0067] It should be noted that a multi-source sensor network is deployed to collect real-time monitoring data from the entire tea processing process. This network integrates a high-precision infrared thermal imager, a multispectral imaging unit, a triaxial vibration sensor, and a temperature and humidity sensing module. It simultaneously captures processing environment parameters (such as the three-dimensional temperature distribution inside the drum and the ambient humidity gradient), equipment operating status (such as motor current fluctuations and mechanical drive shaft speed waveforms), and tea characteristics (including leaf surface wrinkling and near-infrared spectra). The collected raw signal is then subjected to adaptive wavelet packet decomposition, using specific wavelet basis functions to perform multi-scale frequency band segmentation on the non-stationary time series data. By calculating the energy distribution entropy of each subband, characteristic frequency bands containing key process information are identified. Noise-dominated subbands with entropy values ​​below a threshold are suppressed using dynamic soft threshold filtering, eliminating random interference while preserving effective signal edge features. The multi-source data stream that has completed frequency domain filtering enters the spatiotemporal alignment module, which eliminates the phase deviation caused by device heterogeneous sampling through millisecond-level timestamp matching, and adopts a composite filtering strategy to deal with sensor-specific noise: a median filter eliminates the pulse interference of the temperature sensor, and an anisotropic Gaussian kernel compensates for the motion blur effect of the optical imaging unit in high-speed rotation scenes, restoring the true morphological characteristics of the leaf texture.

[0068] Next, a dynamic sliding window mechanism is introduced during the data cleaning phase. This mechanism scans the data stream with a window length adaptive to process timing characteristics, calculating in-window statistics (mean and standard deviation) in real time to construct a data quality assessment matrix. A repair mechanism is triggered for detected anomalous intervals (such as outliers or data discontinuities caused by transient sensor failures). Short-term omissions are compensated using linear interpolation, while persistent anomalies are replaced by redundant sensor data. After cleaning, the heterogeneous data undergoes cross-modal normalization. Continuous parameters (such as temperature, pressure, and moisture content) are mapped to a uniform numerical range using dynamic range normalization to eliminate dimensional differences. Discrete state parameters (such as equipment fault codes and process stage identifiers) are converted into standard feature vectors using equal probability binning. Ultimately, cross-modal tea processing monitoring information is generated, providing a highly consistent and relevant input foundation for subsequent intelligent decision-making models.

[0069] Furthermore, in a preferred embodiment of the present invention, the cross-modal correlation relationship of each monitoring data is analyzed based on the cross-modal tea processing monitoring information, and a multimodal fusion feature is constructed to obtain a multimodal tea processing fusion feature, which specifically includes:

[0070] Obtaining cross-modal tea processing monitoring information, obtaining processed tea monitoring data through the cross-modal tea processing monitoring information and extracting processed tea image monitoring data, and importing the processed tea image monitoring data into the HSV color space to obtain a histogram distribution;

[0071] The histogram distribution is used to calculate the local area hue mean and saturation variance, and then a weighted fusion is performed to form the browning degree index. Subsequently, a morphological algorithm is used to extract the contour of the maximum connected area of ​​the processed tea image. The ratio between the contour perimeter and the tea area is calculated as the curling index, and the first feature information is generated by combining it with the texture characteristics of the processed tea.

[0072] Extracting near-infrared spectral data from processed tea monitoring data and calculating the first-order derivative to generate a near-infrared derivative spectrum, inputting the near-infrared derivative spectrum into a pre-trained least squares regression model to invert the chemical composition of the tea leaves and generate a spectral domain feature vector of the processed tea leaves, thereby obtaining second feature information;

[0073] A multi-head attention mechanism is introduced to use the first feature information as the query vector, the second feature information as the key vector and the value vector, and the scaled click attention score is calculated to perform feature fusion and generate fused feature information;

[0074] Cross-modal tea processing monitoring information is used to extract processing environment monitoring data and processing equipment operation data, which are combined with the fusion feature information and input into the BiGRU network for multimodal feature fusion to obtain multimodal tea processing fusion features.

[0075] It should be noted that tea image data is first separated from cross-modal monitoring information and converted to the HSV color space to analyze the distribution characteristics of hue (H) and saturation (S). Histograms are constructed by statistically analyzing the pixel density of different hue intervals. Based on the histogram distribution, a weighted fusion value of the hue mean and saturation variance of local leaf regions is calculated (the weight coefficient is dynamically adjusted according to the process stage) to form a quantitative index representing the degree of enzymatic browning. Simultaneously, a morphological dilation-erosion algorithm is used to extract the contour of the largest connected region in the image. The degree of leaf curl is quantified by the ratio of the contour perimeter to the projected area. Combined with texture contrast and energy values ​​calculated using the gray-level co-occurrence matrix (GLCM), primary feature information describing the changes in the tea's appearance and morphology is generated. Furthermore, the extracted near-infrared spectral data is subjected to Savitzky-Golay smoothing filtering, and its first-order derivative spectrum is calculated to highlight characteristic absorption peaks. This derivative spectrum is then input into a pre-trained least-squares regression model (trained on historical tea sample data) to infer the concentration distribution of key chemical components such as tea polyphenols and water, forming secondary feature information reflecting the internal substance transformation of the tea leaves. A multi-head attention mechanism is used to establish an association model between appearance features and chemical composition: the first feature information is used as the query vector (Query), and the second feature information as the key vector (Key) and value vector (Value). The scaled dot product similarity score of each attention head is calculated, and high-correlation feature nodes are screened (for example, the negative correlation between high curl and low tea polyphenol concentration). After weighted fusion, a fused feature vector is generated that jointly represents the evolution of tea's internal and external quality. Processing environment parameters (temperature and humidity gradient, airflow velocity) and equipment operating status (motor load factor, drum speed timing) are further integrated and input into a bidirectional gated recurrent network (BiGRU). The update gate and reset gate dynamically adjust the fusion weights of cross-modal information, ultimately outputting a multimodal fusion feature that simultaneously encodes appearance, chemical composition, environmental disturbances, and equipment operating conditions, providing a highly discriminative state representation for intelligent decision-making.

[0076] Furthermore, in a preferred embodiment of the present invention, the cross-modal tea processing monitoring information is used to extract processing environment monitoring data and processing equipment operation data, and the fusion feature information is combined and input into the BiGRU network for multimodal feature fusion, specifically including:

[0077] In the time dimension, the data input into the BiGRU network is converted into several time series feature vectors, where each time series feature vector contains the tea texture details, spectral chemical composition, equipment operating parameters, and processing environment parameters at the current moment;

[0078] The forward propagation layer captures the immediate changes of the blade state caused by the adjustment of the processing parameters through the update gate, and the backward propagation layer uses the reset gate to filter the lagged changes of the blade state caused by the adjustment of the processing parameters to generate temporal causal features;

[0079] In the spatial dimension, the data input to the BiGRU network is spatially convolved using a convolutional layer to obtain a convolutional feature map, and the convolutional feature map is globally averaged pooled in the channel dimension to generate a spatial significance weight vector;

[0080] Obtain the temporal hidden state of the BiGRU, perform matrix multiplication with the spatial significance weight vector, calculate the importance score of each channel in the temporal context, perform channel weighting on the original convolution feature map, and obtain the spatial response enhancement feature;

[0081] The temporal causal features and spatial response enhancement features are input into the fully connected layer for feature vector fusion to generate multimodal tea processing fusion features when the target processing equipment is processing tea.

[0082] It's important to note that during the multimodal feature fusion process for tea processing, cross-modal monitoring information is first expanded temporally into a sequence of continuous time-series feature vectors. Each vector integrates the current tea surface texture details (such as contrast calculated from the gray-level co-occurrence matrix), chemical component concentrations (such as tea polyphenol content) derived from near-infrared spectroscopy, equipment operating parameters (such as drum speed and motor current waveform), and environmental parameters (such as the temperature and humidity gradient within the drum). This information is fed into a bidirectional gated recurrent network (BiGRU). The forward propagation layer dynamically adjusts the proportion of historical state information retained through update gates, capturing the immediate impact of processing parameter adjustments on leaf physical and chemical properties. For example, a minute-long response to a 5°C increase in withering temperature leads to a change in chlorophyll degradation rate. The backward propagation layer utilizes reset gates to screen key nodes in historical processing events and explore the lag effects of parameter adjustments, such as the persistent impact of the cumulative rolling pressure gradient on polyphenol oxidase activity during subsequent fermentation. This generates latent state features that encode temporal causal relationships.

[0083] In spatial dimension processing, the original multimodal data is processed through a three-dimensional convolutional layer to extract local spatial features, generating a feature map that includes leaf morphology, equipment structure texture, and thermal field distribution. Global average pooling is performed on this feature map in the channel dimension, compressing the spatial information to generate a weight vector that represents the significance of each channel. For example, the weight of the heat conduction feature channel in the high-temperature area is automatically increased during the withering process. The temporal latent state matrix output by the BiGRU is matrix-multiplied with the spatial weight vector to calculate the importance score of each feature channel in the temporal context. After the scoring result is Sigmoid normalized, the original convolution feature map is channel-weighted to enhance spatial responses that are strongly correlated with the current processing stage, such as focusing on enhancing the characteristic response intensity of the leaf wrinkled edge at the end of drying. Finally, the temporal causal features generated by the BiGRU and the spatially enhanced convolution features are input into a fully connected network for cross-dimensional fusion, eliminating information redundancy between modalities through a weight sharing mechanism. A residual connection structure is introduced in the fusion process to retain the physical interpretability of the original features. The generated multimodal fusion features simultaneously encode the spatiotemporal propagation laws of process parameter adjustments, the boundary constraints of equipment operating status, and the cross-scale correlation of blade quality evolution, providing feature expressions with high discrimination and strong generalization capabilities for intelligent decision-making.

[0084] Furthermore, in a preferred embodiment of the present invention, the tea processing quality perception model is constructed, and the multimodal tea processing fusion features are input into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, specifically including:

[0085] Acquire historical processing instances of different tea processing qualities based on historical data retrieval, extract historical tea processing features when tea is processed in each historical processing instance according to the acquired historical processing instances, and obtain historical tea processing feature information;

[0086] Extracting the tea processing flow corresponding to each historical processing instance, performing feature partitioning on the historical tea processing feature information, and generating tea processing feature subsets of several different historical processing instances, each tea processing feature subset corresponding to a different tea processing flow;

[0087] The historical tea processing feature information after feature partitioning is associated with the corresponding historical tea processing quality, and the tea processing feature is used as the first node, the tea processing quality is used as the second node, and the tea processing process is defined as the node label of the first node to establish a topological structure diagram;

[0088] A tea processing quality perception model was constructed based on a long short-term memory network and a graph neural network. The historical tea processing feature information after feature partitioning was input into the long short-term memory network to extract multi-scale temporal features of different processing qualities.

[0089] The adjacency matrix is ​​obtained through the established topological structure graph, and the nodes are aggregated and updated based on the message passing mechanism to obtain the updated node representation. The updated node representation of the entire graph is globally pooled to obtain spatial features, and the multi-scale temporal features are fused with the spatial features to generate spatiotemporal fusion features.

[0090] The spatiotemporal fusion features are input into the fully connected layer to perform tea processing quality perception. The tea processing quality perception results are verified and parameters are adjusted using a preset validation set. After iterative training, a tea processing quality perception model that meets the expectations is obtained.

[0091] Acquire multimodal tea processing fusion features, input the multimodal tea processing fusion features into a trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtain tea processing quality perception information.

[0092] It should be noted that processing instance data of different quality levels is first retrieved from the historical database. The multi-dimensional processing features corresponding to each instance (such as the temperature curve during the withering stage, the rolling pressure gradient, and the humidity fluctuation of the fermentation environment) are extracted. The feature sequence is then dynamically partitioned based on the processing stage (spreading, withering, rolling, fermentation, and drying) to form feature subsets strongly associated with specific processes. A topological relationship network is constructed using graph structure modeling technology, in which each feature subset serves as a graph node. The node attributes encode feature statistics, and the edge weights between nodes are determined by the process connection relationship and the mutual information of features. The historical processing quality (sensory evaluation scores and physical and chemical index test values) is used as the target attribute of the graph node, and the nonlinear relationship between features and quality is learned using a graph embedding algorithm. A tea processing quality perception model was constructed. Its core is a combination of a long short-term memory (LSTM) network and a graph neural network (GNN). The LSTM network models time series features of a subset of process features, capturing the dynamic evolution of process parameters within each process (e.g., the hourly impact of humidity control during fermentation on polyphenol oxidase activity). The GNN, based on an adjacency matrix, propagates topological correlation information across processes. It aggregates features of adjacent nodes using a message passing mechanism (e.g., a rolling process node receives residual enzyme activity features from the withering process). A node update function generates spatial features encoding the coupling effects between processes. The fusion of spatiotemporal features is achieved through a cascaded approach: the multi-scale temporal features output by the LSTM (including minute-level fluctuations within a process and hour-level trends across processes) are concatenated with the spatial topological features generated by the GNN (reflecting the material and energy transfer relationships between processes) into a joint vector, which is then input into a fully connected network for nonlinear mapping. During training, a curriculum learning strategy is employed to optimize the model in stages: initially, it focuses on capturing temporal patterns within a process, while later, it strengthens the modeling of spatial correlations across processes. A dynamic weighted loss function is used to balance the contribution of spatiotemporal features. Finally, the multimodal fusion features of the current processing batch (including real-time process parameters, equipment status and tea characteristics) are input into a fully trained perception model. The model uses a spatiotemporal feature decoder to analyze the similarity between the processing status and historical high-quality cases, and outputs quantitative evaluation results covering immediate quality scores (such as the current withering uniformity index) and final quality predictions (estimated sensory scores of finished tea), providing a basis for subsequent decision-making.

[0093] Figure 2 A flow chart of an intelligent decision-making and auxiliary method for tea processing provided by one embodiment of the present invention;

[0094] like Figure 2 As shown, the present invention provides a flow chart of an intelligent decision-making and auxiliary method for a tea processing process, including:

[0095] S202, obtaining tea processing quality perception information, and generating a tea quality change perception curve of a current processing batch based on the tea processing quality perception information, wherein the tea quality change perception curve represents the tea processing quality of the current processing batch of tea at different time stamps;

[0096] S204, obtaining the tea quality requirements for the current processing batch, calculating the deviation between the tea processing quality and the tea quality requirements corresponding to different time stamps based on the tea quality change perception curve, and comparing the calculated deviation with a preset tolerance range;

[0097] S206: If the value is greater than the preset tolerance range, it means that the current tea processing control parameters do not meet the quality requirements of the tea of ​​the current processing batch, and then the preset processing plan of the current processing batch is obtained to optimize the processing plan;

[0098] S208, retrieving historical processing cases that meet the current tea quality requirements and processed tea varieties from a preset historical processing database, and extracting historical processing control plans corresponding to each historical processing case;

[0099] S210, introducing a multi-objective gray wolf optimization algorithm, initializing the population according to the historical processing control scheme corresponding to each historical processing case, presetting the objective function and setting constraints, calculating the objective function value of each individual in the initial gray wolf population through the objective function, and performing non-dominated sorting to divide all individuals in the population into different frontier levels;

[0100] S212, performing congestion calculation on each frontier layer to obtain the congestion distance, selecting the individual with the highest congestion in each layer as the leader wolf, and the remaining individuals as follower wolves, and calculating the direction vector of the leader wolf in the decision space to update its position, and outputting the optimal solution set after repeated iterations until the stopping condition is met;

[0101] S214, generating several candidate processing optimization schemes based on the output optimal solution set, performing feasibility screening on each candidate processing optimization scheme, and selecting the optimal processing optimization scheme to assist in tea processing based on the feasibility screening results.

[0102] It should be noted that a dynamic quality evolution curve is constructed based on real-time perception of tea processing quality. This curve maps the continuous changing trends of key tea quality indicators (such as tea polyphenol oxidation rate, chlorophyll retention, and sensory score predictions) over a timeline. By comparing the actual quality indicators at each process node against the target values ​​(e.g., thresholds such as moisture content ≤6% and phenol-ammonia ratio of 2.5-3.2 in premium tea standards), a multidimensional deviation vector is calculated. When the composite deviation index of a key process (such as the end point of withering and the turning point of fermentation) exceeds the preset tolerance threshold, the current process parameters are determined to be mismatched with the target quality requirements, triggering an adaptive optimization mechanism. High-quality processing examples of the same tea type are retrieved from a historical case library, and the mapping relationship between their process control parameters (temperature gradient, pressure curve, humidity setpoint) and quality compliance rate is extracted as prior knowledge for the optimization algorithm. A multi-objective gray wolf optimization algorithm is used to construct a parameter optimization framework. During the initialization phase, control solutions from historically high-performing cases are encoded as gray wolf individuals, and Latin hypercube sampling is used to expand the population diversity. The objective function is defined as a weighted combination of minimizing quality deviation, maximizing energy efficiency, and minimizing equipment loss rate. Constraints include equipment physical limits (such as maximum drum speed and maximum heating plate power) and process specifications (such as minimum fermentation time). During the algorithm iteration, a non-dominated sorting algorithm divides the population into multiple Pareto front levels. The crowding distance of individuals in each level is calculated to select the most representative leader individuals, guiding the population towards the optimal solution domain. Each iteration implements a leader-guided position update strategy: the leader wolf is selected as the individual with the highest crowding degree in each level, and the remaining individuals are followers. Based on their parameter combinations, a gravitational vector in the decision space is calculated, driving the followers to explore along the optimal quality-energy-efficiency equation. An adaptive mutation operator is designed to address the physical constraints of the equipment. When a parameter combination exceeds the feasible region, a reflective correction is applied along the normal vector of the constraint boundary to ensure the engineering feasibility of the solution. After a preset number of generations of evolutionary calculations, a set of candidate solutions covering multi-objective balance is output. 3-5 optimal solutions are screened out through the feasibility verification module, and finally the entropy weight-TOPSIS decision-making method is used to select the control parameter set with the highest comprehensive score. The solution is then sent to the processing equipment actuator in real time through the edge computing unit, forming a closed-loop control link of "perception-evaluation-optimization-execution", realizing adaptive and precise control of the processing process.

[0103] Figure 3 An embodiment of the present invention provides a tea processing assistance system 3 based on intelligent decision-making. The system includes: a memory 31 and a processor 32. The memory 31 contains a tea processing assistance method program based on intelligent decision-making. When the tea processing assistance method program based on intelligent decision-making is executed by the processor 32, the following steps are implemented:

[0104] Performing tea processing monitoring on target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information;

[0105] Analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing a multimodal fusion feature to obtain a multimodal tea processing fusion feature;

[0106] Constructing a tea processing quality perception model, inputting the multimodal tea processing fusion features into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtaining tea processing quality perception information;

[0107] Based on the tea processing quality perception information, it is determined whether the quality of the tea in the current processing batch meets expectations. If not, the current processing plan is optimized to assist in tea processing.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0109] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0110] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0111] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0112] Alternatively, if the integrated units described above are implemented as software modules and sold or used as standalone products, they can also be stored on a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product, stored on a storage medium, includes instructions for enabling a computer device (such as a personal computer, server, or network device) to execute all or part of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as removable storage devices, ROM, RAM, magnetic disks, or optical disks.

[0113] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A tea processing auxiliary method based on intelligent decision-making, characterized in that: include: Performing tea processing monitoring on target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information; Analyzing the cross-modal correlation relationship of each monitoring data based on the cross-modal tea processing monitoring information, and constructing a multimodal fusion feature to obtain a multimodal tea processing fusion feature; Constructing a tea processing quality perception model, inputting the multimodal tea processing fusion features into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtaining tea processing quality perception information; Determining whether the quality of the tea leaves of the current processing batch meets expectations based on the tea processing quality perception information, and if not, optimizing the current processing plan to assist in tea processing; The process of constructing a tea processing quality perception model and inputting the multimodal tea processing fusion features into the trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch specifically includes: Acquire historical processing instances of different tea processing qualities based on historical data retrieval, extract historical tea processing features when tea is processed in each historical processing instance according to the acquired historical processing instances, and obtain historical tea processing feature information; Extracting the tea processing flow corresponding to each historical processing instance, performing feature partitioning on the historical tea processing feature information, and generating tea processing feature subsets of several different historical processing instances, each tea processing feature subset corresponding to a different tea processing flow; The historical tea processing feature information after feature partitioning is associated with the corresponding historical tea processing quality, and the tea processing feature is used as the first node, the tea processing quality is used as the second node, and the tea processing process is defined as the node label of the first node to establish a topological structure diagram; A tea processing quality perception model was constructed based on a long short-term memory network and a graph neural network. The historical tea processing feature information after feature partitioning was input into the long short-term memory network to extract multi-scale temporal features of different processing qualities. The adjacency matrix is ​​obtained through the established topological structure graph, and the nodes are aggregated and updated based on the message passing mechanism to obtain the updated node representation. The updated node representation of the entire graph is globally pooled to obtain spatial features, and the multi-scale temporal features are fused with the spatial features to generate spatiotemporal fusion features. The spatiotemporal fusion features are input into the fully connected layer to perform tea processing quality perception. The tea processing quality perception results are verified and parameters are adjusted using a preset validation set. After iterative training, a tea processing quality perception model that meets the expectations is obtained. Acquire multimodal tea processing fusion features, input the multimodal tea processing fusion features into a trained tea processing quality perception model to perform processing quality perception on the tea of ​​the current processing batch, and obtain tea processing quality perception information.

2. The tea processing auxiliary method based on intelligent decision-making according to claim 1, characterized in that: The tea processing monitoring of the target processing equipment to obtain tea processing monitoring data, and preprocessing the obtained tea processing monitoring data to obtain cross-modal tea processing monitoring information specifically includes: When the target processing equipment is processing tea, the installed sensor array is used to monitor the target processing equipment for tea processing to obtain tea processing monitoring data, wherein the tea processing monitoring data includes processing environment monitoring data, processing equipment operation data, and processed tea monitoring data; Performing wavelet packet decomposition on the tea processing monitoring data using a preset wavelet basis function to obtain different sub-bands and calculating the energy entropy of each sub-band, and performing soft threshold filtering based on the calculated energy entropy to identify effective characteristic frequency bands; After filtering, the tea processing monitoring data was time-series aligned, noise was removed using median filtering, and motion blur was eliminated through Gaussian smoothing. A sliding window algorithm was introduced to slide based on a preset window and calculate the mean and standard deviation of the data within the window. A linear interpolation algorithm was then used for data cleaning. Maximum-minimum normalization is performed on the tea processing monitoring data after cleaning, and the continuous data is mapped to the interval [0,1]. The discrete data is standardized through binning to generate cross-modal tea processing monitoring information.

3. The tea processing auxiliary method based on intelligent decision-making according to claim 1, characterized in that: The cross-modal correlation relationship of each monitoring data is analyzed based on the cross-modal tea processing monitoring information, and a multi-modal fusion feature is constructed to obtain a multi-modal tea processing fusion feature, which specifically includes: Obtaining cross-modal tea processing monitoring information, obtaining processed tea monitoring data through the cross-modal tea processing monitoring information and extracting processed tea image monitoring data, and importing the processed tea image monitoring data into the HSV color space to obtain a histogram distribution; The histogram distribution is used to calculate the local area hue mean and saturation variance, and then a weighted fusion is performed to form the browning degree index. Subsequently, a morphological algorithm is used to extract the contour of the maximum connected area of ​​the processed tea image. The ratio between the contour perimeter and the tea area is calculated as the curling index, and the first feature information is generated by combining it with the texture characteristics of the processed tea. Extracting near-infrared spectral data from processed tea monitoring data and calculating the first-order derivative to generate a near-infrared derivative spectrum, inputting the near-infrared derivative spectrum into a pre-trained least squares regression model to invert the chemical composition of the tea leaves and generate a spectral domain feature vector of the processed tea leaves, thereby obtaining second feature information; A multi-head attention mechanism is introduced to use the first feature information as the query vector, the second feature information as the key vector and the value vector, and the scaled click attention score is calculated to perform feature fusion and generate fused feature information; Cross-modal tea processing monitoring information is used to extract processing environment monitoring data and processing equipment operation data, which are combined with the fusion feature information and input into the BiGRU network for multimodal feature fusion to obtain multimodal tea processing fusion features.

4. The tea processing auxiliary method based on intelligent decision-making according to claim 3, characterized in that: The cross-modal tea processing monitoring information is used to extract processing environment monitoring data and processing equipment operation data, and the fused feature information is combined and input into the BiGRU network for multimodal feature fusion, specifically including: In the time dimension, the data input into the BiGRU network is converted into several time series feature vectors, where each time series feature vector contains the tea texture details, spectral chemical composition, equipment operating parameters, and processing environment parameters at the current moment; The forward propagation layer captures the immediate changes of the blade state caused by the adjustment of the processing parameters through the update gate, and the backward propagation layer uses the reset gate to filter the lagged changes of the blade state caused by the adjustment of the processing parameters to generate temporal causal features; In the spatial dimension, the data input to the BiGRU network is spatially convolved using a convolutional layer to obtain a convolutional feature map, and the convolutional feature map is globally averaged pooled in the channel dimension to generate a spatial significance weight vector; Obtain the temporal hidden state of the BiGRU, perform matrix multiplication with the spatial significance weight vector, calculate the importance score of each channel in the temporal context, perform channel weighting on the original convolution feature map, and obtain the spatial response enhancement feature; The temporal causal features and spatial response enhancement features are input into the fully connected layer for feature vector fusion to generate multimodal tea processing fusion features when the target processing equipment is processing tea.

5. The tea processing auxiliary method based on intelligent decision-making according to claim 1, characterized in that: The method of judging whether the quality of the tea leaves of the current processing batch meets expectations based on the tea leaves processing quality perception information, and if not, optimizing the current processing plan to assist in tea leaves processing, specifically includes: Acquiring tea processing quality perception information, and generating a tea quality change perception curve of a current processing batch based on the tea processing quality perception information, wherein the tea quality change perception curve represents the tea processing quality of the current processing batch of tea at different time stamps; Obtaining the tea quality requirements for the current processing batch, calculating the deviation between the tea processing quality and the tea quality requirements corresponding to different time stamps based on the tea quality change perception curve, and comparing the calculated deviation with a preset tolerance range; If it is greater than the preset tolerance range, it means that the current tea processing control parameters are not suitable for the tea quality requirements of the current processing batch, then the preset processing plan of the current processing batch is obtained to optimize the processing plan; Retrieving historical processing cases that meet the current tea quality requirements and processed tea varieties from a preset historical processing database, and extracting historical processing control plans corresponding to each historical processing case; A multi-objective gray wolf optimization algorithm is introduced. The population is initialized according to the historical processing control scheme corresponding to each historical processing case. The objective function is preset and the constraints are set. The objective function value of each individual in the initial gray wolf population is calculated through the objective function, and all individuals in the population are divided into different frontier levels through non-dominated sorting. Congestion calculation is performed on each frontier layer to obtain the congestion distance. The individual with the highest congestion in each layer is selected as the leader wolf, and the remaining individuals are used as follower wolves. The direction vector of the leader wolf in the decision space is calculated to update its position. The optimal solution set is output after repeated iterations until the stopping condition is met. Based on the output optimal solution set, several candidate processing optimization schemes are generated, and the feasibility of each candidate processing optimization scheme is screened. Based on the feasibility screening results, the optimal processing optimization scheme is selected to assist tea processing.

6. A tea processing auxiliary system based on intelligent decision-making, characterized in that: The system includes: a memory and a processor, wherein the memory contains a tea processing auxiliary method program based on intelligent decision-making, and when the tea processing auxiliary method program based on intelligent decision-making is executed by the processor, any one of the steps of the tea processing auxiliary method based on intelligent decision-making described in claims 1-5 is implemented.

Images