Tea leaf fixation method based on deep learning

Abstract

The invention discloses a tea leaf fixation method based on deep learning, which comprises the following steps: S1, collecting tea leaf data in a fixation process, and preprocessing; s2, color change features, leaf curl features and temperature and humidity change features are extracted respectively, and feature alignment operation is executed; s3, executing a time sequence resampling operation on the cross-modal fusion feature representation, extracting a key fixation change frame and deleting a redundant frame; s4, extracting local and global features from the compressed state sequence; s5, inputting the state vector and the stage division label into an autoregression controller, and outputting a control parameter adjustment amount; s6, the heating temperature, the hot air moisture removal proportion and the stir-frying rotating speed of the fixation equipment are adjusted; and S7, periodically inputting the incremental training samples into the network model for parameter updating. According to the invention, multi-mode state accurate identification and intelligent closed-loop control in the tea leaf fixation process are realized, and the consistency and control precision of fixation quality are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of food processing technology, and in particular to a tea withering method based on deep learning. Background Art

[0002] Withering tea leaves is a crucial step in the entire tea processing process. Its function is to inactivate the enzyme activity in the fresh leaves, preventing the tea from oxidizing and deteriorating. At the same time, high temperatures destroy the tissue structure of the fresh leaves, promoting the formation of aroma and fixing the color, thus affecting the quality and style of the final tea. Traditional withering operations rely mainly on manual control, including setting the heating temperature, determining the frequency of stir-frying, and controlling the withering time. The accumulation of long-term manual operations has formed a relatively fixed withering process. However, this process relies on human judgment, is highly subjective, has poor replicability, is difficult to achieve standardized control, and is not adaptable to large-scale, automated production.

[0003] Currently, some modern tea processing factories have begun to implement semi-automated tea withering processes using temperature control equipment, stir-frying machinery, and hot air systems. However, most systems still rely on fixed control logic based on preset parameters and lack the ability to perceive the tea leaves' state in real time. Traditional control methods often use equipment operating time as the primary control variable, while ignoring multi-dimensional state characteristics such as tea color changes, leaf curl, and moisture content. This can easily lead to insufficient or excessive withering, resulting in unstable tea quality. Furthermore, traditional systems lack effective feedback mechanisms, making it impossible to adaptively optimize control parameters based on real-time changes during the withering process, which can affect the consistency of the finished tea leaves in terms of aroma, flavor, and appearance.

[0004] In terms of identifying the state of tea leaves, existing studies have attempted to use cameras to capture leaf images or infrared sensors to obtain temperature data, and use shallow image processing methods or expert systems to make rough judgments. However, such methods are difficult to accurately model the dynamic evolution trend of tea leaves during the withering process, and their processing capabilities are low, making it difficult to respond in real time. In terms of control strategy, some systems use rule-driven control algorithms to adjust the heating temperature and stir-frying rate. However, this type of control method is essentially a static mapping and cannot perceive the stage differences in the withering process, let alone make personalized adjustments for different tea varieties, environmental humidity, or initial moisture content. In addition, existing methods have serious shortcomings in multimodal information fusion, and it is difficult to simultaneously process multiple types of data such as images, temperature and humidity, and odor, resulting in insufficient understanding of the withering state of tea leaves and limited control accuracy.

[0005] To address the state recognition and control issues during the withering process, some studies have introduced convolutional neural networks to extract tea image features and fine-tune parameters in combination with simple control strategies. However, these models generally lack time series processing capabilities and are unable to capture the continuous dynamic changes during the withering process. They also suffer from problems such as inaccurate identification of "frame breakpoints" and fuzzy "stage divisions." Especially during the withering process, the state of tea leaves in different time periods has obvious stage characteristics, such as the heating period, the main withering period, and the cooling period. Current deep learning models lack the ability to accurately model these evolutionary stages, resulting in delayed response or over-adjustment of control strategies, seriously affecting product quality stability.

[0006] Furthermore, in terms of feedback closed-loop control, a complete intelligent cycle mechanism of state perception-control response-feedback update has not yet been formed. Most systems only have a one-way control path, making it difficult to achieve real-time synchronization between the operating status of the withering equipment and the actual state of the tea leaves. In the absence of effective feedback, the control strategy cannot self-update and iterate, and the overall intelligence level of the system is low, lacking the ability to learn online and continuously optimize. At the same time, in terms of structured management of equipment operation data, traditional solutions only record data in a log format and lack the ability to integrate key control parameters such as temperature, humidity, wind speed, and rotation speed in a unified format, hindering the subsequent training and updating of data-driven models.

[0007] Therefore, how to provide a tea withering method based on deep learning is an urgent problem that those skilled in the art need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a tea withering method based on deep learning. The present invention fully integrates convolutional neural networks, time series resampling mechanism and Transformer model, constructs a multimodal state recognition and intelligent control closed-loop system, and describes in detail the dynamic extraction method of color change characteristics, leaf curling characteristics and temperature and humidity change characteristics during the tea withering process. It has the advantages of high recognition accuracy, fast control response and strong stability of withering quality.

[0009] The tea withering method based on deep learning according to an embodiment of the present invention includes the following steps:

[0010] S1, collect tea data during the withering process through sensors and perform preprocessing;

[0011] S2. Input the preprocessed tea data into the convolutional neural network to extract color change features, leaf curl features, and temperature and humidity change features, perform feature alignment operations, and construct a cross-modal fusion feature representation;

[0012] S3. Perform temporal resampling on the cross-modal fusion feature representation, use the attention mechanism to extract key change frames and delete redundant frames, and output the compressed state sequence after temporal alignment;

[0013] S4. Input the compressed state sequence into the Transformer model, extract the local dynamic structure features and global evolution trend features, and output the state vector and stage division label of the killing process;

[0014] S5. Input the state vector and the stage division label into the autoregressive controller, combine it with the preprocessed tea data, perform nonlinear control mapping, and output the control parameter adjustment amount;

[0015] S6. Adjust the heating temperature, hot air dehumidification ratio, and stir-frying speed of the fixing equipment, and simultaneously record the adjusted equipment operating status and fixing process feedback data;

[0016] S7. Combine the tea data and feedback data into label-aligned incremental training samples, and periodically input them into the convolutional neural network and Transformer model for parameter update.

[0017] Optionally, the tea data includes image data, temperature and humidity data, color and texture data, and odor signal data.

[0018] Optionally, the preprocessing includes filling missing data, removing outliers, normalizing values ​​and aligning timestamps.

[0019] Optionally, the S3 specifically includes:

[0020] S31, expanding the cross-modal fusion feature representation in the time dimension to form a multi-channel time feature set consisting of a color change feature frame sequence, a leaf curl feature frame sequence, and a temperature and humidity change feature frame sequence;

[0021] S32. Constructing an inter-frame weight matrix based on the multi-channel temporal feature set, and calculating the importance score of each frame based on the color change amplitude, curl morphology change rate, and temperature and humidity fluctuation trend;

[0022] S33, extracting key completion change frames according to the importance score sorting, and eliminating invalid frames with scores lower than a set threshold to form a preliminary key frame sequence;

[0023] S34, performing adjacent frame similarity detection on the preliminary key frame sequence, calculating the inter-frame distance based on color similarity, morphological proximity, and thermal and humidity state overlap, and merging redundant frames when the inter-frame distance is lower than a set threshold;

[0024] S35, performing time window sliding sampling on the key frame sequence after de-redundancy, setting the window length and sliding step size, and forming a compressed time series with balanced distribution between frames;

[0025] S36 , performing a temporal consistency check on the compressed time sequence, removing frames with discontinuous timestamps and skipped frames, and outputting a compressed state sequence with consistent temporal order.

[0026] Optionally, the S4 specifically includes:

[0027] S41. Construct a Transformer model structure including a multi-layer encoder and decoder, wherein the encoder is composed of a position encoding module, a multi-head attention module, and a feedforward calculation module, and the decoder is composed of a context backpropagation module, a residual connection module, and a normalization module;

[0028] S42, performing position encoding processing on the compressed state sequence, embedding the position information of each frame on the time axis into the feature sequence, and forming a vector sequence with a time mark;

[0029] S43. Input the vector sequence with time stamp into the multi-head attention module, calculate the attention relationship between different attention heads, and extract the correlation information of color change structure, leaf edge disturbance structure and temperature and humidity mutation structure;

[0030] S44, the output of the multi-head attention module is residually connected to the original input and input into the feedforward calculation module to perform nonlinear transformation to extract local dynamic structure features;

[0031] S45. The encoder output is passed to the decoder structure to reversely model the cross-temporal variation characteristics between the states of each stage to form a global evolution trend feature covering the entire killing process;

[0032] S46, combining the local dynamic structure features with the global evolution trend features, and inputting the combined features into a classification mapping unit, thereby outputting a state vector representing the state of the fixing process;

[0033] S47. Perform time period boundary detection on the state vector, automatically divide the completion stage interval according to the feature change density, and output the stage division label.

[0034] Optionally, the stage division label is generated based on an adaptive boundary detection algorithm. The boundary detection algorithm performs a multi-scale sliding window operation on the global evolution trend features output by the Transformer decoder, calculates the color change gradient mean, the leaf curling feature density change value and the temperature and humidity fluctuation stability index in each time period, and constructs a feature change function sequence. The first-order derivative turning point of the feature change function is used as a candidate segmentation point, and the stable boundary point set is screened by the density clustering method. Finally, the state vector sequence is divided into multiple stage intervals, and the stage division label corresponding to each stage is output.

[0035] Optionally, the S5 specifically includes:

[0036] S51, constructing an autoregressive controller structure, wherein the autoregressive controller is composed of a feature extraction unit, a stage label encoding unit, a historical action backtracking unit, and a nonlinear mapping unit;

[0037] S52, inputting the state vector into a feature extraction unit, extracting the color offset, curling trend fluctuation value, and temperature and humidity gradient change value of the current fixing state, and constructing a continuous state feature sequence;

[0038] S53, inputting the stage division labels into the stage label encoding unit, using a fixed encoding rule to convert each stage label into a stage control weight vector to form a stage identification control matrix;

[0039] S54: Input the pre-processed tea data into the historical action backtracking unit, perform difference calculation based on the current state vector and the previous round of control parameters, and generate a historical difference vector for adjusting the target;

[0040] S55, concatenating the state feature sequence, the stage identification control matrix, and the historical difference vector in the feature dimension, and inputting the concatenated concatenation into a nonlinear mapping unit, wherein the nonlinear mapping unit is composed of a multi-layer activation function;

[0041] S56. In the nonlinear mapping unit, perform feature weighting calculation, normalization mapping and parameter projection operations in sequence to generate a fixing control parameter adjustment amount, which is used to adjust the operating instructions of the fixing device.

[0042] Optionally, the feature extraction unit performs continuous feature trajectory modeling on the state vector based on a multi-layer convolution structure, color offset, curl trend fluctuation value and temperature and humidity gradient change value, and the stage label encoding unit adopts a fixed-dimensional vector encoding scheme to map the stage division label into a stage control weight vector.

[0043] Optionally, the S6 specifically includes:

[0044] S61. Drive the fixing device according to the control parameter adjustment amount, adjust the output power of the heating unit to the target temperature range, and continuously monitor the temperature change curve of the heating plate surface, recording the temperature response time and the steady-state maintenance time;

[0045] S62, adjusting the fan speed and the opening and closing range of the dehumidification valve, setting the hot air output flow rate and the dehumidification time window, dynamically correcting the humidity and wind speed parameters in the hot air circulation channel, and recording the change sequence of the temperature and humidity coupling value at the hot air outlet;

[0046] S63, comprehensively control the stir-frying speed, turning amplitude and stir-frying interval, and record the mechanical response delay and operation stability index of each cycle of the fixing equipment in real time;

[0047] S64. Based on the heating temperature recording data, hot air dehumidification data and stir-frying speed data, a device operation status set is constructed, and a two-way recording data set is formed with the status feedback data corresponding to the current fixing stage, and stored in the local data cache unit.

[0048] Optionally, the equipment operation status set is derived from heating temperature recording data, hot air dehumidification data and stir-frying speed data. The specific parameters include heating temperature sequence, hot air flow trajectory, humidity adjustment parameters, wind speed change deviation, stir-frying speed curve, flipping amplitude range and stir-frying interval time, and are synchronously calibrated in chronological order to form a structured state set.

[0049] The beneficial effects of the present invention are:

[0050] First, the present invention significantly improves the accuracy of identifying state changes during the tea withering process by constructing a cross-modal feature extraction mechanism that integrates color change features, leaf curling features, and temperature and humidity change features, avoiding the error accumulation problem of traditional judgment relying on a single image or temperature data, and effectively solving the technical bottlenecks of inaccurate judgment of the degree of withering and fuzzy state evolution.

[0051] Secondly, the present invention introduces a time-series resampling algorithm based on the attention mechanism and a Transformer model structure, which can dynamically extract key killing change frames and model the evolution trend of the entire process. It has the ability to automatically divide the killing stages and accurately express the state trends, further enhancing the decision-making rationality and response agility of the system in multi-time period control, and ensuring the stability and consistency of the killing process.

[0052] Finally, the present invention combines the autoregressive control strategy to map the state vector and stage division label into continuous control parameter adjustment quantities, driving the tea withering equipment to achieve adaptive regulation in multiple dimensions such as heating, dehumidification and stir-frying. At the same time, it constructs a complete feedback record and sample update mechanism, and has the ability of continuous optimization and autonomous learning, realizing the transformation of the tea withering process from experience-driven to intelligent closed-loop control. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0054] Figure 1 This is a flow chart of the tea withering method based on deep learning proposed by the present invention;

[0055] Figure 2 This is a schematic diagram of the convolutional neural network feature extraction and fusion of the tea withering method based on deep learning proposed in the present invention;

[0056] Figure 3 This is a schematic diagram of the Transformer model structure flow of the tea withering method based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0057] 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.

[0058] refer to Figure 1-3 The tea withering method based on deep learning includes the following steps:

[0059] S1, collect tea data during the withering process through sensors and perform preprocessing;

[0060] S2. Input the preprocessed tea data into the convolutional neural network to extract color change features, leaf curl features, and temperature and humidity change features, perform feature alignment operations, and construct a cross-modal fusion feature representation;

[0061] S3. Perform temporal resampling on the cross-modal fusion feature representation, use the attention mechanism to extract key change frames and delete redundant frames, and output the compressed state sequence after temporal alignment;

[0062] S4. Input the compressed state sequence into the Transformer model, extract the local dynamic structure features and global evolution trend features, and output the state vector and stage division label of the killing process;

[0063] S5. Input the state vector and the stage division label into the autoregressive controller, combine it with the preprocessed tea data, perform nonlinear control mapping, and output the control parameter adjustment amount;

[0064] S6. Adjust the heating temperature, hot air dehumidification ratio, and stir-frying speed of the fixing equipment, and simultaneously record the adjusted equipment operating status and fixing process feedback data;

[0065] S7. Combine the tea data and feedback data into label-aligned incremental training samples, and periodically input them into the convolutional neural network and Transformer model for parameter update.

[0066] The present invention realizes dynamic perception and closed-loop optimization control of the entire tea withering process by constructing a complete process from tea data collection, deep feature extraction, time series modeling, intelligent control to model updating, significantly improving the consistency of withering quality and the intelligence level of the system.

[0067] In this embodiment, the tea data includes image data, temperature and humidity data, color and texture data, and odor signal data.

[0068] The present invention enhances the multi-angle understanding ability of tea state by introducing multimodal inputs such as image data, temperature and humidity data, color and texture data, and odor signal data, and improves the robustness of the network to the withering process in complex environments.

[0069] In this embodiment, the preprocessing includes missing data filling, outlier removal, value normalization and timestamp alignment.

[0070] The present invention ensures the continuity and structural consistency of input data by performing preprocessing operations such as missing data filling, outlier removal, numerical normalization and timestamp alignment, providing a reliable input basis for subsequent deep learning models.

[0071] In this embodiment, S3 specifically includes:

[0072] S31, expanding the cross-modal fusion feature representation in the time dimension to form a multi-channel time feature set consisting of a color change feature frame sequence, a leaf curl feature frame sequence, and a temperature and humidity change feature frame sequence;

[0073] S32. Constructing an inter-frame weight matrix based on the multi-channel temporal feature set, and calculating the importance score of each frame based on the color change amplitude, curl morphology change rate, and temperature and humidity fluctuation trend;

[0074] S33, extracting key completion change frames according to the importance score sorting, and eliminating invalid frames with scores lower than a set threshold to form a preliminary key frame sequence;

[0075] S34, performing adjacent frame similarity detection on the preliminary key frame sequence, calculating the inter-frame distance based on color similarity, morphological proximity, and thermal and humidity state overlap, and merging redundant frames when the inter-frame distance is lower than a set threshold;

[0076] S35, performing time window sliding sampling on the key frame sequence after de-redundancy, setting the window length and sliding step size, and forming a compressed time series with balanced distribution between frames;

[0077] S36 , performing a temporal consistency check on the compressed time sequence, removing frames with discontinuous timestamps and skipped frames, and outputting a compressed state sequence with consistent temporal order.

[0078] The present invention uses multi-channel temporal feature expansion and attention mechanism key frame extraction to accurately capture the core change moments in the completion process and eliminate redundant frames, thereby improving the efficiency of timing modeling and the accuracy of discrimination.

[0079] In this embodiment, the S4 specifically includes:

[0080] S41. Construct a Transformer model structure including a multi-layer encoder and decoder, wherein the encoder is composed of a position encoding module, a multi-head attention module, and a feedforward calculation module, and the decoder is composed of a context backpropagation module, a residual connection module, and a normalization module;

[0081] S42, performing position encoding processing on the compressed state sequence, embedding the position information of each frame on the time axis into the feature sequence, and forming a vector sequence with a time mark;

[0082] S43. Input the vector sequence with time stamp into the multi-head attention module, calculate the attention relationship between different attention heads, and extract the correlation information of color change structure, leaf edge disturbance structure and temperature and humidity mutation structure;

[0083] S44, the output of the multi-head attention module is residually connected to the original input and input into the feedforward calculation module to perform nonlinear transformation to extract local dynamic structure features;

[0084] S45. The encoder output is passed to the decoder structure to reversely model the cross-temporal variation characteristics between the states of each stage to form a global evolution trend feature covering the entire killing process;

[0085] S46, combining the local dynamic structure features with the global evolution trend features, and inputting the combined features into a classification mapping unit, thereby outputting a state vector representing the state of the fixing process;

[0086] S47. Perform time period boundary detection on the state vector, automatically divide the completion stage interval according to the feature change density, and output the stage division label.

[0087] The present invention constructs a Transformer model with an encoder and decoder structure, integrates the local dynamic structure with the global evolution trend, effectively characterizes the timing laws and stage characteristics of the tea withering process, and realizes fine state modeling.

[0088] In this embodiment, the stage division label is generated based on an adaptive boundary detection algorithm. The boundary detection algorithm performs a multi-scale sliding window operation on the global evolution trend characteristics output by the Transformer decoder, calculates the color change gradient mean, the leaf curling feature density change value and the temperature and humidity fluctuation stability index in each time period, and constructs a feature change function sequence. The first-order derivative turning point of the feature change function is used as a candidate segmentation point, and the stable boundary point set is screened by the density clustering method. Finally, the state vector sequence is divided into multiple stage intervals, and the stage division label corresponding to each stage is output.

[0089] The present invention constructs an adaptive boundary detection algorithm based on a multi-scale sliding window and the first-order derivative of the feature change function to accurately divide the stage intervals of the killing process, thereby improving the pertinence and response rationality of the controller.

[0090] In this embodiment, the S5 specifically includes:

[0091] S51, constructing an autoregressive controller structure, wherein the autoregressive controller is composed of a feature extraction unit, a stage label encoding unit, a historical action backtracking unit, and a nonlinear mapping unit;

[0092] S52, inputting the state vector into a feature extraction unit, extracting the color offset, curling trend fluctuation value, and temperature and humidity gradient change value of the current fixing state, and constructing a continuous state feature sequence;

[0093] S53, inputting the stage division labels into the stage label encoding unit, using a fixed encoding rule to convert each stage label into a stage control weight vector to form a stage identification control matrix;

[0094] S54: Input the pre-processed tea data into the historical action backtracking unit, perform difference calculation based on the current state vector and the previous round of control parameters, and generate a historical difference vector for adjusting the target;

[0095] S55, concatenating the state feature sequence, the stage identification control matrix, and the historical difference vector in the feature dimension, and inputting the concatenated concatenation into a nonlinear mapping unit, wherein the nonlinear mapping unit is composed of a multi-layer activation function;

[0096] S56. In the nonlinear mapping unit, perform feature weighting calculation, normalization mapping and parameter projection operations in sequence to generate a fixing control parameter adjustment amount, which is used to adjust the operating instructions of the fixing device.

[0097] The present invention realizes the dynamic update of parameter control and nonlinear mapping response by constructing an autoregressive controller composed of state characteristics, stage labels and historical control differences, thereby enhancing the accuracy and flexibility of control decisions.

[0098] In this embodiment, the feature extraction unit performs continuous feature trajectory modeling on the state vector based on a multi-layer convolution structure, including color offset, curling trend fluctuation value and temperature and humidity gradient change value. The stage label encoding unit adopts a fixed-dimensional vector encoding scheme to map the stage division label into a stage control weight vector.

[0099] The present invention continuously models the state vector through a multi-layer convolution structure and uses a vector encoding method to map the stage label, thereby enhancing the controller's ability to understand the stage-by-stage behavior characteristics and improving the discrimination ability of the control mapping.

[0100] In this embodiment, S6 specifically includes:

[0101] S61. Drive the fixing device according to the control parameter adjustment amount, adjust the output power of the heating unit to the target temperature range, and continuously monitor the temperature change curve of the heating plate surface, recording the temperature response time and the steady-state maintenance time;

[0102] S62, adjusting the fan speed and the opening and closing range of the dehumidification valve, setting the hot air output flow rate and the dehumidification time window, dynamically correcting the humidity and wind speed parameters in the hot air circulation channel, and recording the change sequence of the temperature and humidity coupling value at the hot air outlet;

[0103] S63, comprehensively control the stir-frying speed, turning amplitude and stir-frying interval, and record the mechanical response delay and operation stability index of each cycle of the fixing equipment in real time;

[0104] S64. Based on the heating temperature recording data, hot air dehumidification data and stir-frying speed data, a device operation status set is constructed, and a two-way recording data set is formed with the status feedback data corresponding to the current fixing stage, and stored in the local data cache unit.

[0105] The present invention realizes fine control and complete recording of the physical process through multi-parameter dynamic adjustment and data collection of three key behaviors of heating, dehumidification and stir-frying in the fixing equipment, providing data support for subsequent optimization.

[0106] In this embodiment, the equipment operation status set is derived from the heating temperature recording data, hot air dehumidification data and stir-frying speed data. The specific parameters include the heating temperature sequence, hot air flow trajectory, humidity adjustment parameters, wind speed change deviation, stir-frying speed curve, flipping amplitude range and stir-frying interval time, and are synchronously calibrated in chronological order to form a structured state set.

[0107] The present invention constructs a set of equipment operating states, unifies the multi-source operating parameters of heating, hot air and stir-frying, and synchronously calibrates them in a structured manner, providing a high-precision and time-consistent data foundation for model training and equipment diagnosis.

[0108] Example 1:

[0109] To verify the feasibility of this invention, we applied it to a tea production and processing scenario where semi-automated equipment was used for the withering process. The original system, which employed fixed heating temperatures and timed stir-frying logic, lacked the ability to perceive the tea's real-time state. This often led to problems such as uneven withering, delayed stir-frying, and unbalanced hot air dehumidification, directly impacting the tea's color, aroma, and morphological consistency. In this scenario, the present invention deployed an image acquisition device, a temperature and humidity sensor module, and a stir-frying drive controller to achieve data collection and intelligent control of the entire withering process.

[0110] In practical applications, the system first uses multimodal sensors to collect tea image data, temperature and humidity information, and equipment operating parameters in real time. After preprocessing, the collected data is fed into a deep convolutional network to extract the color change characteristics of the tea leaves, the degree of leaf curl, and the current temperature and humidity change trends, and fuse them into a unified feature representation. Subsequently, the system performs a temporal resampling operation through an attention mechanism, automatically extracts key change frames, and constructs a compressed state sequence. This sequence is input into the Transformer model to analyze the dynamic evolution trend of the current withering process and output a state vector and stage division labels. Based on these output values, the controller dynamically adjusts the heating temperature, hot air dehumidification ratio, and stir-frying speed, thereby achieving closed-loop optimization control.

[0111] During the one-week operation cycle, the system carried out withering operations on four batches of fresh leaves of different categories. Compared with the traditional control system, the method of the present invention shows obvious advantages in leaf color uniformity, moisture content compliance rate and stir-frying balance. Actual measurement data show that the variance of leaf color distribution in the withered tea samples controlled by the present invention decreased by 42.8%, the average moisture content was controlled in the range of 6.2% to 6.8%, and the fluctuation range was less than 0.6%; the equipment response delay during the stir-frying process decreased from an average of 1.3 seconds to 0.4 seconds, and the mechanical error was reduced by nearly 70%; at the same time, the sensory score of the terminal tea samples increased by 14.6%, and the customer satisfaction evaluation increased by about 25% compared with the original control system.

[0112] In addition, in terms of equipment operation, the structured operating status set established by the present invention records information such as the heating temperature curve, wind speed and humidity change trajectory, and stir-fry parameter adjustment sequence, providing high-quality data support for subsequent model iterative training. The actual sampling frequency is set to 20 frames per second, and the equipment operation data and the finishing status feedback data are stored synchronously. Each batch of operations generates an average of 148,000 structured sample data. The system automatically completes the label alignment process, which significantly improves the density and accuracy of the training data. After incremental learning updates, the model inference accuracy is further improved by 4.3%, and the overall response efficiency of the system is improved by nearly 30%.

[0113] In summary, the present invention demonstrates good adaptability and intelligent control capabilities in practical applications, especially in the multi-category and multi-batch withering production process, effectively solving the problems of inaccurate state recognition, control parameter lag and equipment response incoordination existing in traditional methods, and significantly improving the stability and automation level of tea withering quality.

[0114] Table 1 Comparison of tea-killing intelligent control effects

[0115]

[0116] This table shows the actual data of the control accuracy, action coordination and product quality improvement of the control system of the present invention during the withering process of different tea batches, further verifying the beneficial effects of the present invention in practical applications.

[0117] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A tea withering method based on deep learning, characterized in that: The steps include: S1, collect tea data during the withering process through sensors and perform preprocessing; S2. Input the preprocessed tea data into the convolutional neural network to extract color change features, leaf curl features, and temperature and humidity change features, perform feature alignment operations, and construct a cross-modal fusion feature representation; S3. Perform temporal resampling on the cross-modal fusion feature representation, use the attention mechanism to extract key change frames and delete redundant frames, and output the compressed state sequence after temporal alignment; S4. Input the compressed state sequence into the Transformer model, extract the local dynamic structure features and global evolution trend features, and output the state vector and stage division label of the killing process; S5. Input the state vector and the stage division label into the autoregressive controller, combine it with the preprocessed tea data, perform nonlinear control mapping, and output the control parameter adjustment amount; S6. Adjust the heating temperature, hot air dehumidification ratio, and stir-frying speed of the fixing equipment, and simultaneously record the adjusted equipment operating status and fixing process feedback data; S7. Combine the tea data and feedback data into label-aligned incremental training samples, and periodically input them into the convolutional neural network and Transformer model for parameter update.

2. The tea leaf fixing method based on deep learning according to claim 1, characterized in that: The tea data includes image data, temperature and humidity data, color and texture data, and odor signal data.

3. The tea leaf fixing method based on deep learning according to claim 1, characterized in that: The preprocessing includes filling missing data, removing outliers, normalizing values ​​and aligning timestamps.

4. The tea leaf fixing method based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31, expanding the cross-modal fusion feature representation in the time dimension to form a multi-channel time feature set consisting of a color change feature frame sequence, a leaf curl feature frame sequence, and a temperature and humidity change feature frame sequence; S32. Constructing an inter-frame weight matrix based on the multi-channel temporal feature set, and calculating the importance score of each frame based on the color change amplitude, curl morphology change rate, and temperature and humidity fluctuation trend; S33, extracting key completion change frames according to the importance score sorting, and eliminating invalid frames with scores lower than a set threshold to form a preliminary key frame sequence; S34, performing adjacent frame similarity detection on the preliminary key frame sequence, calculating the inter-frame distance based on color similarity, morphological proximity, and thermal and humidity state overlap, and merging redundant frames when the inter-frame distance is lower than a set threshold; S35, performing time window sliding sampling on the key frame sequence after de-redundancy, setting the window length and sliding step size, and forming a compressed time series with balanced distribution between frames; S36 , performing a temporal consistency check on the compressed time sequence, removing frames with discontinuous timestamps and skipped frames, and outputting a compressed state sequence with consistent temporal order.

5. The tea leaf fixing method based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Construct a Transformer model structure including a multi-layer encoder and decoder, wherein the encoder is composed of a position encoding module, a multi-head attention module, and a feedforward calculation module, and the decoder is composed of a context backpropagation module, a residual connection module, and a normalization module; S42, performing position encoding processing on the compressed state sequence, embedding the position information of each frame on the time axis into the feature sequence, and forming a vector sequence with a time mark; S43. Input the vector sequence with time stamp into the multi-head attention module, calculate the attention relationship between different attention heads, and extract the correlation information of color change structure, leaf edge disturbance structure and temperature and humidity mutation structure; S44, the output of the multi-head attention module is residually connected to the original input and input into the feedforward calculation module to perform nonlinear transformation to extract local dynamic structure features; S45. The encoder output is passed to the decoder structure to reversely model the cross-temporal variation characteristics between the states of each stage to form a global evolution trend feature covering the entire killing process; S46, combining the local dynamic structure features with the global evolution trend features, and inputting the combined features into a classification mapping unit, thereby outputting a state vector representing the state of the fixing process; S47. Perform time period boundary detection on the state vector, automatically divide the completion stage interval according to the feature change density, and output the stage division label.

6. The tea leaf fixing method based on deep learning according to claim 5, characterized in that: The stage division labels are generated based on an adaptive boundary detection algorithm. The boundary detection algorithm performs a multi-scale sliding window operation on the global evolution trend features output by the Transformer decoder, calculates the color change gradient mean, the leaf curl feature density change value and the temperature and humidity fluctuation stability index in each time period, and constructs a feature change function sequence. The first-order derivative turning point of the feature change function is used as a candidate segmentation point, and the stable boundary point set is screened through a density clustering method. Finally, the state vector sequence is divided into multiple stage intervals, and the stage division label corresponding to each stage is output.

7. The tea leaf fixing method based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, constructing an autoregressive controller structure, wherein the autoregressive controller is composed of a feature extraction unit, a stage label encoding unit, a historical action backtracking unit, and a nonlinear mapping unit; S52, inputting the state vector into a feature extraction unit, extracting the color offset, curling trend fluctuation value, and temperature and humidity gradient change value of the current fixing state, and constructing a continuous state feature sequence; S53, inputting the stage division labels into the stage label encoding unit, using a fixed encoding rule to convert each stage label into a stage control weight vector to form a stage identification control matrix; S54: Input the pre-processed tea data into the historical action backtracking unit, perform difference calculation based on the current state vector and the previous round of control parameters, and generate a historical difference vector for adjusting the target; S55, concatenating the state feature sequence, the stage identification control matrix, and the historical difference vector in the feature dimension, and inputting the concatenated concatenation into a nonlinear mapping unit, wherein the nonlinear mapping unit is composed of a multi-layer activation function; S56. In the nonlinear mapping unit, perform feature weighting calculation, normalization mapping and parameter projection operations in sequence to generate a fixing control parameter adjustment amount, which is used to adjust the operating instructions of the fixing device.

8. The tea leaf fixing method based on deep learning according to claim 7, characterized in that: The feature extraction unit performs continuous feature trajectory modeling on the state vector based on a multi-layer convolution structure, including color offset, curling trend fluctuation value and temperature and humidity gradient change value. The stage label encoding unit adopts a fixed-dimensional vector encoding scheme to map the stage division label into a stage control weight vector.

9. The tea leaf fixing method based on deep learning according to claim 1, characterized in that: The S6 specifically includes: S61. Drive the fixing device according to the control parameter adjustment amount, adjust the output power of the heating unit to the target temperature range, and continuously monitor the temperature change curve of the heating plate surface, recording the temperature response time and the steady-state maintenance time; S62, adjusting the fan speed and the opening and closing range of the dehumidification valve, setting the hot air output flow rate and the dehumidification time window, dynamically correcting the humidity and wind speed parameters in the hot air circulation channel, and recording the change sequence of the temperature and humidity coupling value at the hot air outlet; S63, comprehensively control the stir-frying speed, turning amplitude and stir-frying interval, and record the mechanical response delay and operation stability index of each cycle of the fixing equipment in real time; S64. Based on the heating temperature recording data, hot air dehumidification data and stir-frying speed data, a device operation status set is constructed, and a two-way recording data set is formed with the status feedback data corresponding to the current fixing stage, and stored in the local data cache unit.

10. The tea withering method based on deep learning according to claim 9, characterized in that: The equipment operation status set is derived from the heating temperature recording data, hot air dehumidification data and stir-frying speed data. The specific parameters include the heating temperature sequence, hot air flow trajectory, humidity adjustment parameters, wind speed change deviation, stir-frying speed curve, flipping amplitude range and stir-frying interval time, and are synchronously calibrated in chronological order to form a structured state set.

Images