Isolated low-temperature jasmine flower distillation system based on multi-modal perception and AI decision

Abstract

This invention relates to a jasmine flower isolated low-temperature scenting system based on multimodal perception and AI decision-making, belonging to the field of tea processing technology, including a sensing and monitoring module. The invention constructs a distributed sensor array to scan the reflected wavelengths of each grating, converting wavelength drift into temperature values ​​to achieve millimeter-level spatial resolution temperature distribution monitoring. This completely solves the problem of missed detection of local hot spots that traditional single-point temperature measurement cannot detect. It inversely reconstructs the three-dimensional transient temperature field data within the scenting chamber to generate directional airflow adjustment commands. Based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, it inputs them into a preset digital prediction model to predict the spatial temperature distribution within the scenting chamber in the future time period, generating airflow pre-adjustment commands. Through a comparative learning algorithm, it horizontally identifies the differences in scenting effects between different scenting chambers, extracts the optimal process features to form a strategy template library, and searches for and matches the optimal strategy for new batches based on raw material labels.

Description

Technical Field

[0001] This invention relates to the field of tea processing technology, and in particular to a jasmine flower isolated low-temperature scenting system based on multimodal perception and AI decision-making. Background Technology

[0002] Jasmine tea is a reprocessed tea made by scenting finely processed tea leaves with fresh jasmine flowers. It combines the mellow taste of tea with the fresh and fragrant aroma of jasmine. It accounts for more than 90% of the flower tea market in my country and has a history and cultural heritage of over a thousand years. The scenting process is the core step in forming the quality of jasmine tea. Its essence is to utilize the adsorption properties of tea leaves, so that the tea leaves can fully absorb the volatile aromatic compounds released by jasmine flowers through the mixing of tea and flowers. In recent years, scenting equipment using low-temperature environmental control technology has emerged. The temperature of the scenting chamber is maintained at a low level through a refrigeration system to inhibit the respiration and metabolism rate of the flowers and prolong the fragrance release time.

[0003] The existing technology for low-temperature scenting of jasmine tea has the following drawbacks: Existing equipment typically uses single-point or a small number of discretely arranged temperature sensors to monitor the temperature inside the chamber. However, due to factors such as heat generated by the respiration of the jasmine flower layer and the tea base layer, as well as uneven airflow distribution, there is a significant temperature gradient inside the chamber. Discrete single-point temperature measurement can only reflect local temperature information and cannot capture the true temperature field distribution inside the chamber. It is easy to miss local hot spots, resulting in local temperature runaway and aroma deterioration, while other areas have too low a temperature and insufficient aroma. Existing technologies lack a cross-box knowledge transfer mechanism, limiting the improvement of quality consistency under large-scale production conditions. Traditional scenting processes heavily rely on the personal experience of individual master craftsmen, and the process parameter settings for different operators, different scenting boxes, and different batches lack standardization and transferability. High-quality process parameters optimized for one box are difficult to directly replicate to other boxes due to equipment differences, resulting in reduced processing efficiency during scenting.

[0004] To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0005] The purpose of this invention is to significantly improve the consistency and stability of the molding quality and achieve generational evolution of the control strategy by deeply integrating four technologies: gas sensor self-calibration, fiber optic grating distributed temperature measurement, digital twin prediction, and multi-box comparative learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a jasmine flower isolated low-temperature curing system based on multimodal perception and AI decision-making, including a perception and monitoring module, a monitoring and analysis module, an AI decision-making module, a curing optimization module, and a control execution module; The sensing and monitoring module is used to monitor the concentration data of characteristic aroma components of jasmine. By analyzing the difference in the response spectrum of the sensor to different gases at different temperatures, it outputs drift-resistant continuous aroma concentration time-series data. The monitoring and analysis module constructs a distributed sensor array, scans the reflection wavelength of each grating, converts the wavelength drift into temperature values, inverts and reconstructs the three-dimensional transient temperature field data inside the enclosure, performs temperature difference analysis, and generates directional airflow adjustment commands. The AI ​​decision-making module, based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, inputs the data into a preset digital prediction model to predict the temperature spatial distribution inside the chamber in the future time period and generate airflow pre-adjustment instructions. The scenting optimization module uses a comparative learning algorithm to identify the differences in scenting effects of different scenting boxes, extracts the optimal process features to form a strategy template library, and searches for and matches the optimal strategy for new batches based on raw material labels. The control execution module feeds back the optimal strategy across the boxes to the AI ​​decision-making module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, thereby achieving the generational evolution of the control strategy.

[0007] Furthermore, by analyzing the differences in the response spectra of sensors to different gases at different temperatures, drift-resistant continuous aroma concentration time-series data are output. The specific process is as follows: Based on the deployment of a gas sensor array with integrated micro-hot plate technology in the flower layer and tea base layer of the scenting box, a database of response curves of each unit at multiple operating temperature points is established. During initialization, after clean air is introduced, the baseline response values ​​of each sensor in clean air are recorded and stored in non-volatile memory as a reference for subsequent calibration. After each scenting batch is completed, the baseline recovery program is automatically triggered. The sensor array is heated using a micro-hot plate to completely desorb the residual aroma molecules adsorbed on the sensor surface. After cooling to the set temperature, clean air is introduced again, and the baseline response value is measured again. Calculate the baseline drift between adjacent batches. The drift amount is subtracted from all original response values ​​in subsequent batches to achieve baseline zeroing.

[0008] Furthermore, a distributed sensor array is constructed to scan the reflected wavelengths of each grating, converting the wavelength drift into temperature values, and reconstructing the three-dimensional transient temperature field data inside the enclosure. The specific process is as follows: Single sensing optical fibers are laid out in a three-dimensional grid in the inner wall of the scenting box and in the flower layer and tea blank layer. Bragg gratings are engraved on the optical fibers at equal intervals to form a distributed temperature sensing network covering the entire box. Based on the initial reflection wavelength of the grating, it is compared with the factory calibration value to correct the small wavelength shift caused by installation stress and establish the initial wavelength reference library for each grating node. The demodulator scans the reflected wavelength of each grating at a set frequency, calculates the temperature of each flower layer and tea base layer based on the wavelength drift, and binds the spatial coordinates to form a discrete temperature point cloud. Spatial interpolation is performed on the discrete temperature point cloud to reconstruct the continuous temperature field of the fresh flower layer and tea base layer within the entire scenting box. The discrete point cloud is then expanded into a continuous three-dimensional temperature field using interpolation. Let the points to be interpolated be... Temperature estimate : In the above formula, N is the total number of known discrete temperature points used for interpolation. The weight coefficient corresponding to the iith known point. Let i be the measured temperature value of the iith known discrete temperature point, where i is the index of the temperature point. The chamber space is discretized to generate a voxel grid temperature distribution matrix, which is then stitched together according to time frames to form a transient temperature field sequence. The temperature difference between adjacent voxels is scanned point by point and marked. Hot topics and cold topics; Further, temperature difference analysis is performed to generate directional airflow adjustment commands. The specific process is as follows: Calculate the temperature difference between the flower layer and the tea base layer. If the temperature difference exceeds the temperature difference threshold, it is marked as an abnormal heat flow across layers. Track the temperature rise rate at a single point. When the temperature rise exceeds the temperature rise threshold, an early warning of abnormal heating is issued. Three-level response based on the severity of the anomaly: The value is D, and the interlayer temperature is C. First-level local fine-tuning; , ; Secondary regional regulation: , ; Level 3 full-field intervention: ≤; The instruction is mapped to the micro fan combination corresponding to the hotspot spatial coordinates.

[0009] Furthermore, based on the continuous aroma concentration time-series data and the three-dimensional transient temperature field data, the data are input into a preset digital prediction model to predict the spatial temperature distribution inside the chamber in the future time period and generate airflow pre-adjustment commands. The specific process is as follows: By combining continuous aroma concentration time-series data with three-dimensional transient temperature field data, key statistical features are extracted: aroma concentration time-series data, overall mean, hotspot maximum temperature, temperature difference between fresh flower layer and tea base layer, and temperature field standard deviation, forming a feature vector for each moment. The feature vector is input into the digital prediction model. Using the feature vector at the current moment as the initial condition, the state evolution within the future time window is solved, and the model outputs the forecast sequence for the future time period. Based on the warning level and the predicted location of hotspots, airflow adjustment instructions are generated in advance.

[0010] Furthermore, by using a comparative learning algorithm to identify the differences in the scenting effect of different scenting boxes, the optimal process features are extracted to form a strategy template library. For new batches, the optimal strategy is searched and matched according to the raw material label. The specific process is as follows: All historical scenting data from all scenting boxes were standardized in a unified format, including a comprehensive quality score that evaluates the finished tea from three dimensions: aroma concentration, freshness, and mellowness. Based on the comprehensive quality score, each batch was divided into excellent batches, ordinary batches, and poor batches. To train the comparative learning model, positive and negative sample pairs were included. Positive sample pairs consisted of two batches from different boxes at the same excellent grade, while negative sample pairs consisted of excellent batches and poor batches. A Siamese network framework is used as the contrastive learning algorithm. A pair of samples is input, and the encoder maps the two samples to the embedding space through shared weights to obtain the feature vector. The batches selected based on their overall quality score include parameters from the molding process, including... Process parameter time series data, actual perceived data trajectory during strategy execution, aroma concentration sequence and temperature field data, digital prediction model prediction trajectory of strategy and final product quality score, generate strategy templates for excellent batches, and store all strategy templates in strategy template library. When a new batch of scenting tasks begins, the raw material label and environmental conditions of the batch are input to construct an initial feature vector. The embedding vector is calculated by the encoder, and templates that match the cosine similarity are retrieved from the strategy template library and sorted in descending order of similarity.

[0011] Furthermore, the optimal strategy across the boxes is fed back into the AI ​​decision-making module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, thereby achieving intergenerational evolution of the control strategy. The specific process is as follows: From the optimal strategy template library, extract the complete parameter feature set of the currently selected optimal strategy, perform Bayesian incremental updates on parameters with high uncertainty in the digital prediction model, select the parameter set with the highest sensitivity but low calibration confidence in the model, and reconstruct the objective function virtually based on the execution data of multiple high-scoring strategies in the optimal strategy template library. Transform the learned reward function into the virtual inference objective function, replace or weight and fuse the original heuristic objective function, and deploy the updated twin model and the new objective function to the online version of the digital twin analysis module. In the next batch, execute the optimized strategy after back-injection.

[0012] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This jasmine flower isolation low-temperature scenting system, based on multimodal perception and AI decision-making, analyzes the difference in response spectra of different gases by sensors at different temperatures. It outputs drift-resistant, continuous time-series data of aroma concentration. With a self-calibrating sensor array, the sensor baseline automatically returns to its initial state, solving the problem of long-term sensor drift and enabling accurate identification of characteristic components in the gas. By constructing a distributed sensor array and scanning the reflected wavelengths of each grating, the system converts wavelength drift into temperature values, achieving millimeter-level spatial resolution temperature distribution monitoring. This completely solves the problem of missed detection of local hot spots that traditional single-point temperature measurement cannot detect. The system also inverts and reconstructs the three-dimensional transient temperature field data within the scenting chamber, performs temperature difference analysis, and generates directional airflow. The system provides adjustment commands to achieve second-level early warning and targeted control of hotspots. Based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, the data is input into a preset digital prediction model to predict the spatial temperature distribution within the scenting chamber in the future time period. This generates airflow pre-adjustment commands, which are initiated before the actual temperature exceeds the limit. Through comparative learning algorithms, the system horizontally identifies the differences in scenting effects among different scenting chambers, extracts the optimal process features to form a strategy template library, and retrieves and matches the optimal strategy for new batches based on raw material labels. The optimal strategy across chambers is then fed back into the AI ​​decision module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, achieving intergenerational evolution of control strategies and enabling cross-chamber knowledge accumulation and strategy self-optimization. Attached Figure Description

[0013] Figure 1 A schematic diagram of the overall system steps of the present invention is shown. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] Example 1: like Figure 1 As shown, the jasmine flower isolated low-temperature sealing system based on multimodal perception and AI decision-making is characterized by including a perception and monitoring module, a monitoring and analysis module, an AI decision-making module, a sealing optimization module, and a control execution module; The sensing and monitoring module is used to monitor the concentration data of characteristic aroma components of jasmine. By analyzing the difference in the response spectrum of the sensor to different gases at different temperatures, it outputs drift-resistant continuous aroma concentration time-series data. The monitoring and analysis module constructs a distributed sensor array, scans the reflection wavelength of each grating, converts the wavelength drift into temperature values, inverts and reconstructs the three-dimensional transient temperature field data inside the enclosure, performs temperature difference analysis, and generates directional airflow adjustment commands. The AI ​​decision-making module, based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, inputs the data into a preset digital prediction model to predict the temperature spatial distribution inside the chamber in the future time period and generate airflow pre-adjustment instructions. The scenting optimization module uses a comparative learning algorithm to identify the differences in scenting effects of different scenting boxes, extracts the optimal process features to form a strategy template library, and searches for and matches the optimal strategy for new batches based on raw material labels. The control execution module feeds back the optimal strategy across the boxes to the AI ​​decision-making module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, thereby achieving the generational evolution of the control strategy.

[0016] By analyzing the differences in the response spectra of sensors to different gases at different temperatures, drift-resistant continuous time-series data of aroma concentration are output. The specific process is as follows: Based on the deployment of a gas sensor array with integrated micro-hot plate technology in the flower layer and tea base layer of the scenting box, a database of response curves of each unit at multiple operating temperature points is established. During initialization, after clean air is introduced, the baseline response values ​​of each sensor in clean air are recorded and stored in non-volatile memory as a reference for subsequent calibration. After each scenting batch is completed, the baseline recovery program is automatically triggered. The sensor array is heated using a micro-hot plate to completely desorb the residual aroma molecules adsorbed on the sensor surface. After cooling to the set temperature, clean air is introduced again, and the baseline response value is measured again. Calculate the baseline drift between adjacent batches. The drift amount is subtracted from all original response values ​​in subsequent batches to achieve baseline zeroing.

[0017] A distributed sensor array is constructed, the reflected wavelengths of each grating are scanned, the wavelength drift is converted into temperature values, and the three-dimensional transient temperature field data inside the enclosure is reconstructed. The specific process is as follows: Single sensing optical fibers are laid out in a three-dimensional grid in the inner wall of the scenting box and in the flower layer and tea blank layer. Bragg gratings are engraved on the optical fibers at equal intervals to form a distributed temperature sensing network covering the entire box. Based on the initial reflection wavelength of the grating, it is compared with the factory calibration value to correct the small wavelength shift caused by installation stress and establish the initial wavelength reference library for each grating node. The demodulator scans the reflected wavelength of each grating at a set frequency, calculates the temperature of each flower layer and tea base layer based on the wavelength drift, and binds the spatial coordinates to form a discrete temperature point cloud. Spatial interpolation is performed on the discrete temperature point cloud to reconstruct the continuous temperature field of the fresh flower layer and tea base layer within the entire scenting box. The discrete point cloud is then expanded into a continuous three-dimensional temperature field using interpolation. Let the points to be interpolated be... Temperature estimate : In the above formula, N is the total number of known discrete temperature points used for interpolation. The weight coefficient corresponding to the iith known point. Let i be the measured temperature value of the iith known discrete temperature point, where i is the index of the temperature point. The chamber space is discretized to generate a voxel grid temperature distribution matrix, which is then stitched together according to time frames to form a transient temperature field sequence. The temperature difference between adjacent voxels is scanned point by point and marked. Hot topics and cold topics; Temperature difference analysis is performed to generate directional airflow control commands. The specific process is as follows: Calculate the temperature difference between the flower layer and the tea base layer. If the temperature difference exceeds the temperature difference threshold, it is marked as an abnormal heat flow across layers. Track the temperature rise rate at a single point. When the temperature rise exceeds the temperature rise threshold, an early warning of abnormal heating is issued. Three-level response based on the severity of the anomaly: The value is D, and the interlayer temperature is C. First-level local fine-tuning; , ; Secondary regional regulation: , ; Level 3 full-field intervention: ≤; The instruction is mapped to the micro fan combination corresponding to the hotspot spatial coordinates.

[0018] Based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, the data are input into a preset digital prediction model to predict the spatial temperature distribution inside the chamber over future time periods and generate airflow pre-adjustment commands. The specific process is as follows: By combining continuous aroma concentration time-series data with three-dimensional transient temperature field data, key statistical features are extracted: aroma concentration time-series data, overall mean, hotspot maximum temperature, temperature difference between fresh flower layer and tea base layer, and temperature field standard deviation, forming a feature vector for each moment. The feature vector is input into the digital prediction model. Using the feature vector at the current moment as the initial condition, the state evolution within the future time window is solved, and the model outputs the forecast sequence for the future time period. Based on the warning level and the predicted location of hotspots, airflow adjustment instructions are generated in advance.

[0019] The digital prediction model employs an encoder-predictor-decoder architecture. The encoder consists of a two-layer convolutional neural network and a one-layer LSTM, with an input window of 30 minutes and an output embedding vector dimension of 64. The predictor consists of a two-layer LSTM, receiving the embedding vector and outputting the dimensionality-reduced temperature field features and aroma concentration sequence for the next 30 minutes. The decoder uses a two-layer deconvolutional neural network to recover the temperature field distribution from the predicted features. Using historical running data as the training set, with mean squared error as the loss function, Adam optimizer, learning rate 0.001, training for 50 epochs; during online operation, during the low load period in the early morning of each day, incremental fine-tuning of the model with data from the most recent 7 days, learning rate 0.0001, training for 10 epochs. When it is predicted that the temperature of any voxel inside the chamber will exceed a set threshold within the next 15 minutes with a confidence level >70%, a pre-adjustment command is generated. The strength of the pre-adjustment command is proportional to the difference between the predicted hot spot temperature and the threshold. The command is sent to the edge controller, which superimposes the pre-adjustment command and the actual temperature difference adjustment command within its execution cycle, and takes the larger value as the final control quantity.

[0020] By using a comparative learning algorithm to identify the differences in the scenting effect of different scenting boxes, the optimal process features are extracted to form a strategy template library. For new batches, the optimal strategy is searched and matched according to the raw material label. The specific process is as follows: All historical scenting data from all scenting boxes were standardized in a unified format, including a comprehensive quality score that evaluates the finished tea from three dimensions: aroma concentration, freshness, and mellowness. Based on the comprehensive quality score, each batch was divided into excellent batches, ordinary batches, and poor batches. To train the comparative learning model, positive and negative sample pairs were included. Positive sample pairs consisted of two batches from different boxes at the same excellent grade, while negative sample pairs consisted of excellent batches and poor batches. A Siamese network framework is used as the contrastive learning algorithm. A pair of samples is input, and the encoder maps the two samples to the embedding space through shared weights to obtain the feature vector. The batches selected based on their overall quality score include parameters from the molding process, including... Process parameter time series data, actual perceived data trajectory during strategy execution, aroma concentration sequence and temperature field data, digital prediction model prediction trajectory of strategy and final product quality score, generate strategy templates for excellent batches, and store all strategy templates in strategy template library. When a new batch of scenting tasks begins, the raw material label and environmental conditions of the batch are input to construct an initial feature vector. The embedding vector is calculated by the encoder, and templates that match the cosine similarity are retrieved from the strategy template library and sorted in descending order of similarity.

[0021] The optimal strategy across the boxes is fed back into the AI ​​decision-making module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, thereby achieving the intergenerational evolution of the control strategy. The specific process is as follows: From the optimal strategy template library, extract the complete parameter feature set of the currently selected optimal strategy, perform Bayesian incremental updates on parameters with high uncertainty in the digital prediction model, select the parameter set with the highest sensitivity but low calibration confidence in the model, and reconstruct the objective function virtually based on the execution data of multiple high-scoring strategies in the optimal strategy template library. Transform the learned reward function into the virtual inference objective function, replace or weight and fuse the original heuristic objective function, and deploy the updated twin model and the new objective function to the online version of the digital twin analysis module. In the next batch, execute the optimized strategy after back-injection.

[0022] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0023] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. In the two embodiments provided in this application, it should be understood that the disclosed apparatus and system can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, and the division of modules is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; furthermore, the coupling or direct coupling or communication connection between the shown or discussed mutuals can be through some interfaces, and the indirect coupling or communication connection between the apparatus or modules can be electrical, mechanical or other forms. The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A jasmine flower isolation low-temperature curing system based on multimodal perception and AI decision-making, characterized in that, It includes a perception and monitoring module, a monitoring and analysis module, an AI decision-making module, a control optimization module, and a control execution module; The sensing and monitoring module is used to monitor the concentration data of characteristic aroma components of jasmine. By analyzing the difference in the response spectrum of the sensor to different gases at different temperatures, it outputs drift-resistant continuous aroma concentration time-series data. The monitoring and analysis module constructs a distributed sensor array, scans the reflection wavelength of each grating, converts the wavelength drift into temperature values, inverts and reconstructs the three-dimensional transient temperature field data inside the enclosure, performs temperature difference analysis, and generates directional airflow adjustment commands. The AI ​​decision-making module, based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, inputs the data into a preset digital prediction model to predict the temperature spatial distribution inside the chamber in the future time period and generate airflow pre-adjustment instructions. The scenting optimization module uses a comparative learning algorithm to identify the differences in scenting effects of different scenting boxes, extracts the optimal process features to form a strategy template library, and searches for and matches the optimal strategy for new batches based on raw material labels. The control execution module feeds back the optimal strategy across the boxes to the AI ​​decision-making module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, thereby achieving the generational evolution of the control strategy.

2. The jasmine flower isolated low-temperature curing system based on multimodal perception and AI decision-making according to claim 1, characterized in that, By analyzing the differences in the response spectra of sensors to different gases at different temperatures, drift-resistant continuous time-series data of aroma concentration are output. The specific process is as follows: Based on the deployment of a gas sensor array with integrated micro-hot plate technology in the flower layer and tea base layer of the scenting box, a database of response curves of each unit at multiple operating temperature points is established. During initialization, after clean air is introduced, the baseline response values ​​of each sensor in clean air are recorded and stored in non-volatile memory as a reference for subsequent calibration. After each scenting batch is completed, the baseline recovery program is automatically triggered. The sensor array is heated using a micro-hot plate to completely desorb the residual aroma molecules adsorbed on the sensor surface. After cooling to the set temperature, clean air is introduced again, and the baseline response value is measured again. Calculate the baseline drift between adjacent batches. The baseline drift is subtracted from all original response values ​​in subsequent batches to achieve baseline zeroing.

3. The jasmine flower isolation low-temperature curing system based on multimodal perception and AI decision-making according to claim 1, characterized in that, A distributed sensor array is constructed, the reflected wavelengths of each grating are scanned, the wavelength drift is converted into temperature values, and the three-dimensional transient temperature field data inside the enclosure is reconstructed. The specific process is as follows: Single sensing optical fibers are laid out in a three-dimensional grid in the inner wall of the scenting box and in the flower layer and tea blank layer. Bragg gratings are engraved on the optical fibers at equal intervals to form a distributed temperature sensing network covering the entire box. Based on the initial reflection wavelength of the grating, it is compared with the factory calibration value to correct the small wavelength shift caused by installation stress and establish the initial wavelength reference library for each grating node. The demodulator scans the reflected wavelength of each grating at a set frequency, calculates the temperature of each flower layer and tea base layer based on the wavelength drift, and binds the spatial coordinates to form a discrete temperature point cloud. Spatial interpolation is performed on the discrete temperature point cloud to reconstruct the continuous temperature field of the fresh flower layer and tea base layer within the entire scenting box. The discrete point cloud is then expanded into a continuous three-dimensional temperature field using interpolation. Let the points to be interpolated be... Temperature estimate : In the above formula, The total number of known discrete temperature points used for interpolation. The weight coefficient corresponding to the iith known point. Let i be the measured temperature value of the iith known discrete temperature point, where i is the index of the temperature point. The chamber space is discretized to generate a voxel grid temperature distribution matrix, which is then stitched together according to time frames to form a transient temperature field sequence. The temperature difference between adjacent voxels is scanned point by point and marked. Hot topics and cold topics.

4. The jasmine flower isolation low-temperature curing system based on multimodal perception and AI decision-making according to claim 1, characterized in that, Temperature difference analysis is performed to generate directional airflow control commands. The specific process is as follows: Calculate the temperature difference between the flower layer and the tea base layer. If the temperature difference exceeds the temperature difference threshold, it is marked as an abnormal heat flow across layers. Track the temperature rise rate at a single point. When the temperature rise exceeds the temperature rise threshold, an early warning of abnormal heating is issued. Three-level response based on the severity of the anomaly: The value is D, and the interlayer temperature is C. First-level local fine-tuning; , ; Secondary regional regulation: , ; Level 3 full-field intervention: ≤; The instruction is mapped to the micro fan combination corresponding to the hotspot spatial coordinates.

5. The jasmine flower isolation low-temperature curing system based on multimodal perception and AI decision-making according to claim 1, characterized in that, Based on continuous aroma concentration time-series data and three-dimensional transient temperature field data, the data are input into a preset digital prediction model to predict the spatial temperature distribution inside the chamber over future time periods and generate airflow pre-adjustment commands. The specific process is as follows: By combining continuous aroma concentration time-series data with three-dimensional transient temperature field data, key statistical features are extracted: aroma concentration time-series data, overall mean, hotspot maximum temperature, temperature difference between fresh flower layer and tea base layer, and temperature field standard deviation, forming a feature vector for each moment. The feature vector is input into the digital prediction model. Using the feature vector at the current moment as the initial condition, the state evolution within the future time window is solved, and the model outputs the forecast sequence for the future time period. Based on the warning level and the predicted location of hotspots, airflow adjustment instructions are generated in advance.

6. The jasmine flower isolation low-temperature curing system based on multimodal perception and AI decision-making according to claim 1, characterized in that, By using a comparative learning algorithm to identify the differences in the scenting effect of different scenting boxes, the optimal process features are extracted to form a strategy template library. For new batches, the optimal strategy is searched and matched according to the raw material label. The specific process is as follows: All historical scenting data from all scenting boxes were standardized in a unified format, including a comprehensive quality score that evaluates the finished tea from three dimensions: aroma concentration, freshness, and mellowness. Based on the comprehensive quality score, each batch was divided into excellent batches, ordinary batches, and poor batches. To train the comparative learning model, positive and negative sample pairs were included. Positive sample pairs consisted of two batches from different boxes at the same excellent grade, while negative sample pairs consisted of excellent batches and poor batches. A Siamese network framework is used as the contrastive learning algorithm. A pair of samples is input, and the encoder maps the two samples to the embedding space through shared weights to obtain the feature vector. The batches selected based on their overall quality score include parameters from the molding process, including... Process parameter time series data, actual perceived data trajectory during strategy execution, aroma concentration sequence and temperature field data, digital prediction model prediction trajectory of strategy and final product quality score, generate strategy templates for excellent batches, and store all strategy templates in strategy template library. When a new batch of scenting tasks begins, the raw material label and environmental conditions of the batch are input to construct an initial feature vector. The embedding vector is calculated by the encoder, and templates that match the cosine similarity are retrieved from the strategy template library and sorted in descending order of similarity.

7. The jasmine flower isolation low-temperature curing system based on multimodal perception and AI decision-making according to claim 1, characterized in that, The optimal strategy across the boxes is fed back into the AI ​​decision-making module to correct the parameters of the digital twin model and optimize the objective function of the virtual simulation, thereby achieving the intergenerational evolution of the control strategy. The specific process is as follows: From the optimal strategy template library, extract the complete parameter feature set of the currently selected optimal strategy, perform Bayesian incremental updates on parameters with high uncertainty in the digital prediction model, select the parameter set with the highest sensitivity but low calibration confidence in the model, and reconstruct the objective function virtually based on the execution data of multiple high-scoring strategies in the optimal strategy template library. Transform the learned reward function into the virtual inference objective function, replace or weight and fuse the original heuristic objective function, and deploy the updated twin model and the new objective function to the online version of the digital twin analysis module. In the next batch, execute the optimized strategy after back-injection.

Images