Real-time control method and system for tea fixation machine

Abstract

The invention provides a real-time control method and system for a tea leaf fixation machine, and the method comprises the steps: obtaining the flowing speed and direction data of tea leaves in a fixation cylinder through a flow sensor array, and generating a real-time flow distribution image in combination with a data fusion algorithm; according to the real-time flow distribution image, a convolutional neural network is adopted to extract time sequence features of the tea leaf accumulation area and the non-uniform flow area, and accumulation distribution feature vectors are obtained; according to the stable flow distribution data, obtaining time sequence data of multi-point temperature in the fixation cylinder through an infrared thermometer, and generating a temperature distribution matrix by combining with time sequence data smoothing processing; according to the temperature distribution matrix, an anomaly detection algorithm is adopted to judge whether an area with local temperature exceeding a dynamic threshold value set exists or not, and coordinates of the temperature anomaly area are obtained; and according to the new process parameter combination, controlling the fixation equipment to execute stirring and feeding operations through a feedback control system to obtain uniform heating data. The tea leaf heating uniformity is obviously improved, the local overheating risk is reduced, and the fixation quality and efficiency are optimized.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a real-time control method and system for a tea fixing machine. Background Technology

[0002] Tea processing is a crucial sector in agriculture and the food industry. Its core lies in preserving the color, aroma, and taste of tea through precise process control, ensuring consistent quality and enhancing market value. The initial fixation (or "kill-green") step, a critical process, directly impacts the physicochemical properties and final quality of tea. However, existing fixation equipment largely relies on manual experience or simple automation, making it difficult to handle the dynamic differences in tea caused by variations in variety, moisture content, and environment during processing. This leads to delayed adjustments in processing parameters, uneven heating, or over-fixing, affecting quality consistency. Furthermore, existing methods often neglect the combination of real-time monitoring and dynamic adjustment when addressing tea flowability and distribution uniformity, easily causing tea accumulation or localized overheating, reducing processing efficiency.

[0003] During the withering process, the flow state and distribution density of tea leaves are core technical attributes affecting quality stability. Uneven distribution of tea leaves within the withering drum can lead to some leaves being underheated or overheated, thus affecting the formation of color and aroma. For example, during the high-temperature withering stage, if tea leaves accumulate at the bottom of the drum, localized excessive heat may cause scorching, while other areas may retain too much moisture due to insufficient heating, affecting subsequent shaping. The instability of the flow state further exacerbates the difficulty of controlling process parameters, as changes in tea leaf distribution density directly affect the working effect of the agitator and the efficiency of temperature conduction. This makes real-time monitoring and precise control of the tea leaf flow state in a dynamic processing environment a pressing technical challenge that needs to be addressed.

[0004] Therefore, how to monitor the flow state and distribution density of tea leaves in real time, and adaptively adjust the stirring parameters and feeding speed based on these dynamic data to achieve uniform heating of tea leaves and smooth transition of process parameters during the fixation process, has become a key issue in improving the quality and efficiency of tea processing. Summary of the Invention

[0005] In a first aspect, the present invention provides a real-time control method for a tea fixing machine, mainly comprising: The flow velocity and direction data of tea leaves inside the fixing drum are acquired through a flow sensor array, and a real-time flow distribution image is generated using a data fusion algorithm. Based on the real-time flow distribution image, a convolutional neural network is used to extract the temporal features of tea leaf accumulation areas and uneven flow areas, resulting in an accumulation distribution feature vector. If the proportion of accumulation areas in the accumulation distribution feature vector exceeds a preset threshold, a deep reinforcement learning algorithm is used to optimize the stirring speed and angle adjustment mechanism based on the real-time flow distribution image and temporal features, resulting in optimized stirring parameters. Based on the optimized stirring parameters and the real-time flow distribution image, a fuzzy control algorithm is used with flow distribution uniformity as input to dynamically adjust the feeding speed, obtaining stable flow distribution data. Based on the stable flow distribution data, a multi-point temperature temporal data is acquired using an infrared thermometer inside the fixing drum. A temperature distribution matrix is ​​generated by smoothing time-series data. Based on the temperature distribution matrix, an anomaly detection algorithm is used to determine whether there are areas where the local temperature exceeds the dynamic threshold, and the coordinates of the temperature anomaly areas are obtained. If the coordinates of the temperature anomaly areas are not empty, the stirring speed, angle adjustment mechanism, and feeding speed are adjusted by a control parameter optimization algorithm based on the real-time flow distribution image, the accumulation distribution feature vector, and the coordinates of the temperature anomaly areas to obtain a new combination of process parameters. Based on the new combination of process parameters, the blanching equipment is controlled to perform stirring and feeding operations through a feedback control system to obtain uniform heating data. The real-time flow distribution image and temperature distribution matrix are continuously acquired through a flow sensor array and an infrared thermometer, and the above steps are repeated in a loop with real-time sequence updates to obtain the final stable flow distribution and uniform heating data.

[0006] On the other hand, this invention provides a real-time control system for a tea processing machine, mainly comprising: a flow data acquisition module, used to acquire the flow speed and direction data of tea leaves in the processing cylinder through a flow sensor array, and generate a real-time flow distribution image by combining a data fusion algorithm; a feature extraction module, used to extract the temporal features of the tea leaf accumulation area and the uneven flow area based on the real-time flow distribution image using a convolutional neural network, to obtain an accumulation distribution feature vector; a stirring parameter optimization module, used to optimize the stirring speed and angle adjustment mechanism based on the real-time flow distribution image and temporal features by using a deep reinforcement learning algorithm if the proportion of the accumulation area in the accumulation distribution feature vector exceeds a preset threshold, to obtain optimized stirring parameters; a flow regulation module, used to dynamically adjust the feeding speed based on the optimized stirring parameters and the real-time flow distribution image using a fuzzy control algorithm with the flow distribution uniformity as input, to obtain stable flow distribution data; and a temperature data acquisition module, used to adjust the temperature based on the stable flow distribution data by... An infrared thermometer acquires time-series temperature data at multiple points within the blanching cylinder, and combines this data with smoothing to generate a temperature distribution matrix. An anomaly detection module uses an anomaly detection algorithm based on the temperature distribution matrix to determine if there are areas where the local temperature exceeds a dynamically set threshold, obtaining the coordinates of these anomaly areas. A process parameter optimization module, if the coordinates of the anomaly areas are not empty, adjusts the stirring speed, angle adjustment mechanism, and feeding speed using a control parameter optimization algorithm based on the real-time flow distribution image, accumulation distribution feature vector, and the coordinates of the anomaly areas, obtaining a new combination of process parameters. An execution control module controls the blanching equipment to perform stirring and feeding operations through a feedback control system based on the new combination of process parameters, obtaining uniform heating data. A cycle monitoring module continuously acquires real-time flow distribution images and temperature distribution matrices through a flow sensor array and an infrared thermometer, and cyclically executes the above steps in conjunction with real-time sequence updates to obtain the final stable flow distribution and uniform heating data.

[0007] In summary, this invention discloses a real-time control method and system for a tea fixing machine. Addressing the technical challenges of uneven tea flow and localized overheating caused by tea accumulation during traditional fixing processes, the method collects real-time data on the flow velocity, direction, and temperature of the tea leaves within the fixing drum. This data is then integrated with a flow sensor array and an infrared thermometer to generate a flow distribution image and a temperature distribution matrix. A convolutional neural network is used to extract the temporal features of accumulation and uneven flow. Combined with deep reinforcement learning and fuzzy control algorithms, the stirring speed, angle, and feeding speed are dynamically optimized to effectively control the uniformity of flow distribution. If a localized temperature anomaly is detected, a control parameter optimization algorithm integrates the flow, accumulation, and temperature data to generate a new combination of process parameters, driving the feedback control system to execute precise operations. This invention significantly improves the uniformity of tea leaf heating, reduces the risk of localized overheating, and optimizes the quality and efficiency of fixing. Attached Figure Description

[0008] Figure 1 This is a flowchart of a real-time control method for a tea fixing machine according to the present invention.

[0009] Figure 2 This is a schematic diagram of a real-time control method for a tea fixing machine according to the present invention.

[0010] Figure 3 This is another schematic diagram of a real-time control method for a tea fixing machine according to the present invention.

[0011] Figure 4 This is a schematic diagram of the structure of a real-time control system for a tea fixing machine according to the present invention.

[0012] Figure 5 This is a schematic diagram showing the state comparison inside the tea fixing cylinder in a real-time control system for a tea fixing machine according to the present invention.

[0013] Figure 6 This is a quantitative comparison chart showing the real-time control method for a tea fixing machine according to the present invention and the prior art. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0015] Example 1 like Figure 1-3 This embodiment of a real-time control method and system for a tea fixing machine may specifically include: Step S101: Obtain the flow speed and direction data of tea leaves in the fixing cylinder through the flow sensor array, and generate a real-time flow distribution image by combining the data fusion algorithm.

[0016] like Figure 1As shown, flow velocity and direction data of tea leaves inside the processing drum are collected using a flow sensor array. A Kalman filter algorithm is used to fuse the data, generating a first flow distribution data set. If the variance of the first flow distribution data set is greater than a preset threshold, the velocity and direction data are denoised to obtain a second flow distribution data set. Based on the second flow distribution data set, a mean filter algorithm is used to smooth the data, generating a third flow distribution data set. Using the third flow distribution data set, a two-dimensional convolution algorithm is used to generate an initial flow distribution image. If the edge sharpness of the initial flow distribution image is lower than a preset threshold, edge enhancement processing is performed on the initial flow distribution image to generate a first flow distribution image. Based on the first flow distribution image, a color mapping method is used to generate a real-time flow distribution image. Using the real-time flow distribution image, key feature points of tea leaf flow are extracted to generate the final flow distribution image.

[0017] For example, flow velocity and direction data of tea leaves inside the processing drum are acquired using a flow sensor array. First, a set of high-precision ultrasonic flow sensors is deployed, evenly distributed along the inner wall of the drum. Assuming a drum diameter of 0.8 meters, eight sensors are installed at 45-degree intervals. Each sensor collects flow velocity data at a frequency of 100Hz, with a range of 0-2 m / s and an accuracy of ±0.01 m / s. Simultaneously, the flow direction angle is recorded, ranging from 0-360 degrees. The sensors measure the velocity of tea particles under the influence of hot air using a time-difference method. For example, if a sensor measures a flow velocity of 1.25 m / s and a direction of 30 degrees, the data is transmitted in real-time to the central processing unit via a wireless module. A real-time flow distribution image is generated using a data fusion algorithm, and a Kalman filter algorithm is employed to fuse the sensor data.

[0018] For example, after fusion, the flow velocity at a certain point is corrected to 1.23 m / s, with a direction of 31.5 degrees. Next, an interpolation algorithm (such as inverse distance weighting, with a weighting factor n=2) is used to convert the discrete point data into a continuous flow field, with a grid resolution of 0.05 m × 0.05 m, generating a two-dimensional velocity vector field. Finally, a visualization algorithm maps the vector field to a real-time flow distribution image, using a heatmap to represent the velocity magnitude (color scale range 0-2 m / s, red for high, green for low), and overlaying arrows to indicate direction, with a refresh rate of 10 frames / second. The analysis revealed that the flow velocity is higher at the center of the cylinder (approximately 1.5 m / s) and lower at the edges (approximately 0.8 m / s), indicating uneven hot air distribution, which allows for optimization of the duct design. This process is automatically run through an embedded system to ensure real-time performance and accuracy, with data stored in the cloud to support subsequent process optimization.

[0019] Step S102: Based on the real-time flow distribution image, a convolutional neural network is used to extract the temporal features of the tea accumulation area and the uneven flow area to obtain the accumulation distribution feature vector.

[0020] like Figure 1As shown, initial image data is acquired from real-time traffic distribution images. Preprocessing operations are performed to denoise and standardize the images, yielding the first image data. A convolutional neural network is used to extract features from the first image data, performing region segmentation on tea-accumulation areas and uneven flow areas, resulting in region segmentation results. If the pixel proportion of tea-accumulation areas in the region segmentation results exceeds a preset threshold, temporal features are extracted from these areas to generate a first temporal feature set; if it is below the preset threshold, it is marked as an uneven flow area, generating a second temporal feature set. A spatiotemporal feature fusion algorithm is used to integrate the first and second temporal feature sets to obtain a fused feature set. Based on the fused feature set, principal component analysis is used to reduce the dimensionality of the features, generating a dimensionality-reduced feature vector. Dynamic distribution analysis is performed on the dimensionality-reduced feature vector to detect changes in traffic patterns, obtaining an accumulation distribution feature vector. Key distribution parameters are extracted from the accumulation distribution feature vector to generate the final feature vector representation.

[0021] For example, real-time images of the flow distribution in a tea processing workshop are acquired. Assuming the images are in RGB format with a resolution of 1920×1080 and a frame rate of 30 frames per second, a convolutional neural network (CNN) is used for feature extraction. First, the images are preprocessed, using Gaussian blur (kernel size 5×5, standard deviation 1.0) to remove noise and converting the images to grayscale to reduce computational complexity. Then, the Otsu thresholding method (adaptive thresholding, assuming that after segmentation, the pixel value of the tea-stacking area is >150 and the non-stacking area is <100) is used to separate the tea-stacking area from the background, generating a binary mask to mark the stacking area and the uneven flow area (defined as areas with a flow change rate >20%). Next, a CNN model is designed, containing three convolutional layers (3×3 kernels, with 32, 64, and 128 channels respectively, stride 1, ReLU activation), followed by two fully connected layers (512 and 128 neurons), outputting a 32-dimensional temporal feature vector. To extract temporal features, 10 consecutive frames (approximately 0.33 seconds) of images are input into a CNN. Convolutional layers capture spatial features (such as the edges and density of stacking regions), pooling layers (2×2 max pooling) downsample to retain key features, and fully connected layers integrate temporal information to generate feature vectors. To analyze the stacking distribution, the mean and variance of each dimension in the feature vector are calculated (assuming a mean range of [0.1, 0.9], and a variance <0.05 for stable stacking). K-means clustering (K=3, initial centers random) is used to divide the feature vector into high, medium, and low stacking regions, resulting in a distribution feature vector. The analysis process incorporates business considerations, assuming that high stacking regions (mean >0.7) require adjustment of the conveyor belt speed (reduced by 10%), and low stacking regions (mean <0.3) require an increase in the material supply (increased by 15%), achieving dynamic control. The final output is a 32-dimensional vector representing the temporal characteristics of stacking distribution and uneven flow, providing insights for subsequent optimization decisions.

[0022] Step S103: If the proportion of the accumulation area in the accumulation distribution feature vector exceeds a preset threshold, then the stirring speed and angle adjustment mechanism are optimized based on the real-time flow distribution image and time-series features using a deep reinforcement learning algorithm to obtain optimized stirring parameters.

[0023] Real-time flow distribution images and time-series feature sequences are acquired from sensors to generate a stacking distribution feature vector. If the region proportion of the stacking distribution feature vector exceeds a preset threshold, the real-time flow distribution image is denoised and segmented using image processing algorithms to obtain a processed flow distribution image. Based on the processed flow distribution image, a convolutional neural network is used to extract spatial features, which are then combined with the time-series feature sequence to generate a comprehensive feature vector. Using a deep reinforcement learning algorithm, the stirring speed and angle adjustment parameters are optimized based on the comprehensive feature vector to obtain preliminary optimized control parameters. If the simulation results of the preliminary optimized control parameters deviate from the target control parameters beyond a preset range, the weights of the deep reinforcement learning model are adjusted using a gradient descent algorithm to obtain updated optimized control parameters. Based on the updated optimized control parameters, control commands for the stirring equipment are generated and output to the actuator. The time-series feature sequence is updated using real-time operating data fed back from the actuator, generating a new stacking distribution feature vector.

[0024] For example, firstly, image processing techniques are used to analyze the feature vector of the accumulation distribution. Assuming an image of the accumulation area is acquired from a real-time monitoring camera, a convolutional neural network (CNN) is used to extract features and obtain the proportion of the accumulation area. A preset threshold of 70% is set; if the accumulation area is detected to occupy 75% of the entire container cross-section, the optimization process is triggered. Image processing uses the OpenCV library; after grayscale conversion, the accumulation area is segmented using an edge detection algorithm (such as Canny). The pixel area proportion is calculated to be 75 / 100 = 0.75, exceeding the threshold of 0.7. Next, based on the real-time flow distribution image, fluid dynamics simulation software (such as ANSYS Fluent) is used to generate flow distribution features, extracting velocity and pressure field data. Assuming the flow distribution shows a flow velocity of 0.5 m / s in the central area and 0.2 m / s at the edges, it indicates uneven mixing. Subsequently, combining time-series characteristics, stirring speed data (sampled per minute, mean 300 rpm, standard deviation 20 rpm) and angle data (mean 45°, standard deviation 5°) from the past 10 minutes were collected. A Long Short-Term Memory (LSTM) network was used to predict the flow trend for the next 5 minutes, and the prediction results showed that the flow rate deviation would increase by 10%. Finally, stirring parameters were optimized using a Deep Reinforcement Learning (DRL) algorithm (such as DDPG), with the reward function set as flow rate uniformity (target deviation <5%) and energy consumption (<500W). After 1000 iterations of the DDPG algorithm, the optimized parameters were output: stirring speed adjusted to 350 rpm, angle adjusted to 50°, and simulation verification showed that the flow rate deviation decreased to 4.8% and energy consumption was 480W. The optimized parameters were automatically sent to the stirring equipment via the PLC controller, completing closed-loop control.

[0025] Step S104: Based on the optimized stirring parameters and real-time flow distribution image, a fuzzy control algorithm is used with the flow distribution uniformity as input to dynamically adjust the feed rate and obtain stable flow distribution data.

[0026] like Figure 2As shown, real-time flow distribution data is acquired by collecting flow distribution information within the pipeline through sensors. Noise is removed using data preprocessing methods to obtain a first flow distribution dataset. Based on this dataset, a flow uniformity index is calculated. If the uniformity index is lower than a preset threshold, a fuzzy control algorithm is triggered to obtain an adjustment signal. The fuzzy control algorithm processes the adjustment signal, calculating the feed rate adjustment based on a predefined fuzzy rule base to obtain a first speed adjustment parameter. Using this first speed adjustment parameter, the operating speed of the feeding device is dynamically adjusted to update the flow distribution within the pipeline, resulting in a second flow distribution dataset. For the second flow distribution dataset, the flow uniformity index is recalculated. If the uniformity index is still lower than the preset threshold, the fuzzy rule base is iteratively adjusted to obtain optimized rule parameters. Based on the optimized rule parameters, the fuzzy control algorithm is updated, and the feed rate adjustment is recalculated to obtain a second speed adjustment parameter. Using this second speed adjustment parameter, the feed rate is continuously adjusted, and the flow distribution within the pipeline is monitored to obtain stable flow distribution data.

[0027] For example, based on optimized stirring parameters and real-time flow distribution images, a fuzzy control algorithm is used to dynamically adjust the feed rate to achieve uniform flow distribution. First, flow distribution data within the stirring tank is collected in real-time using sensors, such as an ultrasonic flow meter collecting 100 sets of flow data per second, recording the flow velocity at 10 key points in m / s. Assuming the initial flow velocity distribution is [0.5, 0.7, 0.6, 0.8, 0.9, 0.4, 0.5, 0.6, 0.7, 0.8], the uniformity index is calculated as the standard deviation σ = 0.15, and the goal is to reduce σ to below 0.05. Next, a fuzzy control algorithm is constructed. The inputs are the flow uniformity error E (the difference between the target uniformity and the actual uniformity, E = 0.05 - 0.15 = -0.1) and the error change rate EC (the change in uniformity per second, e.g., EC = 0.02). The output is the feed rate adjustment ΔV (in m³ / s). 3 / h). The fuzzy controller uses triangular membership functions to divide E and EC into five fuzzy sets: {negative large, negative small, zero, positive small, positive large}. The output ΔV is divided into {-0.5, -0.2, 0, 0.2, 0.5}. According to fuzzy rules (such as "if E is negative large and EC is positive small, then ΔV is negative small"), fuzziness is resolved through Mamdani inference and the centroid method, yielding ΔV = -0.2m. 3The feed pump frequency was adjusted from 50Hz to 48Hz per hour. Real-time flow distribution images were generated by the data visualization module, plotting a flow velocity heatmap based on the collected data at a refresh rate of 1Hz. Analysis showed that the flow velocity distribution tended to be uniform, with the adjusted flow velocity being [0.65, 0.67, 0.66, 0.68, 0.67, 0.64, 0.65, 0.66, 0.67, 0.66], and σ decreasing to 0.01. The system continuously optimized through a feedback loop, recalculating E and EC every 5 seconds and dynamically adjusting ΔV until σ stabilized below 0.05. The entire process was automated via a PLC controller and SCADA system, with data stored in a cloud database for subsequent analysis of the correlation between stirring parameters and flow distribution, ensuring process stability.

[0028] Step S105: Based on the stable flow distribution data, time-series data of multiple temperatures inside the blanching cylinder are obtained using an infrared thermometer, and a temperature distribution matrix is ​​generated by smoothing the time-series data.

[0029] Infrared thermometers were used to collect real-time temperature data at multiple points within the blanching drum, obtaining multi-point temperature time-series data. The multi-point temperature time-series data was smoothed using a moving average algorithm to obtain smoothed time-series data. Based on the stable flow distribution characteristics, spatiotemporal features were extracted from the smoothed time-series data to obtain the temperature change trend. If the temperature change trend exceeded a preset threshold, local weighted regression analysis was performed on the smoothed time-series data to obtain an optimized temperature sequence. Based on the optimized temperature sequence, a temperature distribution matrix within the blanching drum was constructed to obtain the spatiotemporal temperature distribution. A high-resolution temperature distribution matrix was obtained by performing gridded interpolation on the spatiotemporal temperature distribution. Principal component analysis was used to reduce the dimensionality of the high-resolution temperature distribution matrix to obtain dimensionality-reduced temperature features.

[0030] For example, based on stable flow distribution data, multi-point temperature time-series data inside the blanching cylinder can be obtained using an infrared thermometer, and a temperature distribution matrix can be generated by combining smoothing processing. This can be achieved through the following method. Assume the blanching cylinder is a cylinder with a diameter of 0.5 meters and a length of 1.2 meters, with 10 infrared temperature measurement points evenly distributed on the cylinder wall. The sampling frequency is 1Hz, the acquisition time is 300 seconds, generating 3000 sets of temperature data, with each point's temperature range being 100-250°C. First, data is collected using an infrared thermometer (such as the FLIR T series, with an accuracy of ±2°C). Assume the time-series data for a certain point is [150.2, 151.1, 149.8, ..., 152.3]°C, and the data is stored in CSV format. To ensure data stability, a moving average algorithm is used for smoothing, with a window size of 5. The calculation formula is T_smooth(t) = (T(t-2) + T(t-1) + T(t) + T(t+1) + T(t+2)) / 5. For example, for the above data, the smoothed value at a certain point is (150.2 + 151.1 + 149.8 + 150.5 + 152.3) / 5 = 150.78°C, generating a smoothed time series dataset. To construct the temperature distribution matrix, the 10 temperature measurement points are mapped to a two-dimensional plane. Assuming the cylinder wall unfolded diagram is used as the grid (5×2 matrix), each point corresponds to a grid position. Combined with the time dimension, a three-dimensional matrix T(x,y,t) with a size of 5×2×3000 is generated. If the temperature at a certain point in time is missing, it is filled in using K-nearest neighbor interpolation (K=3). The calculation formula is T_missing = Σ(w_i * T_i) / Σw_i, where w_i is the reciprocal of the distance, based on Euclidean distance.

[0031] For example, if a certain point is missing, and the temperatures of its three neighboring points are 150.5°C, 151.2°C, and 149.7°C, with distances of 0.1m, 0.15m, and 0.2m respectively, and weights w_i of 10, 6.67, and 5 respectively, the interpolation result is (10*150.5 + 6.67*151.2 + 5*149.7) / (10 + 6.67 + 5) ≈ 150.47°C. The final matrix can be used to analyze temperature gradients. Combined with business requirements (such as the uniformity of tea processing), the mean and standard deviation of regional temperatures can be calculated by slicing the matrix. For example, if the mean of a certain region is 150.5°C and the standard deviation is 2.3°C, the processing parameters for tea processing can be optimized.

[0032] Step S106: Based on the temperature distribution matrix, use an anomaly detection algorithm to determine whether there are areas where the local temperature exceeds the dynamic threshold setting, and obtain the coordinates of the temperature anomaly area.

[0033] The data preprocessing steps involve obtaining cleaned temperature data from the temperature distribution matrix. A standardization method is used to denoise and normalize the temperature distribution matrix, resulting in a first temperature dataset. Based on this first temperature dataset, a region segmentation method is employed to divide the data into local temperature regions. A grid partitioning algorithm is used to segment the first temperature dataset into multiple local temperature regions, resulting in a set of local temperature regions. For each local temperature region set, a threshold calculation rule is used to generate dynamic thresholds. The statistical characteristics of each local temperature region are calculated, resulting in a set of dynamic thresholds. An isolated forest algorithm is used to identify regions in the set of local temperature regions that exceed the dynamic thresholds. If the temperature value of a local temperature region exceeds the corresponding dynamic threshold, it is marked as part of the first abnormal region set. Based on the first abnormal region set, a coordinate extraction process is used to obtain temperature anomaly coordinates. A geometric center calculation method is used to extract the center coordinates of each anomaly region from the first abnormal region set, resulting in a first coordinate set. For the first coordinate set, a clustering analysis method is used to optimize the anomaly region identification. The DBSCAN algorithm is used to cluster the first coordinate set, resulting in a second coordinate set. Based on the second coordinate set, the final temperature anomaly coordinates are generated. By using a coordinate mapping method, the second coordinate set is converted into the original coordinate system of the temperature distribution matrix, thus obtaining the final temperature anomaly coordinate set.

[0034] For example, the coordinates of the temperature anomaly region are (3,4). The analysis process first loads a 10×10 temperature matrix. Example data includes a temperature of 34.5℃ at point (3,4) and a temperature of 20.8℃ at point (7,8). The Z-score anomaly detection algorithm is used to calculate the Z-score for each point, with the formula Z = (x - μ) / σ, where x is the point temperature, μ is the matrix mean, and σ is the standard deviation. The matrix mean μ is calculated to be 28.2℃, and the standard deviation σ is 3.5. For point (3,4), Z = (34.5 - 28.2) / 3.5 ≈ 1.8; for point (7,8), Z = (20.8 - 28.2) / 3.5 ≈ -2.11. The dynamic threshold is set to |Z|>2, indicating that a temperature deviation from the mean exceeding two times the standard deviation is considered an anomaly. Point (7,8) has |Z|=2.11>2, thus it is determined to be an anomaly region with coordinates (7,8). The Z=1.8 at point (3,4) does not exceed the threshold and is not abnormal. The algorithm automatically traverses the matrix and outputs all coordinates where |Z|>2, forming a list of abnormal regions. This method is logically rigorous, combining matrix statistical characteristics with threshold judgment, and is suitable for dynamic temperature monitoring systems. It can be correlated with equipment fault detection to ensure that equipment operates within the normal temperature range.

[0035] Step S107: If the coordinates of the abnormal temperature region are not empty, then based on the real-time flow distribution image, the accumulation distribution feature vector and the coordinates of the abnormal temperature region, the stirring speed, the angle adjustment mechanism and the feed speed are adjusted by the control parameter optimization algorithm to obtain a new combination of process parameters.

[0036] If the coordinates of the temperature anomaly region are not empty, pixel intensity values ​​are obtained from the real-time flow distribution image. Combined with the accumulation distribution feature vector, key feature vectors are extracted using principal component analysis (PCA) to obtain feature dimensionality reduction results. Based on the feature dimensionality reduction results, the Euclidean distance between the feature vector and the preset process parameter template is calculated. If the distance exceeds a preset threshold, the stirring speed and feed speed are optimized using gradient descent to obtain adjusted speed parameters. Using the adjusted speed parameters and the coordinates of the temperature anomaly region, the rotation matrix of the angle adjustment mechanism is calculated using a geometric transformation algorithm to obtain optimized angle parameters. From the optimized angle and speed parameters, a new combination of process parameters is constructed. The parameter combination is classified using a support vector machine (SVM) algorithm to determine its stability and obtain classification results. If the classification result is unstable, dynamic feature vectors are re-extracted from the real-time flow distribution image. Combined with the coordinates of the temperature anomaly region, the stirring speed and feed speed are optimized again using gradient descent to obtain updated speed parameters. Based on the updated speed parameters and optimized angle parameters, a new combination of process parameters is reconstructed. The stability of the parameter combination is verified using a SVM algorithm to obtain the final process parameter combination. From the final combination of process parameters, the specific values ​​of stirring speed, angle adjustment mechanism and feeding speed are extracted and stored in the process parameter database to obtain the process parameter record.

[0037] For example, receiving data with non-empty coordinates for a temperature anomaly region, such as coordinates (x1, y1) = (50, 30) and (x2, y2) = (70, 40), indicates a temperature anomaly region within the reactor. First, a real-time flow distribution image is read. Assuming the image is a 256x256 pixel grayscale image with flow values ​​ranging from 0 to 100, features are extracted using a convolutional neural network (CNN) to obtain a 64-dimensional flow feature vector, representing the spatial characteristics of the flow distribution. Next, the packing distribution feature vector is analyzed, assuming it is a 32-dimensional vector containing the packing density (e.g., 0.85 g / cm³). 3Parameters such as particle size (mean 2.5 mm) and packing height (10 cm) were used to reduce the dimensionality to 8 dimensions using principal component analysis (PCA), retaining 95% of the variance to reduce computational complexity. Then, combining the coordinates of the temperature anomaly region, the center point of the anomaly region (x_c, y_c) = ((x1+x2) / 2, (y1+y2) / 2) = (60, 35) was calculated, and the flow rate and packing characteristics within an influence radius r = 15 pixels were determined according to the distance formula d = ((x-x_c)+(y-y_c)). Based on this, a genetic algorithm was used to optimize the control parameters, setting the initial population to 100 and iterating 50 times. The objective function was to minimize the temperature anomaly area (initial area 100 cm²) and the flow rate deviation (target deviation < 5%). Through calculation, new process parameters were obtained: the stirring speed was adjusted from 500 rpm to 550 rpm, the angle adjustment mechanism was increased from 30° to 35°, and the feed rate was reduced from 2 kg / min to 1.8 kg / min. During the verification phase, by simulating the reactor's thermodynamic model, it was predicted that the temperature anomaly area would be reduced to 50 cm² and the flow deviation would be reduced to 3% under the new parameters, thus meeting the optimization objectives.

[0038] In practice, all the above steps can be implemented using Python scripts, data processing relies on NumPy and TensorFlow, the optimization algorithm calls the DEAP library to automatically complete parameter adjustment, and this application does not specifically limit the specific program development and implementation method.

[0039] Step S108: Based on the new process parameter combination, control the blanching equipment to perform stirring and feeding operations through the feedback control system to obtain uniform heating data.

[0040] like Figure 3 As shown, sensors collect material temperature and stirring speed data within the blanching equipment to obtain real-time monitoring data. If the temperature deviation in the real-time monitoring data exceeds a preset threshold, the stirring speed is adjusted through a feedback control system to achieve a stable temperature distribution. Based on the stable temperature distribution, the correlation parameters between the feeding rate and stirring speed are calculated to determine the feeding operation command. The control system sends the feeding operation command to the blanching equipment to execute material input and obtain updated heat distribution data. A support vector machine algorithm is used to classify the heat distribution data and determine the uniform heating state. If the uniform heating state does not meet the preset standard, the stirring and feeding parameters are optimized through the feedback control system to obtain adjusted operation commands. Based on the adjusted operation commands, the blanching equipment is driven to perform stirring and feeding operations to obtain adjusted uniform heating data.

[0041] For example, based on the new combination of process parameters, the blanching equipment executes stirring and feeding operations through a feedback control system to obtain uniform heating data. First, the system receives process parameter input, such as a target temperature set at 120°C, a stirring speed of 300 rpm, and a feeding rate controlled at 2 kg / min. The feedback control system uses a PID algorithm (proportional-integral-derived algorithm) for real-time adjustment, with the proportional coefficient Kp set to 0.8, the integral time Ti to 10 seconds, and the derivative time Td to 2 seconds. The sensor collects the blade surface temperature once per second. Assuming an initial temperature of 80°C, the system calculates the error e = 120 - 80 = 40°C. Using the PID formula u(t) = Kp*e + Ki*∫e(t)dt + Kd*de(t) / dt, a control signal is output to adjust the heating power to 1800 watts and the stirring motor frequency to 50 Hz, ensuring uniform heating of the blades. The feeding system is controlled by a servo motor and calibrates the feeding rate every minute based on data from the hopper weight sensor (accuracy 0.01 kg), with the error controlled within ±0.05 kg. If uneven temperature distribution is detected (standard deviation > 5°C), the system automatically increases the stirring speed by 10% to 330 rpm and analyzes the temperature field through infrared thermal imaging to identify cold spots, dynamically adjusting the feeding position to optimize heat distribution. The data recording module stores the timestamp, temperature, stirring speed, and feeding rate of each operation in JSON format, for example, {"time":"2025-07-25 04:02:00","temp":120.5,"stir_speed":330,"feed_rate":2.02}, and performs real-time analysis through a cloud computing platform to calculate uniformity indicators (temperature standard deviation < 3°C is considered acceptable). If uniformity is not met, the system retrieves historical data and uses a machine learning model (Support Vector Machine, kernel function RBF, C=1.0) to predict the optimal parameter combination, iteratively optimizing until the standard deviation is reduced to 2.8°C to ensure process stability. All operations are automatically executed by a PLC (Programmable Logic Controller), relying on the Modbus protocol to achieve communication between devices and ensure real-time synchronization of process parameters.

[0042] Step S109: Continuously acquire real-time flow distribution images and temperature distribution matrices through the flow sensor array and infrared thermometer, and cyclically execute the above steps in combination with real-time sequence updates to obtain the final stable flow distribution and uniform heating data.

[0043] like Figure 3As shown, after obtaining uniform heating data, real-time flow distribution data and temperature distribution matrices are further collected synchronously using a flow sensor array and an infrared thermometer. Data fusion processing is then used to generate a comprehensive distribution dataset. If the flow distribution deviation in the comprehensive distribution dataset exceeds a preset threshold, the feed rate is adjusted through a feedback control system to obtain optimized flow distribution data. Based on the optimized flow distribution data, a k-means clustering algorithm is used to classify the flow distribution areas and determine the flow distribution equilibrium state. If the flow distribution equilibrium state does not meet the preset standard, time series analysis is used to predict the flow change trend and determine the feed rate adjustment parameters. Based on the feed rate adjustment parameters, a feed control command is generated and sent to the equipment through the control system to obtain updated flow distribution data. The infrared thermometer collects the temperature distribution matrix for the corresponding area of ​​the updated flow distribution data, and a support vector machine algorithm is used to classify the temperature distribution state to determine the uniform heating state. If the uniform heating state does not meet the preset standard, the feed rate and stirring speed parameters are optimized through the feedback control system, generating adjusted operation commands to drive the equipment to execute operations, resulting in the final stable flow distribution and uniform heating data.

[0044] For example, the system collects material flow and temperature distribution data within the blanching equipment every 2 seconds using a flow sensor array (accuracy 0.1 kg / min) and an infrared thermometer (accuracy 0.2°C), generating a real-time flow distribution image and temperature distribution matrix. The flow sensor array consists of 8 nodes, covering the cross-section of the equipment. The collected flow data is represented in matrix form, such as [[2.1, 2.3, 2.0], [2.2, 2.4, 2.1]] (unit: kg / min). The uniformity of the flow distribution is analyzed using a convolutional neural network (CNN, such as 3x3 kernel, stride 1), and the standard deviation of the matrix is ​​calculated, with a target standard deviation < 0.15 kg / min. The infrared thermometer generates a temperature matrix, such as [[118.5, 119.2, 117.8], [119.0, 120.3, 118.2]] (unit: °C). Principal component analysis (PCA) is used to extract the principal features of the temperature field to determine temperature uniformity, with a target standard deviation < 2.5 °C. If the flow rate standard deviation is > 0.15 kg / min, the system uses a gradient descent algorithm (learning rate 0.01) to optimize the feed valve opening, adjusting it until the flow rate matrix standard deviation drops to 0.12 kg / min. If the temperature standard deviation is > 2.5 °C, the system uses K-means clustering (K=3) to identify abnormal temperature areas and automatically adjusts the hot air distribution valve angle (range 0-45°, 5° increments) to direct hot air to cold areas. Data is stored in JSON format, for example, {"time":"2025-07-25 04:04:00", "flow_std":0.12, "temp_std":2.3}, and uploaded to edge computing nodes via the MQTT protocol for real-time analysis of uniformity trends. If uniformity is not met, the system calls a random forest model (100 trees, maximum depth 10) to predict the optimal valve opening and hot air power (range 1000-2000 watts), iteratively adjusting the above operations until the flow rate standard deviation is 0.11 kg / min and the temperature standard deviation is 2.2°C. All operations are coordinated and executed via PLC using the EtherCAT protocol, with a communication latency of <10 milliseconds between sensors and actuators, ensuring real-time performance.

[0045] Example 2 like Figure 4As shown, this invention provides a real-time control system for a tea processing machine, mainly comprising: a flow data acquisition module, used to acquire the flow speed and direction data of tea leaves in the processing cylinder through a flow sensor array, and generate a real-time flow distribution image by combining a data fusion algorithm; a feature extraction module, used to extract the temporal features of the tea leaf accumulation area and the uneven flow area based on the real-time flow distribution image using a convolutional neural network, to obtain an accumulation distribution feature vector; a stirring parameter optimization module, used to optimize the stirring speed and angle adjustment mechanism based on the real-time flow distribution image and temporal features by using a deep reinforcement learning algorithm if the proportion of the accumulation area in the accumulation distribution feature vector exceeds a preset threshold, to obtain optimized stirring parameters; a flow regulation module, used to dynamically adjust the feeding speed based on the optimized stirring parameters and the real-time flow distribution image using a fuzzy control algorithm with the flow distribution uniformity as input, to obtain stable flow distribution data; and a temperature data acquisition module, used to adjust the temperature based on the stable flow distribution data by using an infrared sensor array to obtain a real-time flow distribution image; and a temperature data acquisition module, used to adjust the flow rate based on the stable flow distribution data by using an infrared sensor array to obtain a real-time flow distribution image; and a flow rate adjustment module, used to adjust the flow rate based on the real-time flow distribution image by using an infrared sensor array to obtain a real-time flow distribution image; and a flow rate adjustment module, used to adjust the flow rate based on the real-time flow distribution image by using an infrared sensor array to obtain a real-time flow distribution image; and a flow rate data acquisition ... data acquisition module, used to adjust the flow rate based on the real-time flow distribution An external thermometer acquires time-series temperature data from multiple points within the blanching cylinder, and combines this data with smoothing to generate a temperature distribution matrix. An anomaly detection module uses an anomaly detection algorithm based on the temperature distribution matrix to determine if there are areas where the local temperature exceeds a dynamic threshold, obtaining the coordinates of these anomaly areas. A process parameter optimization module, if the coordinates of the anomaly areas are not empty, adjusts the stirring speed, angle adjustment mechanism, and feeding speed using a control parameter optimization algorithm based on the real-time flow distribution image, accumulation distribution feature vector, and anomaly area coordinates to obtain a new combination of process parameters. An execution control module controls the blanching equipment to perform stirring and feeding operations through a feedback control system based on the new combination of process parameters, obtaining uniform heating data. A cycle monitoring module continuously acquires real-time flow distribution images and temperature distribution matrices through a flow sensor array and an infrared thermometer, and cyclically executes the above steps in conjunction with real-time sequence updates to obtain the final stable flow distribution and uniform heating data.

[0046] Example 3 This embodiment uses the withering process of premium West Lake Longjing tea as an example to specifically illustrate the application of the method of the present invention. West Lake Longjing tea has a flat appearance and high moisture content, and the requirements for withering temperature and uniformity of stirring are extremely stringent, making it very easy to produce scorched edges or insufficient withering.

[0047] System initialization: A drum-type tea-fixing machine with a diameter of 0.8 meters and a length of 1.5 meters was selected. Twelve ultrasonic flow sensors and twelve infrared thermal imaging thermometers were spirally deployed along the inner wall of the drum, forming a high-density sensor network. The ideal tea-fixing process parameters for West Lake Longjing tea were set as follows: target temperature 180°C-220°C, stirring speed 25-45 rpm, and initial moisture content of the tea leaves approximately 75%.

[0048] Steps S101-S102: Flow Status Perception and Accumulation Identification. Feeding is initiated, with fresh Longjing tea leaves entering the processing drum at a rate of 2.5 kg / min. The system collects data from all sensors at a frequency of 10 Hz. After processing by a data fusion algorithm (based on Kalman filtering), a real-time flow distribution map of 128x128 pixels is generated. Due to the flat and easily adhered characteristics of Longjing tea leaves, after approximately 30 seconds of operation, the CNN model (a pre-trained lightweight MobileNetV2 model) identifies, from five consecutive frames, a region below the drum near the discharge outlet where the tea flow velocity is significantly lower than the average (<0.1 m / s), forming a "non-uniform flow region" covering 18% of the area. The system extracts the temporal features of this region, generating a 32-dimensional accumulation distribution feature vector. (See attached image.) Figure 5 As shown, the left image dynamically displays the flow speed and direction of tea leaves using a heat map and vector arrows, and highlights the slow-flowing accumulation areas; the right image uses a heat map to show the temperature distribution, and marks local hot spots with a star-shaped marker.

[0049] Steps S103-S104: Flow-based stirring and initial feed adjustment. Since the 18% accumulation area exceeds the preset 15% threshold, the system triggers the Deep Reinforcement Learning (DRL) optimization module. The reward function of this DRL model is designed as Reward = w1 * (1 - flow_std) - w2 * energy_consumption, where flow_std is the standard deviation of the flow rate. Based on the current accumulation feature vector, the model outputs the optimal action: increasing the stirring speed from 30 rpm to 38 rpm, and simultaneously adjusting the stirring blade angle from 45 degrees to 55 degrees to enhance the peeling and lifting effect on the tea leaves on the cylinder wall. Simultaneously, based on the input of "low flow distribution uniformity," the fuzzy controller temporarily reduces the feed rate by 5% to 2.375 kg / min to reduce the system load and facilitate the dispersion of the accumulated tea leaves.

[0050] Steps S105-S107: Temperature Anomaly Detection and Coordinated Control. After adjusting the stirring parameters, the flow uniformity was improved. However, after about 1 minute, as... Figure 5As shown in the temperature distribution matrix, a localized high-temperature zone with a diameter of approximately 5 cm appears near coordinates (x=0.8, y=π / 2), reaching a temperature of 225°C, exceeding the dynamic threshold set for Longjing tea (current average temperature 190°C + 20% standard deviation). Anomaly detection algorithms (such as the Isolation Forest algorithm) immediately mark this area as a temperature anomaly. At this point, the control parameter optimization algorithm receives three key inputs: 1) a flow distribution map (showing that the overall flow has stabilized); 2) an accumulation feature vector (showing that there is no severe accumulation); and 3) the coordinates of the temperature anomaly. The algorithm determines that this hot spot may be due to abnormal power of a single heating element or a small amount of tea leaves sticking together.

[0051] Steps S108-S109: Precise Execution and Continuous Optimization. The system generates a new combination of process parameters: 1) Maintaining the optimized stirring speed of 38 rpm, but finely adjusting the angles of several sets of stirring blades near the overheating point to increase local disturbance; 2) Restoring the feed rate to 2.5 kg / min; 3) Slightly reducing the power of the heating elements in the corresponding area via the PID controller. After the operation is executed, the system continuously monitors the temperature. Within 15 seconds, the temperature of the overheating point drops to 210°C, returning to a safe range. The entire process is repeated until the moisture content of the tea leaves drops to the target value, completing the fixation process.

[0052] like Figure 6 As shown, compared with the traditional method relying on manual operation by experienced tea masters, the West Lake Longjing tea produced using the withering machine control system of this embodiment reduces the rate of scorched edges and bursting points in the finished product from an average of 8% to below 1%, while increasing the rate of high-quality products (bright green color and pure aroma) by approximately 15%. Simultaneously, by avoiding ineffective heating and excessive stirring, total energy consumption is reduced by 12%. This invention successfully transforms artificial intelligence technology into an effective tool for solving specific technological challenges, achieving standardized and intelligent production of high-quality tea. The above are merely preferred embodiments of this invention and do not limit the patent scope of this invention. Any equivalent structural or procedural transformations made based on the description and drawings of this invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this invention.

Claims

1. A real-time control method for a tea fixing machine, characterized in that, The method includes: The flow velocity and direction data of tea leaves inside the processing drum are acquired using a flow sensor array, and a real-time flow distribution image is generated by combining the data fusion algorithm. Based on the real-time flow distribution image, a convolutional neural network is used to extract the temporal features of the tea leaf accumulation area and the uneven flow area, resulting in an accumulation distribution feature vector. If the proportion of the accumulation area in the accumulation distribution feature vector exceeds a preset threshold, a deep reinforcement learning algorithm is used to optimize the stirring speed and angle adjustment mechanism based on the real-time flow distribution image and temporal features, resulting in optimized stirring parameters. Based on the optimized stirring parameters and the real-time flow distribution image, a fuzzy control algorithm is used with the flow distribution uniformity as input to dynamically adjust the feeding speed, obtaining stable flow distribution data. Based on stable flow distribution data, time-series data of multiple temperatures within the blanching cylinder are acquired using an infrared thermometer. This time-series data is then smoothed to generate a temperature distribution matrix. Based on the temperature distribution matrix, an anomaly detection algorithm is used to determine if there are any areas where the local temperature exceeds a dynamically set threshold, thus obtaining the coordinates of the abnormal temperature areas. If the coordinates of the abnormal temperature areas are not empty, the stirring speed, angle adjustment mechanism, and feeding speed are adjusted using a control parameter optimization algorithm based on the real-time flow distribution image, the accumulation distribution feature vector, and the coordinates of the abnormal temperature areas, resulting in a new combination of process parameters. Based on this new combination of process parameters, the blanching equipment is controlled to perform stirring and feeding operations through a feedback control system to obtain uniformly heated data.

2. The real-time control method for a tea fixing machine according to claim 1, characterized in that, The process of acquiring the flow velocity and direction data of tea leaves inside the fixing cylinder through a flow sensor array, and generating a real-time flow distribution image by combining the data fusion algorithm, includes: The flow velocity and direction data of tea leaves in the fixing cylinder are collected by a flow sensor array, and the data are fused by a Kalman filter algorithm to generate the first flow distribution data. If the variance of the first flow distribution data is greater than a preset threshold, then the velocity data and direction data are denoised to obtain the second flow distribution data. Based on the second flow distribution data, the data is smoothed using a mean filtering algorithm to generate the third flow distribution data; using the third flow distribution data, a two-dimensional convolution algorithm is used to generate the initial flow distribution image. If the edge sharpness of the initial flow distribution image is lower than a preset threshold, edge enhancement processing is performed on the initial flow distribution image to generate a first flow distribution image; based on the first flow distribution image, a color mapping method is used to generate a real-time flow distribution image. By extracting key feature points of tea leaf flow from real-time flow distribution images, the final flow distribution image is generated.

3. The real-time control method for a tea fixing machine according to claim 1, characterized in that, The step involves extracting temporal features of tea leaf accumulation areas and uneven flow areas using a convolutional neural network based on real-time flow distribution images, resulting in an accumulation distribution feature vector, including: Initial image data is obtained from real-time traffic distribution images. The images are then denoised and standardized through preprocessing operations to obtain the first image data. A convolutional neural network is used to extract features from the first image data, and region segmentation is performed on the tea accumulation area and the uneven flow area to obtain the region segmentation results. If the proportion of pixels in the tea-leaf accumulation area in the region segmentation result exceeds a preset threshold, then temporal features are extracted from that area to generate a first temporal feature set; if it is below the preset threshold, then it is marked as an uneven flow area to generate a second temporal feature set. The first and second time-series feature sets are integrated using a spatiotemporal feature fusion algorithm to obtain a fused feature set. Based on the fused feature set, the features are dimensionality-reduced using a principal component analysis algorithm to generate dimensionality-reduced feature vectors. Dynamic distribution analysis is performed on the dimensionality-reduced feature vector to detect changes in traffic patterns and obtain a stacked distribution feature vector. Key distribution parameters are extracted from the stacked distribution feature vector to generate the final feature vector representation.

4. The real-time control method for a tea fixing machine according to claim 1, characterized in that, If the proportion of the accumulation region in the accumulation distribution feature vector exceeds a preset threshold, then through a deep reinforcement learning algorithm, based on the real-time flow distribution image and temporal features, an optimized stirring speed and angle adjustment mechanism is used to obtain optimized stirring parameters, including: Real-time traffic distribution images and time-series feature sequences are acquired from sensors to generate a stacked distribution feature vector. If the proportion of the region of the stacked distribution feature vector exceeds the preset threshold, the real-time traffic distribution image is denoised and segmented by the image processing algorithm to obtain the processed traffic distribution image. Based on the processed traffic distribution image, a convolutional neural network is used to extract spatial features, which are then combined with temporal feature sequences to generate a comprehensive feature vector. By using deep reinforcement learning algorithms and based on comprehensive feature vectors, the stirring speed and angle adjustment parameters are optimized to obtain preliminary optimized control parameters; If the deviation between the simulation results of the initially optimized control parameters and the target control parameters exceeds the preset range, the weights of the deep reinforcement learning model are adjusted using the gradient descent algorithm to obtain the updated optimized control parameters. Based on the updated optimized control parameters, control commands for the mixing equipment are generated and output to the actuator; the time-series feature sequence is updated using real-time operating data fed back by the actuator, and a new accumulation distribution feature vector is generated.

5. The real-time control method for a tea fixing machine according to claim 1, characterized in that, The process involves using a fuzzy control algorithm, with flow distribution uniformity as input, to dynamically adjust the feed rate based on optimized stirring parameters and real-time flow distribution images, thereby obtaining stable flow distribution data. This includes: Real-time flow distribution data is obtained by collecting flow distribution information in the pipeline through sensors and removing noise using data preprocessing methods to obtain the first flow distribution dataset. Based on the first traffic distribution dataset, the traffic uniformity index is calculated. If the uniformity index is lower than the preset threshold, the fuzzy control algorithm is triggered to obtain an adjustment signal. The adjustment signal is processed by a fuzzy control algorithm. Based on a predefined fuzzy rule base, the feed speed adjustment amount is calculated to obtain the first speed adjustment parameter. By adjusting the parameters of the first speed, the operating speed of the feeding device is dynamically adjusted, the flow distribution in the pipeline is updated, and a second flow distribution dataset is obtained. For the second traffic distribution dataset, the traffic uniformity index is recalculated. If the uniformity index is still lower than the preset threshold, the fuzzy rule base is iteratively adjusted to obtain the optimized rule parameters. Based on the optimized rule parameters, the fuzzy control algorithm is updated, the feed speed adjustment is recalculated, and the second speed adjustment parameter is obtained. By using the second speed adjustment parameter, the speed of the feeding device is continuously adjusted, and the flow distribution in the pipeline is monitored to obtain stable flow distribution data.

6. The real-time control method for a tea fixing machine according to claim 1, characterized in that, The process involves acquiring time-series temperature data at multiple points within the blanching cylinder using an infrared thermometer based on stable flow distribution data, and then smoothing the time-series data to generate a temperature distribution matrix, including: Infrared thermometers were used to collect real-time temperature data at multiple points inside the blanching cylinder, resulting in multi-point temperature time-series data. The multi-point temperature time series data is smoothed by using the moving average algorithm to obtain smoothed time series data; Based on the stable flow distribution characteristics, spatiotemporal features are extracted from smooth time-series data to obtain temperature change trends; If the temperature change trend exceeds the preset threshold, a local weighted regression analysis is performed on the smoothed time series data to obtain an optimized temperature sequence; based on the optimized temperature sequence, a temperature distribution matrix inside the blanching cylinder is constructed to obtain the spatiotemporal temperature distribution. A high-resolution temperature distribution matrix is ​​obtained by performing gridded interpolation on the spatiotemporal temperature distribution; the dimensionality of the high-resolution temperature distribution matrix is ​​then reduced using principal component analysis to obtain the dimensionality-reduced temperature features.

7. The real-time control method for a tea fixing machine according to claim 1, characterized in that, The step of determining whether there are areas where the local temperature exceeds the dynamic threshold setting based on the temperature distribution matrix and obtaining the coordinates of the temperature anomaly areas includes: The cleaned temperature data is obtained from the temperature distribution matrix through a data preprocessing step. The temperature distribution matrix was denoised and normalized using a standardization method to obtain the first temperature dataset; Based on the first temperature dataset, a region segmentation method is used to divide local temperature regions. The first temperature dataset is divided into multiple local temperature regions using a grid partitioning algorithm, resulting in a set of local temperature regions. For a set of local temperature regions, a dynamic threshold is generated using threshold calculation rules; Calculate the statistical characteristics of each local temperature region to obtain a dynamic threshold set; The isolated forest algorithm is used to identify regions that exceed the dynamic threshold set from the set of local temperature regions; If the temperature value of a local temperature region exceeds the corresponding dynamic threshold, it is marked as the first abnormal region set; Based on the first set of abnormal regions, the coordinates of the temperature anomalies are obtained using a coordinate extraction process; By using the geometric center calculation method, the center coordinates of each abnormal region are extracted from the first abnormal region set to obtain the first coordinate set; For the first set of coordinates, cluster analysis is used to optimize the identification of abnormal regions; The first coordinate set is clustered using the DBSCAN algorithm to obtain the second coordinate set; Based on the second coordinate set, generate the final temperature anomaly coordinates; By using a coordinate mapping method, the second coordinate set is converted into the original coordinate system of the temperature distribution matrix, thus obtaining the final temperature anomaly coordinate set.

8. The real-time control method for a tea fixing machine according to claim 1, characterized in that, If the coordinates of the temperature anomaly region are not empty, then based on the real-time flow distribution image, the accumulation distribution feature vector, and the coordinates of the temperature anomaly region, the stirring speed, angle adjustment mechanism, and feed rate are adjusted through a control parameter optimization algorithm to obtain a new combination of process parameters, including: If the coordinates of the temperature anomaly area are not empty, then the pixel intensity value is obtained from the real-time flow distribution image, combined with the stacked distribution feature vector, and the key feature vector is extracted by the principal component analysis algorithm to obtain the feature dimensionality reduction result. Based on the feature reduction results, the Euclidean distance between the feature vector and the preset process parameter template is calculated. If the distance exceeds the preset threshold, the stirring speed and feeding speed are optimized by the gradient descent algorithm to obtain the adjusted speed parameters. By combining the adjusted velocity parameters with the coordinates of the temperature anomaly area, the rotation matrix of the angle adjustment mechanism is calculated using a geometric transformation algorithm to obtain the optimized angle parameters. Based on the optimized angle and speed parameters, a new combination of process parameters is constructed. The support vector machine algorithm is used to classify the parameter combination, determine the stability of the parameter combination, and obtain the classification result. If the classification result is unstable, the dynamic feature vector is re-extracted based on the real-time flow distribution image, and combined with the coordinates of the temperature anomaly area, the stirring speed and feeding speed are optimized again through the gradient descent algorithm to obtain the updated speed parameters. Based on the updated speed parameters and optimized angle parameters, a new combination of process parameters is reconstructed. The stability of the parameter combination is verified by the support vector machine algorithm to obtain the final combination of process parameters. From the final combination of process parameters, the specific values ​​of stirring speed, angle adjustment mechanism and feeding speed are extracted and stored in the process parameter database to obtain the process parameter record.

9. The real-time control method for a tea fixing machine according to claim 1, characterized in that, After obtaining the uniform heating data, the process further includes: continuously acquiring real-time flow distribution images and temperature distribution matrices through a flow sensor array and an infrared thermometer, and cyclically executing the above steps in conjunction with real-time sequence updates to obtain the final stable flow distribution and uniform heating data.

10. A real-time control system for a tea fixing machine, characterized in that, The system includes: a flow data acquisition module, used to acquire the flow velocity and direction data of tea leaves in the fixing drum through a flow sensor array, and generate a real-time flow distribution image by combining a data fusion algorithm; a feature extraction module, used to extract the temporal features of the tea leaf accumulation area and the uneven flow area based on the real-time flow distribution image using a convolutional neural network, to obtain an accumulation distribution feature vector; a stirring parameter optimization module, used to optimize the stirring speed and angle adjustment mechanism based on the real-time flow distribution image and temporal features using a deep reinforcement learning algorithm if the proportion of the accumulation area in the accumulation distribution feature vector exceeds a preset threshold, to obtain optimized stirring parameters; a flow regulation module, used to dynamically adjust the feeding speed based on the optimized stirring parameters and the real-time flow distribution image using a fuzzy control algorithm with the flow distribution uniformity as input, to obtain stable flow distribution data; and a temperature data acquisition module, used to acquire the temperature data of the tea leaves in the fixing drum using an infrared thermometer based on the stable flow distribution data. The system generates a temperature distribution matrix by combining time-series temperature data from multiple points with smoothing processing. An anomaly detection module uses an anomaly detection algorithm based on the temperature distribution matrix to determine if there are areas where the local temperature exceeds a dynamic threshold, obtaining the coordinates of the temperature anomaly areas. A process parameter optimization module, if the coordinates of the temperature anomaly areas are not empty, adjusts the stirring speed, angle adjustment mechanism, and feed speed using a control parameter optimization algorithm based on the real-time flow distribution image, accumulation distribution feature vector, and temperature anomaly area coordinates to obtain a new combination of process parameters. An execution control module controls the blanching equipment to perform stirring and feeding operations through a feedback control system based on the new combination of process parameters, obtaining uniform heating data. A cyclic monitoring module continuously acquires real-time flow distribution images and temperature distribution matrices through a flow sensor array and an infrared thermometer, and cyclically executes the above steps in conjunction with real-time sequence updates to obtain the final stable flow distribution and uniform heating data.

Images