Flower medicinal material drying process self-adaptive regulation and control method and intelligent drying machine

Abstract

The invention belongs to the technical field of intelligent processing of agricultural products, and provides a self-adaptive regulation and control method for a flower medicinal material drying process and an intelligent drying machine. The method comprises the steps of data acquisition, image segmentation and historical state sequence updating, quality state prediction, prediction evaluation matrix construction and normalization, comprehensive evaluation, target set value updating and drying cavity parameter regulation and control. According to the method, a technical closed loop of real-time sensing, dynamic prediction, optimization decision-making and self-adaptive control is constructed, dynamic optimization of drying process parameters according to state changes of materials is achieved, the accuracy and the intelligent level of the drying process are improved, the final quality of the flower medicinal materials is guaranteed, and meanwhile the quality of the flower medicinal materials is improved. The drying efficiency is improved, and the unit energy consumption is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent agricultural product processing technology, and in particular to an adaptive control method for drying flower-based medicinal materials and an intelligent dryer. Background Technology

[0002] Drying is a key step in the initial processing of flower-based Chinese medicinal materials. The process involves moisture migration, heat and mass transfer, and the coupling of multiple physical fields such as temperature, humidity and wind speed, exhibiting typical black box dynamic response characteristics.

[0003] Current drying processes for floral medicinal materials still heavily rely on operational experience and static process parameter settings. They typically focus only on quality changes before and after drying, neglecting the dynamic evolution of physicochemical properties throughout the drying process. Traditional drying equipment maintains unchanged parameters after process settings, lacking online sensing and dynamic control capabilities for the material's state, and cannot adaptively adjust based on real-time feedback. This contradiction between static process control and the nonlinear, time-varying characteristics inherent in the drying process of floral medicinal materials has become a key bottleneck restricting the stability of dried quality, energy efficiency, and the improvement of processing intelligence. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an adaptive control method for the drying process of floral medicinal materials and an intelligent dryer, which solves the problems of traditional methods neglecting the dynamic evolution of physicochemical properties throughout the drying process and the lack of online perception and dynamic control of material state in static process control.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] An adaptive control method for drying processes of floral medicinal materials, comprising:

[0007] The original image to be detected and the sensor data to be detected in the drying chamber are acquired using a preset image acquisition module and a multi-source sensor information acquisition module.

[0008] The deployed image segmentation model is used to process the original image to be detected to obtain image segmentation data. The image segmentation data and the sensor data to be detected are integrated to obtain a quality state vector. The quality state vector is used to update the historical state sequence in the time series database.

[0009] The updated historical state sequence is input into the deployed quality prediction model for prediction, and a quality state prediction vector is obtained under each candidate scheme condition.

[0010] A prediction evaluation matrix is ​​constructed based on the quality status prediction vector, and the prediction evaluation matrix is ​​normalized to obtain a normalized evaluation matrix.

[0011] The normalized evaluation matrix is ​​calculated using a preset comprehensive evaluation formula to obtain a comprehensive score;

[0012] The combination of process parameters corresponding to the highest comprehensive score will be used as the target setting value for the next control cycle.

[0013] The temperature, humidity, and wind speed inside the drying chamber are controlled using a smooth transition strategy based on the target set values.

[0014] The present invention discloses the following technical effects:

[0015] This invention provides an adaptive control method for drying floral medicinal materials and an intelligent dryer. Through an image segmentation model, it overcomes the shortcomings of traditional drying processes that rely on manual observation and have strong subjectivity in quality monitoring, thereby improving the segmentation accuracy of target regions in irregular floral medicinal materials. Through a quality prediction model, it addresses the problem of poor prediction performance of traditional mechanistic models or static data models, enabling in-depth mining of complex nonlinear relationships hidden in multi-source time-series data. Through multi-index comprehensive evaluation based on the entropy weight method, it solves the complex, multi-objective problem of drying process optimization, achieving dynamic and collaborative optimization of process parameters. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the adaptive control process for drying floral medicinal materials provided in an embodiment of the present invention;

[0018] Figure 2 This is a front view of the intelligent dryer for flower materials provided in an embodiment of the present invention;

[0019] Figure 3 The hardware and software connection diagram of the intelligent drying machine for flower-type medicinal materials based on image and sensor fusion provided in the embodiments of the present invention;

[0020] Figure 4 An improved U-Net model architecture diagram provided for embodiments of the present invention;

[0021] Figure 5 This is a diagram of a CNN-LSTM architecture with a self-attention mechanism provided in an embodiment of the present invention.

[0022] Figure 6 The above is an adaptive control logic diagram of process parameters for the drying process of flower materials provided in an embodiment of the present invention.

[0023] Explanation of reference numerals in the attached figures:

[0024] 1-Industrial camera, 2-LED lighting source, 3-Drying chamber, 4-Water bath, 5-Electronic touch screen, 6-Wind speed sensor, 7-Circulating centrifugal fan, 8-Electric heating element, 9-Tray support, 10-Exhaust outlet, 11-Return air duct, 12-Atomizing nozzle, 13-Weighing sensor, 14-Weighing tray, 15-Temperature and humidity sensor, 16-Turbulence fan, 17-Thermal insulation layer. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] The purpose of this invention is to provide an adaptive control method for the drying process of floral medicinal materials and an intelligent dryer, which solves the problems of traditional methods neglecting the dynamic evolution of physicochemical properties throughout the drying process and the lack of online perception and dynamic control of material state in static process control.

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Figure 1 This is a schematic diagram of the adaptive control process for drying floral medicinal materials provided in an embodiment of the present invention, as shown below. Figure 1 As shown, this invention provides an adaptive control method for the drying process of floral medicinal materials, comprising:

[0029] Step 100: Use the preset image acquisition module and multi-source sensor information acquisition module to acquire the original image to be detected and the sensor data to be detected in the drying chamber;

[0030] Step 200: Process the original image to be detected using the deployed image segmentation model to obtain image segmentation data, integrate the image segmentation data and the sensor data to be detected to obtain a quality state vector, and use the quality state vector to update the historical state sequence in the time series database;

[0031] Step 300: Input the updated historical state sequence into the deployed quality prediction model for prediction to obtain the quality state prediction vector under each candidate scheme condition;

[0032] Step 400: Construct a prediction evaluation matrix based on the quality status prediction vector, and normalize the prediction evaluation matrix to obtain a normalized evaluation matrix;

[0033] Step 500: Calculate the normalized evaluation matrix using a preset comprehensive evaluation formula to obtain a comprehensive score;

[0034] Step 600: Use the combination of process parameters corresponding to the highest comprehensive score as the target set value for the next control cycle;

[0035] Step 700: Adjust the temperature, humidity, and wind speed in the drying chamber according to the target set value using a smooth transition strategy.

[0036] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the embodiments of this application will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0037] Example 1: A smart flower dryer. (Refer to...) Figure 2 This embodiment provides an intelligent dryer for floral medicinal materials based on image and multi-source sensor fusion. This device provides hardware support and an execution platform for the adaptive process control method in Embodiment 2. The device mainly includes a drying module, an image acquisition module, a multi-source sensor information acquisition module, and a control module.

[0038] Specifically, in this embodiment, the drying module is primarily a drying chamber 3 with good thermal insulation properties. The interior of the drying chamber 3 is divided into left and right drying chambers by a vertically arranged partition. Multiple layers of tray supports 9 are evenly and parallelly arranged along the height direction on the inner walls of both sides of the drying chambers to support trays containing flower-like materials. To further enhance the uniformity of the drying process and effectively break the static humid air boundary layer on the material surface, multiple small turbulence fans 16 are installed at positions corresponding to each layer of trays. These turbulence fans 16 generate localized micro-airflows during operation, significantly improving heat and mass exchange efficiency, thereby increasing the drying rate and the uniformity of the finished product.

[0039] Furthermore, the drying module further includes a hot air circulation system. In this embodiment, the system mainly includes a circulating centrifugal fan 7 and an electric heating element 8. During operation, air inside the drying chamber is drawn into the drying chamber by the centrifugal fan through the return air inlet at one end of the chamber. The airflow is rapidly heated to the target temperature as it flows over the surface of the electric heating element 8. Subsequently, the high-temperature airflow is evenly blown into the drying chamber, passing through the materials on each tray from top to bottom, and efficiently transferring heat and mass to the materials via convection. After the exchange is completed, the air, with its temperature decreasing and humidity increasing, finally flows back into the drying chamber through the return air inlet, thus forming a continuous and stable hot air circulation.

[0040] Furthermore, the drying module includes a humidity control device for precise control of the relative humidity within the drying chamber. In this embodiment, the device includes a humidification component and a dehumidification component. The humidification component specifically includes a water bath 4, an electromagnetic pump, and an atomizing nozzle 12. The electromagnetic pump is located inside the water bath 4. When the control module determines that the humidity within the chamber needs to be increased, the electromagnetic pump starts, pumping water (preferably warm water to reduce disturbance to the temperature within the chamber) from the water bath 4 to the atomizing nozzle 12 installed in the air duct. The spray volume is precisely controlled by the humidification solenoid valve, and the sprayed fine water mist evaporates rapidly under the action of the high-temperature airflow, thereby effectively increasing the humidity of the circulating gas. The dehumidification component mainly includes an independently installed dehumidification centrifugal fan. When the control module determines that the humidity within the chamber needs to be reduced, the dehumidification fan starts, discharging some of the high-humidity air within the drying chamber to the outside of the chamber through the dehumidification outlet 10. At the same time, relatively dry air from the outside can be replenished into the circulation system through the circulating air inlet, thereby achieving a rapid reduction in the relative humidity within the chamber.

[0041] Specifically, the image acquisition module is designed to acquire the visual appearance characteristics of materials during the drying process online. An observation window is located at the top of the drying chamber 3, and an industrial camera 1 is mounted thereon. To prevent fogging of the lens in humid and hot environments, the lens is preferably positioned behind a sealed glass cover. To ensure the stability and consistency of image acquisition quality under different lighting conditions and drying stages, multiple LED lighting sources 2 are arranged around the lens of the industrial camera 1. The industrial camera 1 is directly connected to the workstation in the control module via a Gigabit Ethernet (GigE) data cable for real-time transmission of high-resolution digital images.

[0042] Furthermore, a multi-source sensor information acquisition module is used to comprehensively and in real-time perceive the multi-dimensional physical state during the drying process. In this embodiment, this module integrates multiple sensors:

[0043] 1) Temperature and humidity sensor 15: at least one, preferably with its probe placed near the return air vent of the drying chamber, because the air parameters at this location can most accurately reflect the average state after heat and moisture exchange with all material layers.

[0044] 2) Wind speed sensor 6: Installed in the main air supply duct that sends the heated airflow into the drying chamber, used to monitor the circulating wind speed in real time.

[0045] 3) Weight sensor: Preferably, a weighing sensor 13 (e.g., a cantilever beam weighing sensor 13) is installed below at least one layer of pallet support 9. This sensor forms an integrated weighing unit with the pallet support 9, used to monitor the weight change of the material due to moisture evaporation in real time and without disturbance, providing basic data for accurate calculation of the drying rate.

[0046] 4) Electronic touch screen 5: Integrated and installed on the outer side panel of the drying chamber 3, serving as a human-machine interface for operators to set initial process parameters, monitor drying data in real time (such as temperature, humidity, and weight curves), and make manual interventions when necessary.

[0047] All of the above sensors communicate with the control module stably and reliably via an industrial fieldbus (such as RS-485).

[0048] Specifically, the control module is the core of the entire intelligent drying device's computation and control. In this embodiment, it adopts a layered hardware architecture, consisting of three cooperating parts:

[0049] 1) Workstation: As the host computer, it is usually a high-performance industrial control computer (IPC) or server. It is responsible for performing high-level, computationally intensive tasks, including: aggregating and processing multi-dimensional data from image acquisition modules and multi-source sensor information acquisition modules; running data preprocessing algorithms, deep learning semantic segmentation models, multivariate time series prediction models, and multi-objective process optimization decision algorithms; and finally generating optimal process control instructions.

[0050] 2) Microcontroller: As a lower-level machine, such as the STM32 series microcontroller, it is responsible for executing low-level device drive and control tasks with high real-time requirements. It receives macroscopic process control instructions from the workstation through standard communication interfaces (such as UART, CAN), and outputs precise drive signals (such as PWM signals, relay switching signals) through its I / O ports based on the instruction values ​​and real-time feedback values ​​from sensors, to perform precise closed-loop control of actuators such as electric heating tubes, centrifugal fans, and humidity control devices.

[0051] 3) Data storage unit: This is a non-volatile storage medium, such as a solid-state drive (SSD). It is used to store the operating system, upper / lower-level control programs, trained deep learning model files, historical process databases, and operation logs during the drying process.

[0052] Preferably, Example 2: An adaptive control method for drying floral medicinal materials. This example provides an adaptive control method for the drying process based on the intelligent drying system of Example 1. Figure 3 As shown, this method can be divided into two stages: "offline model construction and preparation" and "online adaptive regulation". The offline stage aims to build and train the core algorithm model required for subsequent online regulation; the online stage is to periodically execute the closed-loop optimization process of "perception-prediction-decision-execution" in the actual drying task.

[0053] Furthermore, offline training and deployment of the model:

[0054] 1) Collection of training data:

[0055] Drying experiments were conducted under various typical combinations of process parameters (e.g., temperature combinations: 40℃, 50℃, 60℃; relative humidity: 20%, 30%, 40%; wind speed: 1.2m / s, 2.0m / s, 2.8m / s). During each experiment:

[0056] a) The image acquisition module is used to acquire original images of flower materials at different drying stages online.

[0057] b) Simultaneously, the real-time weight of the material, temperature, relative humidity, and wind speed inside the drying chamber are collected synchronously through a multi-source sensor information acquisition module. All collected data are accompanied by precise timestamps.

[0058] 2) Construction and training of image segmentation models:

[0059] This embodiment constructs and trains an improved U-Net semantic segmentation model for subsequent online accurate extraction of visual features of floral materials.

[0060] a) Data Labeling and Augmentation: Representative samples are selected from the collected images, and pixel-level background and foreground (flower-like materials) are manually labeled using an image labeling tool (such as LabelMe) to generate corresponding binary label images. Then, data augmentation processing is performed on the labeled images, including but not limited to horizontal / vertical flipping, random rotation, and brightness / contrast adjustment, to expand the dataset size and improve the model's generalization ability. The augmented dataset is then divided into training, validation, and test sets in, for example, an 8:1:1 ratio.

[0061] b) Model Architecture: Reference Figure 4 This embodiment constructs an improved U-Net network that embeds a Squeeze-and-Excitation (SE) attention mechanism module. The model retains the classic encoder-decoder structure and skip connections of U-Net, with its key improvement lying in the deepest part of the U-Net structure (i.e., the bottleneck), where an SE module is embedded. The purpose is to utilize the channel attention mechanism of the SE module to perform adaptive recalibration at key nodes where the feature map is transferred from the encoder to the decoder. This allows the model to explicitly learn the importance of different feature channels, actively enhancing features that contribute significantly to the target region while suppressing background noise features, thus solving the information redundancy problem at the bottleneck. The specific operation flow of the SE module includes:

[0062] Squeeze: Squeezes the input feature map. Global average pooling is performed to compress each two-dimensional feature channel into a single numerical value. This yields a channel description vector. The output of the c-th channel The calculation formula is:

[0063]

[0064] in, This represents the c-th input feature channel; Indicates channel In spatial location The feature values ​​at the location; H and W are the height and width of the feature map, respectively; The global descriptor calculated for the c-th channel after the Squeeze operation; This is the function representation of the Squeeze operation.

[0065] Excitation: The global description vector z is input into a bottleneck structure consisting of two fully connected layers (FC) to learn and generate a normalized importance weight vector for each channel. This process can be represented as:

[0066]

[0067] Where z is the output of the Squeeze operation, i.e., the global description vector; and These are the weight matrices of the two fully connected layers (r is the scaling factor); It is the ReLU activation function; The Sigmoid activation function normalizes the weights to the [0, 1] interval.

[0068] Rescale: The channel weight vector s output from the activation step is multiplied element-wise at the channel level with the original feature map U to obtain the rescaled feature map. The c-th channel The calculation is as follows:

[0069]

[0070] in, The weight scalar for the c-th channel generated in the Excitation step; This represents the c-th channel of the original input feature map U. This is for the c-th feature channel of the recalibrated output; This is the function representation for the recalibration operation.

[0071] c) Model Training Configuration: To ensure that the model can be effectively trained and achieve excellent segmentation performance, this embodiment provides a preferred training configuration scheme:

[0072] Data preprocessing: All input images are uniformly adjusted to a resolution of 256×256 pixels; pixel values ​​are normalized by channel and scaled to the range of [0, 1].

[0073] Training hyperparameters: Batch size is set to 8; the optimizer uses Adam, whose parameters can be set to... =0.9, =0.999, weight decay is 1×10 -4 Initial learning rate: 1×10 -3 It also employs a dynamic learning rate decay strategy.

[0074] Loss function: Pixel-level binary cross-entropy loss is used. ) and Dice loss ( The weighted combination of () is used as the total loss function: , where α and β are the weighting coefficients for balancing the two losses, and in this embodiment, both are taken as 0.5.

[0075] Training strategy: An early stopping mechanism is employed to prevent model overfitting. The Dice coefficient or IoU (Intersection over Union) on the validation set is used as the core monitoring metric. If the metric fails to improve for several consecutive rounds, training is terminated early, and the weight file of the best-performing model on the validation set is saved for subsequent online deployment.

[0076] 3) Data preparation and training for the quality prediction model:

[0077] This embodiment constructs and trains a multivariate time-series prediction model for online prediction of the future quality status of materials. The data preparation and training steps are as follows:

[0078] a) Construction of time-series feature dataset: Using the experimental data collected in the above steps, the following processing is performed:

[0079] Using the semantic segmentation model trained above, all experimental images are processed in batches to obtain binary mask images at each time step.

[0080] Based on the mask, the target region (Region of Interest, ROI) of the material is extracted from the original RGB image, and the color and morphological features of the region are quantized and calculated, as follows:

[0081] Color characteristics: The RGB data of the ROI is gamma-corrected and converted to the CIE-XYZ color space. This conversion is performed under a D65 standard light source, with the normalized white point reference taken as... , , The components of CIELAB are calculated using a nonlinear transformation function.

[0082] The components of the CIELAB color space ( , , The formula for calculating ) is:

[0083]

[0084]

[0085]

[0086] Wherein, the nonlinear transformation function Defined as:

[0087]

[0088] Calculate the values ​​of all pixels in the ROI. , , The arithmetic mean of the two values ​​is used to obtain the average value at that time. , , value.

[0089] For ease of description, , , For time t during the drying process , , average value; , , The initial time of drying ( )of , , average value.

[0090] Furthermore, calculate the color difference ( Whiteness index (WI) and yellowness index (YI):

[0091]

[0092]

[0093]

[0094] Area Shrinkage Ratio (SR): The number of pixels with a value of 255 in a binary mask. Combined with the pre-calibrated camera spatial resolution r (mm) 2 ( / pixel), calculate the current real-time area. Based on the initial area and real-time area Calculate the shrinkage ratio .

[0095] Then, based on real-time weight data, the drying rate (DR) of the material is calculated using the following formula:

[0096]

[0097] in and Drying time and Moisture content (g / g) of seasonal floral medicinal materials on a dry basis.

[0098] Finally, the dynamically changing quality characteristics calculated in the above steps are... Construct a multidimensional feature vector by aligning it with the timestamp. A large number of data points are generated on complete time-series data using the sliding window method. , Sample pairs are used to form the training set, validation set, and test set of the model. The input sequence... Target output Where t is the index at the current time. Let be the five-dimensional eigenvector at time t; d represents the original time step of the sliding window; d represents the feature dimension (here) ); To predict future time steps; The input sequence of the model at time t; Let t be the target vector that the model needs to predict at time t.

[0099] b) Model Architecture: Reference Figure 5 This embodiment constructs a composite prediction model that integrates a convolutional neural network (CNN), a long short-term memory network (LSTM), and a self-attention mechanism. The specific data processing flow is as follows:

[0100] CNN Local Feature Extraction Module: This module is used to extract features from the input sequence. The system automatically extracts salient patterns in the local and temporal dimensions. This embodiment preferably employs a one-dimensional convolutional neural network. For a given... Input sequence The one-dimensional convolution operation is defined as follows:

[0101]

[0102] in, The input is the training sample matrix; This represents a one-dimensional convolution operation; Let be the weights of the c-th convolutional kernel; This is the bias term corresponding to the c-th convolutional kernel; To modify the activation function of the linear unit; This is the feature map extracted by the c-th convolutional kernel. Through parallel processing with multiple convolutional kernels, a set of local feature maps can be obtained and combined into a local feature sequence. .

[0103] To capture the most salient features and shorten the sequence length, a sequence of feature maps composed of convolutional kernels is used. It is fed into a pooling layer. The pooling layer downsamples along the time dimension, reducing the sequence length from... Shortening yields a new, shorter feature sequence. (in ).

[0104] LSTM Long-Term Dependency Modeling Module: In this embodiment, the feature sequences extracted from the pooling layer are... The input is fed into a Long Short-Term Memory (LSTM) network layer. Through its internal input gate, forget gate, and output gate, the LSTM effectively learns and remembers long-term dependencies in the time series. Its core state update equations are as follows:

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111] in, For the Sigmoid function; It is the hyperbolic tangent activation function; The product of Hadamard; The input feature vector is fed into the LSTM module at time t; Let be the hidden state vector at time t-1; Let be the output hidden state vector at time t; Let be the cell state vector at time t-1; Let be the candidate cell state vector at time t; Let be the final cell state vector at time t; , , These are the activation vectors of the forget gate, input gate, and output gate at time t, respectively. , , , These are the weight matrices for the forget gate, input gate, cell state, and output gate, respectively. , , , These are the bias vectors for the forget gate, input gate, cell state, and output gate, respectively.

[0112] Self-attention mechanism module: To enable the model to dynamically evaluate the importance of features extracted by LSTM at different time steps, this embodiment introduces a multi-head self-attention module after the LSTM module. This module receives the LSTM's attention from all time steps. The hidden state sequence output at each time step (dimension is) The sequence serves as the input. It simultaneously provides the query (Q), key (K), and value (V). Its core operations can be summarized as follows:

[0113]

[0114] in, Let be the dimension of the LSTM hidden layer; This is the dimension of the key vector. The final output of this module is an attention-weighted sequence of hidden states. .

[0115] c) Model output and training: The sequence output by the attention module. Transform into a single context vector .

[0116] In a preferred embodiment, by processing the sequence The context vector is obtained by applying global average pooling in the time dimension. As shown in the following formula:

[0117]

[0118] Finally, the context vector Through a fully connected layer mapping, the output is a view of the future. Feature vector at time step Prediction:

[0119]

[0120] in, This is the final predicted output vector of the model; and To predict the weight matrix and bias vector of the output layer.

[0121] To train this model, this embodiment employs a loss function that combines weighted mean squared error with L2 regularization. :

[0122]

[0123] Where N is the number of samples, and J is the total number of targets to be predicted. The loss weight for the j-th predicted target, Let be the model's predicted value for the i-th sample and the j-th target. Let be the true value of the i-th sample and the j-th target; λ be the L2 regularization coefficient; and θ be the set of all learnable parameters of the model. An adaptive learning rate optimizer, such as Adam's, is used to iteratively train the model to minimize this loss function.

[0124] Specifically, the online adaptive control process refers to... Figure 6 After the actual drying task begins, the system puts the deployed offline trained model into use and executes an optimization process of "perception-prediction-decision-execution" in a fixed adjustment cycle (e.g., 15 minutes). Within one adjustment cycle, the specific implementation steps are as follows:

[0125] Step A: Real-time data acquisition and feature extraction:

[0126] At the start of a control cycle, the system acquires real-time raw images, material weight, and temperature, relative humidity, and wind speed data within the drying chamber through the image acquisition module and multi-source sensor information acquisition module. Subsequently, the workstation uses a deployed image segmentation model to process the real-time images and, combined with the sensor data, calculates the current quality state vector. The vector is then stored in the time series database to update the historical state sequence.

[0127] Step B: Predicting Future Quality Status

[0128] The workstation takes historical time series data containing the latest status as input and feeds it into the deployed quality prediction model. This is for a pre-defined set of candidate process parameter combinations. (in, For a specific set of (T, v, RH) parameters, The model will process each candidate solution separately. Perform forward reasoning calculations to obtain the future under each candidate solution condition. Predicted values ​​of various quality indicators at each time point (in ).

[0129] Step C: Multi-indicator comprehensive evaluation and optimization decision-making:

[0130] To select the optimal process under the current state from multiple candidate solutions, this embodiment uses the entropy weight method to comprehensively evaluate and rank the prediction results from multiple dimensions, as detailed below:

[0131] Data normalization: Prediction results of m indicators based on K candidate solutions. Construct a prediction and evaluation matrix (In this embodiment, m=5). To eliminate the influence of the dimensions of each index, each column in the matrix is ​​normalized to obtain a normalized matrix Z, whose elements are... .

[0132] For positive indicators (WI, DR), the formula is used:

[0133]

[0134] For negative indicators ( (YI, SR), using the formula:

[0135]

[0136] Where k is the index of the candidate scheme ( ); j is the index of the indicator ( ); This represents the predicted value of the k-th scheme and the j-th indicator; , These are the minimum and maximum values ​​of the j-th index among all K candidate solutions, respectively. This is the normalized evaluation value.

[0137] Based on the normalized evaluation matrix, calculate the weight of the k-th scheme under the j-th indicator. :

[0138]

[0139] Information entropy of the j-th indicator And entropy weight :

[0140]

[0141]

[0142] For each candidate solution Its overall score The comprehensive score is obtained by summing the normalized index values ​​and their corresponding entropy weights:

[0143]

[0144] The rating The score reflects the quality of drying at the predicted time; the higher the score, the better the overall predicted drying quality.

[0145] Finally, the optimal process decision is executed. The system selects the combination of process parameters with the highest overall score from all candidate solutions. As the target setting value for the next adjustment cycle:

[0146]

[0147]

[0148] in, The index for the optimal solution. This corresponds to the optimal combination of process parameters.

[0149] To prevent the control system from frequently switching operating states due to minor score differences, this embodiment introduces a switching anti-shake threshold. (Preferred, (Values ​​range from 0.01 to 0.02). The difference between the score of the selected optimal solution and the score of the currently executed solution is less than this threshold. At this time, the system will maintain the current process parameters unchanged. Only when the parameters are greater than or equal to... Process switching is only performed at that time.

[0150] Step D: Control command generation and closed-loop execution, the specific steps are as follows:

[0151] Command generation and issuance: When the decision-making module determines that it needs to switch to new optimal process parameters... When updating parameters, a smooth transition strategy is adopted to generate new setpoints. :

[0152]

[0153] in, Set the value vector for the newly generated process parameters; This is the vector of process parameters currently being executed; This is the optimal process parameter vector selected in step C; for Corresponding overall score; for Corresponding overall score; This is a vector of process parameters; This is a vector of smoothing adjustment coefficients for each parameter; This is the element-wise multiplication of vectors.

[0154] The workstation will determine the optimal process parameters The command is converted into a standardized communication command frame (e.g., a frame format containing start bits, address, command, data, checksum, and end bits) and sent to the microcontroller via serial communication. A feasible command frame format is defined as follows:

[0155]

[0156] Wherein, STX: start flag; ADDR: device address; CMD: command code; DATA1~3: temperature, wind speed, and humidity settings; CHK: checksum; ETX: end flag.

[0157] Low-level closed-loop control: After receiving and verifying the instructions, the STM32 microcontroller performs precise closed-loop control on each actuator based on the real-time dry environment parameters (such as temperature, relative humidity, and wind speed) collected by the multi-source sensor module, as follows:

[0158] a) Temperature control: A two-position closed-loop control (ON / OFF control) is adopted. The microcontroller continuously monitors the real-time temperature. and the received target temperature Comparison. When When [the signal is high], a high-level signal is output to the solid-state relay to turn on the heater; when [the signal is low], [the signal is high]. When the temperature is low, a low-level signal is output to turn off the heater. This real-time feedback regulation keeps the cavity temperature fluctuating within a small range around the target value.

[0159] b) Humidity control: A closed-loop switch control with hysteresis threshold is used. When the real-time relative humidity... Below the target value ( When the hysteresis threshold is reached (to prevent frequent start-stop of the electromagnetic pump), the electromagnetic pump is started to pump the warm water in the water bath 4 to the nozzle, where it is atomized and evaporated near the heater to achieve rapid humidification; when When the humidification pump is turned off, the microcontroller turns on the exhaust centrifugal fan to expel the high-humidity air from the chamber.

[0160] c) Wind speed control: The wind speed value inside the cavity is collected in real time by wind speed sensor 6. and the target wind speed The microcontroller compares the deviations. Generate a PWM control signal to dynamically adjust the speed of the circulating fan. Less than the stabilization threshold If the current state remains unchanged, otherwise increase or decrease the speed to make convergence to .

[0161] By combining offline preparation with online control, and through the periodic execution of the online process, this embodiment can achieve dynamic, intelligent, and adaptive control of the entire drying process of floral medicinal materials, ensuring that drying efficiency is guaranteed while maximizing the optimization of various quality indicators of the final product.

[0162] The beneficial effects of this invention are as follows:

[0163] (1) This invention integrates an image acquisition module and a multi-source environmental sensing module, enabling online and lossless quantitative acquisition of multi-dimensional dynamic information such as color, shape shrinkage, drying characteristics, and environmental parameters of floral medicinal materials during the drying process. The U-Net model, which incorporates the Squeeze-and-Excitation mechanism, significantly improves the segmentation accuracy of the target region of irregular floral medicinal materials, providing a high-quality and reliable data foundation for subsequent accurate decision-making, and overcoming the shortcomings of traditional drying processes that rely on manual observation and have strong subjectivity in quality monitoring.

[0164] (2) The CNN-LSTM-Attention combined prediction model constructed in this invention can deeply explore the complex nonlinear relationships hidden in multi-source time series data. By capturing local features with CNN, modeling long-term dependencies with LSTM, and focusing on key moments with the Attention mechanism, this model can accurately predict the future evolution trend of various quality indicators during the drying process, and its prediction ability is better than traditional mechanism models or static data models.

[0165] (3) This invention establishes a complete intelligent control system from data perception and dynamic prediction to optimization decision-making and closed-loop control. Through multi-index comprehensive evaluation based on the entropy weight method, the complex and multi-objective drying process optimization problem is transformed into an optimization problem of a single comprehensive score, realizing dynamic and collaborative optimization of process parameters. This system can optimize the drying process according to the real-time status and future trends of the material itself, effectively shortening the drying cycle and reducing unit energy consumption while ensuring and improving the uniformity of product quality, and significantly improving the automation and intelligence level of flower medicinal material processing.

[0166] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0167] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An adaptive control method for the drying process of floral medicinal materials, characterized in that, include: The original image to be detected and the sensor data to be detected in the drying chamber are acquired using a preset image acquisition module and a multi-source sensor information acquisition module. The deployed image segmentation model is used to process the original image to be detected to obtain image segmentation data. The image segmentation data and the sensor data to be detected are integrated to obtain a quality state vector. The quality state vector is used to update the historical state sequence in the time series database. The updated historical state sequence is input into the deployed quality prediction model for prediction, and a quality state prediction vector is obtained under each candidate scheme condition. A prediction evaluation matrix is ​​constructed based on the quality status prediction vector, and the prediction evaluation matrix is ​​normalized to obtain a normalized evaluation matrix. The normalized evaluation matrix is ​​calculated using a preset comprehensive evaluation formula to obtain a comprehensive score; The combination of process parameters corresponding to the highest comprehensive score will be used as the target setting value for the next control cycle. The temperature, humidity, and wind speed inside the drying chamber are controlled using a smooth transition strategy based on the target set values.

2. The adaptive control method for drying floral medicinal materials according to claim 1, characterized in that, The comprehensive evaluation formula is as follows: ;in, ; ; ; The overall score is as described above; The entropy weight of the j-th index; Let be the normalized value of the k-th candidate solution on the j-th metric; The total number of evaluation indicators; Information entropy; The total number of candidate solutions; The normalized value of the k-th candidate solution on the j-th indicator is the proportion of the sum of the normalized values ​​of all candidate solutions on the j-th indicator.

3. The adaptive control method for drying floral medicinal materials according to claim 1, characterized in that, The expression for the smooth transition strategy is: ;in, Set the value vector for the newly generated process parameters; This is the vector of process parameters currently being executed; This represents the optimal process parameter vector; for The corresponding overall score; for The corresponding overall score; For smoothing adjustment coefficient vector; This represents element-wise multiplication of vectors.

4. The adaptive control method for drying floral medicinal materials according to claim 1, characterized in that, The training process of the image segmentation model includes: A drying experiment was conducted under a preset combination of process parameters, and data was collected during the experiment using the image acquisition module and the multi-source sensor information acquisition module to obtain experimental data; the experimental data included: experimental images and experimental sensor data; An improved U-Net model is obtained by embedding the SE attention mechanism module into the original U-Net model. The background and foreground of the experimentally acquired images are labeled using an image annotation tool to obtain a binary labeled image. The experimentally acquired images and the binary labeled images are then integrated to obtain an annotated image. The labeled images are subjected to data augmentation processing to obtain an augmented image set; The U-Net improved model is iteratively trained using the enhanced image set to obtain the image segmentation model.

5. The adaptive control method for drying floral medicinal materials according to claim 4, characterized in that, The training process of the quality prediction model includes: The image segmentation model is used to process the experimentally acquired images to obtain a binary mask image; The binary mask image is used to perform region segmentation on the experimentally acquired image to obtain the target ROI region; The color difference, whiteness index, yellowness index, area shrinkage ratio, and material drying rate from the experimental sensor data of the target ROI region are extracted to obtain a multidimensional feature vector; The multidimensional feature vector is divided and matched using a sliding window method to obtain the input sequence. Construct a CNN-LSTM-Attention combined prediction model; the CNN-LSTM-Attention combined prediction model includes: a CNN local feature extraction module, an LSTM long-term dependency modeling module, and a self-attention mechanism module connected in sequence; The input sequence is used to extract features using the CNN local feature extraction module to obtain a local feature map; The LSTM long-term dependency modeling module is used to extract features from the local feature map to obtain a hidden state sequence. The self-attention mechanism module is used to perform attention weighting calculation on all the hidden state sequences to obtain an attention weighting sequence. The attention-weighted sequence is aggregated in the time dimension using global average pooling to generate a context vector, and the quality parameter prediction vector is obtained by using a fully connected layer to calculate the context vector. The CNN-LSTM-Attention combined prediction model is iteratively trained using the weighted mean square error with L2 regularization based on the quality parameter prediction vector to obtain the quality prediction model.

6. The adaptive control method for drying floral medicinal materials according to claim 4, characterized in that, The data from the sensor to be tested and the data from the experimental sensor both include: material weight, temperature data, relative humidity data, and wind speed data.

7. An intelligent drying machine for floral medicinal materials, characterized in that, The method for adaptive control of the drying process of floral medicinal materials as described in claim 1 includes: an industrial camera, an LED lighting source, a drying chamber, a water bath, an electronic touch screen, a wind speed sensor, a circulating centrifugal fan, an electric heating tube, a tray support, a dehumidification outlet, a return air duct, an atomizing nozzle, a weighing sensor, a weighing tray, a temperature and humidity sensor, a turbulence fan, and a thermal insulation layer. The drying chamber is divided into two drying chambers by an internally vertically arranged partition; the thermal insulation layer covers the outer wall of the drying chamber; multiple layers of tray supports are evenly and parallelly arranged on the inner walls of both sides of the drying chamber; multiple turbulence fans are installed on the drying chamber at positions corresponding to each layer of tray supports; a circulating centrifugal fan is located at one end of the drying chamber; an electric heating element is located inside the return air duct; a water bath is connected to an atomizing nozzle; the atomizing nozzle is located inside the drying chamber; the exhaust outlet penetrates through the inner wall of the drying chamber; an industrial camera is located inside an observation window at the top of the drying chamber; multiple LED lighting sources are arranged around the lens of the industrial camera; a temperature and humidity sensor is located around the return air inlet of the drying chamber; a wind speed sensor is located inside the main air supply duct of the drying chamber; a weighing tray is located on the weighing sensor; the weighing sensor is fixed to the bottom of the drying chamber; and an electronic touch screen is integrated into the outer panel of the drying chamber. The water bath is used to pump water to the atomizing nozzle; the turbulence fan is used to generate local micro-airflow during operation; and the electric heating tube is used to heat the airflow in the return air duct.

Images