A smart suspended microwave drying system and its material drying method

Abstract

This invention pertains to an intelligent suspended microwave drying system and its material drying method within the field of drying equipment. The intelligent suspended microwave drying system includes a material conveying and collecting device, a honeycomb parallel suspended wind field generator, a three-dimensional microwave field generator and control device, a control system, and an execution module. The control system of this invention is electrically connected to a wind speed sensor and a high-speed camera via a data acquisition module to synchronously receive wind speed data and image data, which are then transmitted to the data processing module. The data processing module incorporates wind field intensity judgment and control models, wind field uniformity judgment and control models, and drying efficiency judgment and control models. It constructs a hybrid architecture combining traditional machine learning and deep learning, integrating model predictive control (MPC) to perform data acquisition, analysis, and processing. It identifies and analyzes key material characteristics, controls and drives the actions of each execution component, achieving coordinated optimization of the wind field and microwave field, and a continuous material drying process.

Description

Technical Field

[0001] This invention relates to the field of drying equipment technology, and in particular to a suspended microwave drying device that combines continuous operation and intelligent control functions, suitable for the deep processing of various materials such as grains, fruit and vegetable granules, and Chinese medicinal materials. Background Technology

[0002] Material drying is a critical processing step to ensure its storage quality and extend shelf life. Existing microwave drying equipment is mainly ordinary microwave dryers, but they generally have technical pain points: First, the drying uniformity is poor, and materials are prone to local overheating or incomplete drying, especially for granular agricultural products, where heat and airflow cannot penetrate when piled up, resulting in inconsistent product quality; Second, the drying efficiency is low, lacking real-time perception and dynamic control of the material state, and process parameters mostly rely on manual experience to set, which cannot adapt to changes in material characteristics; Third, the continuity of operation is insufficient, and most equipment requires batch feeding and shutdown for unloading, limiting production efficiency; Fourth, the control precision is low, making it difficult to adjust energy supply and airflow parameters in real time according to changes in material distribution, resulting in energy waste and a decrease in product qualification rate.

[0003] In recent years, although some suspension drying technologies have attempted to achieve uniform drying by using airflow to detach materials from the contact surface, existing equipment mostly uses a single wind field drive, lacking precise wind speed control and material state sensing mechanisms, resulting in poor suspension stability. While microwave drying technology has the advantage of rapid heating, the microwave field distribution is fixed and cannot be adapted to the dynamic distribution of materials, still exhibiting the problem of localized energy concentration. Furthermore, existing equipment has not achieved closed-loop coordination of sensing, control, and operation, making it difficult to balance drying uniformity, efficiency, and continuity, thus hindering the intelligent upgrading of the drying process.

[0004] Therefore, developing a drying equipment that integrates multiple sensing, dual-field coordinated control, and continuous operation functions to solve the technical bottlenecks of traditional equipment is of great significance for improving product processing quality, reducing energy consumption, and increasing production efficiency. It is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention addresses the problems of uneven drying and the inability to dynamically adjust based on material distribution and drying status in traditional drying equipment. It provides an intelligent suspended microwave drying system and its material drying method, which enables uniform suspension of the material within the drying chamber, uniform heating, and real-time adjustment of the suspension air field and microwave function. This achieves a closed-loop coordinated drying process of perception, dynamic control, and operation, and greatly improves drying efficiency.

[0006] To achieve the above objectives, the present invention first provides an intelligent suspended microwave drying system, characterized in that it includes: a material conveying and collecting device, comprising a feeding bin, a high-temperature resistant glass cavity, a suction assembly, a conveying pipe and a receiving bin connected in sequence, for realizing continuous conveying of materials from feeding, drying to finished product collection.

[0007] The honeycomb parallel suspension wind field generator includes: a honeycomb flow guide grid and a supporting cavity connected sequentially to the lower side of a high-temperature resistant glass cavity; an air duct assembly and multiple sets of blower motors are provided on the lower side of the supporting cavity; the air duct assembly is connected to the supporting cavity; multiple sets of wind speed sensors are distributed in a ring at equal angles on the inner wall of the supporting cavity; each wind speed sensor corresponds to and is electrically connected to a blower motor; the supporting cavity is used to achieve outward expansion of the incoming airflow to cover all the grid holes of the honeycomb flow guide grid; the honeycomb flow guide grid is used to regulate the airflow; the wind speed sensors are used to collect wind speed data in the corresponding area in real time to drive the corresponding blower motor to accurately adjust the wind speed, so that the material is stably suspended in the high-temperature resistant glass cavity; a three-dimensional microwave field generator and control device includes the outer periphery of the high-temperature resistant glass cavity. Several groups of microwave generating components are arranged at equal angles to the walls. Each group of microwave generating components includes a microwave generator and a sliding mechanism that drives the microwave generator to reciprocate along the axial direction of the high-temperature resistant glass cavity. The microwave generating components are connected to a power adjustment module. Three high-speed cameras are used to monitor the interior of the high-temperature resistant glass cavity in real time to capture the material suspension distribution. The control system includes a data acquisition module, a data processing module, and an execution control module. The data acquisition module is electrically connected to the wind speed sensor and the high-speed cameras to synchronously receive wind speed data and image data and transmit them to the data processing module. The data processing module has built-in wind field intensity judgment and control models, wind field uniformity judgment and control models, and drying efficiency judgment and control models, constructing a traditional machine learning + The system employs a hybrid architecture based on deep learning and integrates Model Predictive Control (MPC). It analyzes and processes multi-source data to identify and analyze key features such as material height, distribution uniformity, color, and aggregation state. The execution control module is electrically connected to the adjustment modules of the motor-driven air supply component, microwave generator component, material suction component, and material receiving hopper. Based on the data processing results, it generates control commands to drive the actions of each execution component, thereby achieving coordinated optimization of the air field and microwave field.

[0008] To facilitate the smooth inflow and outflow of materials, the lower outer periphery of the high-temperature resistant glass cavity is provided with a feed inlet, and the feed bin is connected to the feed inlet. The feed bin is shaped like an upward-sloping bucket. The upper end of the high-temperature resistant glass cavity is provided with a top cover assembly, and the suction assembly is located on the lower side of the top cover assembly and is connected to the receiving bin via a conveying pipe.

[0009] To facilitate intelligent drying, the data processing module includes the following control model:

[0010] The first control model, the wind field intensity judgment and control model, takes the precise mapping of "pixel coordinates → physical height" as its core, integrates traditional computer vision algorithms and lightweight deep learning detection models, and combines MPC to realize real-time judgment and dynamic control of the height of upper materials;

[0011] The second control model, the wind field uniformity judgment and control model, takes "bottom material flatness characteristics → wind field uniformity mapping" as its core. It integrates traditional computer vision quantitative analysis and lightweight deep learning semantic segmentation, and combines MPC and wind speed sensor data linkage to achieve real-time judgment and precise control of wind field uniformity.

[0012] The third control model, the drying efficiency judgment and control model, adopts a hybrid architecture of "random forest + gradient boosting tree (XGBoost / LightGBM) + MobileNetV3 + multi-task head". Combined with the model prediction control MPC control module, dynamic control is achieved. The traditional machine learning branch is responsible for processing manually extracted quantitative features to ensure inference speed and interpretability. The deep learning branch is responsible for automatically mining deep image features to improve the generalization ability of complex scenes. After the feature fusion of the two branches, the optimal adjustment instructions for microwave power and wind speed are output through MPC rolling optimization, perfectly adapting to the closed-loop drying link of the equipment of "sensing-analysis-control-feedback".

[0013] Furthermore, the regulation process of the first-mode model includes:

[0014] S1.1 Image Preprocessing and Data Acquisition: Receive material images of the upper part of the cavity from various high-speed cameras, focusing on a specific core area (0-10cm) below the material suction component; perform distortion correction using the Zhang Zhengyou calibration method, and establish a "pixel coordinates - physical height" mapping matrix; use median filtering for noise reduction, segment the material area based on HSV color threshold, and enhance edge features through grayscale conversion and histogram equalization;

[0015] S1.2 Height Detection and Threshold Determination: Canny edge detection is performed on the preprocessed image to extract the upper edge contour of the material. The coordinates of the highest pixel point are fitted by Hough linear transformation and converted into the actual physical height Hreal(t). The YOLOv8n model is used to train the key point detection model of the upper edge of the material and the height value is corrected. The effective adsorption height range Htarget±ΔH of the suction component is pre-calibrated, where Htarget is the physical height of the lower edge of the suction component, i.e., the core threshold, and ΔH is the allowable deviation, which is 5-10 mm, and is dynamically updated based on the historical data of material height-adsorption success rate.

[0016] S1.3 Model Predictive Control (MPC): A lightweight LSTM model is used to predict the material height for the next 5 control cycles, aiming to keep the material height stable within the target range, minimizing height deviation and control action amplitude. The lightweight LSTM model is as follows:

[0017] ;

[0018] In the formula, P is the prediction time domain, which takes the value of 5 control cycles (adapting to the dynamic change characteristics of material height).

[0019] M represents the control time domain, taking values ​​for two control cycles (balancing the timeliness and stability of regulation).

[0020] ΔU is the change in control quantity, microwave power adjustment ΔP (range -1.2 to 1.8 kW), n sets of wind speed adjustment Δv1-Δvn (range -0.3 to 0.5 m / s), and suction component negative pressure adjustment ΔF (range -5 to 8 kPa). Hpred is the future material height predicted by the LSTM model, with a weighting coefficient λ=0.15 to avoid material suspension instability. n refers to the number of blower motors and wind speed sensors.

[0021] Furthermore, the regulation process of the second regulation model includes:

[0022] S2.1 Image Preprocessing and Data Acquisition: Receives images of the material layer at the bottom of the cavity from a high-speed camera, and simultaneously receives data from various wind speed sensors to form a dual-source input; employs median filtering + bilateral filtering for noise reduction, histogram equalization to enhance contrast, and HSV threshold segmentation and U-Net lightweight semantic segmentation to accurately extract the material region;

[0023] S2.2 Flatness Feature Extraction and Quantification: Four core indicators were extracted: ① Particle distribution uniformity index; ② Area ratio of the accumulation region; ③ Edge contour regularity index; ④ Peak height difference index.

[0024] Among them, ① the particle distribution uniformity index extraction method is as follows: calculate the variance σ² of pixel density in the material area - when the wind field is uniform, the material distribution is dense and consistent, the variance value is small, and the preset standard range is σ²≤0.05; when the wind field is uneven, local accumulation leads to large density differences, and the variance value exceeds the standard; ② accumulation area quantification index: detect the protruding or concave areas in the material layer, and define the height difference >3mm as accumulation / concavity, calculate the area ratio of the accumulation area S_accum / S_total as the standard threshold ≤5%, the higher the ratio, the more serious the wind field unevenness; ③ edge contour regularity index: extract the edge contour of the material layer, calculate the smoothness of the contour (contour length / contour enclosed area ratio), when the smoothness is ≤1.2, the contour is smooth and regular when the wind field is uniform, otherwise, the contour is obviously jagged when the wind field is uneven; ④ height difference peak index: based on the pixel-physical mapping relationship, calculate the physical height difference Δh between the highest pixel point and the lowest pixel point in the material layer, the standard threshold Δh≤10mm, if it exceeds, it is judged as local wind speed abnormality;

[0025] Deep learning-assisted feature optimization: MobileNetV3-Small is used as the backbone network. The preprocessed bottom material layer image is input to extract a 256-dimensional deep feature vector, which is used to correct the error of traditional quantification indicators (such as density variance deviation caused by material particle size differences) and improve feature robustness.

[0026] Feature fusion: The four quantified core indicators are concatenated with the deep feature vector of deep learning, and after Min-Max normalization, a 32-dimensional comprehensive flatness feature vector is formed, which serves as the core input for uniformity determination.

[0027] S2.3 Wind Field Uniformity Determination: Input the comprehensive feature vector into the "Random Forest + LightGBM" ensemble classification model, output the preliminary determination result, and output the wind field uniformity determination result: 0 = uniform, 1 = slightly uneven, 2 = severely uneven; Combined with the wind speed standard deviation, when Δv≤0.3m / s, cross-validation is performed to locate abnormal areas, and the determination result and the corresponding abnormal air supply area are output.

[0028] MPC Control Module: Communicatively connected to the uniformity determination module and wind speed sensor, it uses an LSTM model to predict the smoothness index for the next 6 control cycles. The target smoothness vector is σ²=0.03, S_accum / S_total=3%, smoothness=1.0, and Δh=3mm. The output wind speed adjustment Δvn is constrained to -0.3~0.5m / s, with a weighting coefficient λ=0.25. The rate of change of wind speed adjustment between adjacent control cycles is constrained to ≤20%. The LSTM model is as follows:

[0029] P represents the prediction time domain, with a value of 6 control cycles to adapt to the dynamic response characteristics of the wind field. After the wind field is adjusted, it takes 3-5 cycles to stabilize.

[0030] M represents the control time domain, and its value is set to two control cycles to balance the timeliness of regulation with the stability of the wind field.

[0031] Δvn is the wind speed adjustment of the nth group of motor air supply components, with a value range of 1-7 and a unit of m / s, which is the change in the control quantity;

[0032] Φtarget is the target smoothness index vector ([σ²=0.03, S_accum / S_total=3%, smoothness=1.0, Δh=3mm]), representing the ideal smoothness state when the wind field is uniform;

[0033] Φpred is the vector of future smoothness indices predicted by the LSTM model;

[0034] λ is a weighting coefficient with a value of 0.25, used to balance the accuracy of the flatness regression and the wind speed adjustment range, and to avoid sudden changes in wind speed.

[0035] K represents the current time domain;

[0036] Constraints: Δvn∈[−0.3,0.5]m / s to match the speed regulation capability of the motor air supply component; the rate of change of wind speed adjustment between adjacent control cycles ≤20% to avoid drastic fluctuations in the wind field; the standard deviation of n sets of wind speeds Δv≤0.3m / s for overall wind field uniformity constraints.

[0037] Furthermore, the regulation process of the third regulation model includes:

[0038] S3.1 Image Preprocessing and Data Acquisition Module: Receives real-time images from a high-speed camera, simultaneously acquires wind speed, microwave power, and measured moisture content data, and aligns them precisely with timestamps; employs median filtering + Gaussian filtering for noise reduction, converts RGB to HSV color space, and performs data enhancement processing through rotation and scaling;

[0039] S3.2 Feature Extraction and Fusion: The traditional machine learning branch extracts HSV color features, aggregation state features, and process-related features of the raw materials, forming a 30-35 dimensional handcrafted feature vector; the deep learning branch uses MobileNetV3 Small to extract 256 dimensional deep features, and configures a multi-task head to output the predicted moisture content, aggregation level, and aggregation region coordinates; the dual-branch features are fused into a 512 dimensional comprehensive feature vector through "splicing + attention weighting";

[0040] S3.3 Hybrid Prediction Model: The traditional machine learning branch constructs an ensemble model of "Random Forest + XGBoost + LightGBM" to predict drying efficiency Δη / Δt and clustering level; the deep learning branch uses transfer learning for fine-tuning, with a multi-task loss function = 0.5 × moisture content regression loss + 0.3 × clustering level classification loss + 0.2 × clustering region detection loss; the weighted fusion weights of the two branches are: traditional branch weight 0.4, deep learning branch weight 0.6.

[0041] MPC Control Module: Predicts the drying efficiency for the next 6 control cycles, with a target drying efficiency ηtarget (0.3-1.2% / s) as the objective. It outputs microwave power adjustment ΔP and wind speed adjustment Δv1-Δvn, with a weighting coefficient λ=0.2, and constrains the rate of change of control quantities between adjacent control cycles to ≤30%. Each control cycle solves the following optimization problem:

[0042]

[0043] P represents the prediction time domain, with a value taken over 6 control cycles to adapt to the dynamic response characteristics of the drying process.

[0044] M represents the control time domain, taking values ​​for three control cycles to balance the timeliness and stability of regulation.

[0045] ΔU is the change in control quantity, including: microwave power adjustment ΔP, unit: kW; 7 sets of wind speed adjustment Δv1−Δv7, unit: m / s;

[0046] ηtarget is the target drying efficiency, with a preset range of 0.3-1.2% / s, which can be adjusted according to the material type;

[0047] ηpred represents the future drying efficiency predicted by the hybrid model;

[0048] λ is a weighting coefficient (with a value of 0.2), used to balance the accuracy of drying efficiency tracking with the smoothness of control actions;

[0049] Constraints: ΔP∈[−1.2,1.8]kW, used to match the power regulation module capability of the microwave generator; Δvi∈[−0.4,0.6]m / s, used to ensure stable suspension of materials in the cavity; the rate of change of control quantity between adjacent control cycles ≤30% to avoid sudden changes.

[0050] The intelligent suspended microwave drying system of this invention integrates a continuous material conveying, drying, and collection device, a honeycomb parallel suspended airflow generator, a three-dimensional microwave field generator and control device, a multi-sensor system, and further forms a closed-loop intelligent drying link through an intelligent airflow and microwave field control system module. The intelligent control model judges the airflow intensity by visually identifying whether the height of the upper material layer is level with the collection device, judges the airflow uniformity by the flatness of the bottom material layer, and judges the drying efficiency by the material color and whether the material is aggregated. The three intelligent control models (airflow intensity judgment and control, airflow uniformity judgment and control, and drying efficiency judgment and control) are all based on a hybrid architecture of "traditional machine learning + deep learning" and integrate model predictive control (MPC). They follow a unified logic of "data acquisition - preprocessing - feature engineering - hybrid training - actual calibration - closed-loop control" and are adapted to the closed-loop drying link of intelligent sensing dual-field collaborative suspended microwave drying equipment. Through the above-mentioned multi-sensor linkage, the microwave and suspended air field parameters are monitored and adjusted in real time according to the degree of material drying during the microwave drying process, so as to realize intelligent and continuous drying, improve drying efficiency, and be applicable to intelligent drying of various materials such as grains, fruit and vegetable granules, and Chinese medicinal materials.

[0051] The present invention also provides a material drying method for the above-mentioned intelligent suspension microwave drying system, specifically including the following steps:

[0052] Step 1, System Start-up and Parameter Setting: Set the initial wind speed, initial microwave power, target drying efficiency, and effective adsorption height range of the suction component according to the type of material to be dried;

[0053] Step 2, Intelligent wind field pre-construction: Start each group of air supply motors, and the airflow enters the support cavity through the air duct assembly. Through jet expansion, the incoming airflow covers the honeycomb guide grid, and the grid is regulated to form the initial rising airflow. The wind speed sensor collects data in real time. The control system judges the wind field uniformity based on the wind speed sensor data of each group and drives the wind supply motor to adjust the speed until the standard deviation of the wind speed of each group is ≤0.3m / s, and the wind field reaches a stable state.

[0054] Step 3, Gradual Feeding and Dynamic Stabilization of the Airflow: The material is gradually released from the feed hopper. With the help of the upward-sloping bucket-shaped inlet of the feed hopper and the negative pressure effect inside the cavity, the material enters the cavity evenly. The wind speed sensor continuously collects data, and the high-speed camera captures images of the bottom material layer. The control system uses the airflow uniformity judgment and control model to fine-tune the speed of the corresponding blower motor in real time to maintain the flatness index of the material layer to meet the following requirements: σ²≤0.05, the area ratio of the accumulation area≤5%, the smoothness≤1.2, and the height difference Δh≤10mm.

[0055] Step 4, Intelligent Microwave Drying Operation: A high-speed camera captures real-time images of the material's suspension distribution, color, and aggregation state. The image data is pre-processed and then transmitted to the control system. The data processing module analyzes the material's moisture content changes and aggregation level using a drying efficiency judgment and control model, generating control commands. The execution control module drives the microwave generator component, adjusting the axial position and output power of the microwave generator (701) to achieve precise 360° microwave irradiation of the material without dead angles. Simultaneously, the wind field intensity judgment and control model maintains the upper layer of material at a stable height within the effective adsorption range of the suction component.

[0056] Step 5, Drying material collection and continuous discharge: The dried material rises to the height of the suction component, the suction component adsorbs the material under negative pressure, and the material is transferred to the receiving hopper through the conveying pipe. The discharge valve of the receiving hopper is dynamically opened to achieve continuous discharge.

[0057] Step 6, Closed-loop control and system shutdown: The control system continuously receives wind speed and image data synchronously, and dynamically adjusts the wind field and microwave parameters through three intelligent control models; when the material in the feed hopper is exhausted or the receiving hopper is full, the drying ends and the system shuts down.

[0058] Furthermore, in step 1, the preset effective adsorption height range of the suction component is 0-10cm.

[0059] In step 4, during the intelligent microwave drying operation, the image processing algorithm identifies the material color, aggregation area, and density, and links the microwave generating component to achieve dynamic adjustment of power and position.

[0060] Furthermore, the real-time image data from the high-speed camera is received at a frequency of 1 frame per second.

[0061] Compared with existing technologies, the intelligent suspended microwave drying system and its material drying method of the present invention have the following advantages: 1. High precision in airflow control, solving the problem of poor suspension stability: Through the one-to-one design of multiple independent motor air supply components and wind speed sensors, combined with the air jet expansion effect of the supporting cavity and the rectification effect of the honeycomb guide grid, the airflow can be precisely controlled, ensuring stable suspension of materials in the cavity, avoiding the problems of material accumulation or uneven suspension caused by traditional single airflow, and improving drying uniformity; 2. Intelligent adaptation of microwave drying, high energy utilization: Relying on the real-time shooting and image processing algorithms of multiple high-speed cameras, the material distribution area and density are accurately identified, and the power and position of the microwave generating components are dynamically adjusted, solving the problem of... The traditional fixed microwave field causes local energy concentration or irradiation dead zones, significantly improving microwave energy utilization: 3. Strong continuous operation capability and improved production efficiency: Through the collaborative design of the material suction component of the continuous material conveying and collection device, the closed conveying pipe and the funnel-type receiving hopper, the entire process from feeding, drying to discharging can be carried out without downtime, solving the efficiency bottleneck of batch operation of traditional equipment. At the same time, the design of the guide stator and the petal-shaped cutting material hopper ensures smooth material suction and reduces material residue: 4. High degree of intelligent closed-loop control: By integrating multi-source data such as wind speed and images through the control system, a full-link closed-loop control of "sensing-analysis-control-feedback" is formed. It can adapt to changes in material characteristics without manual intervention, reducing the difficulty of operation and ensuring the consistency of product quality. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the intelligent suspension microwave drying system of the present invention.

[0063] Figure 2 This is a schematic diagram of the air duct assembly.

[0064] Figure 3 This is a schematic diagram of a honeycomb airflow guide grille.

[0065] The components include: 1. Blower motor; 2. Air duct assembly; 201 Air inlet duct; 202 Combination plate; 3. Support cavity; 4. Honeycomb guide grid; 401 Grid hole; 5. Feed hopper; 6. High-temperature resistant glass cavity; 7. Microwave generator assembly; 8. Top cover assembly; 9. Suction assembly; 10. Conveying pipe; 11. Receiving hopper; 12. High-speed camera. Detailed Implementation

[0066] The intelligent suspended microwave drying system and drying method of the present invention will be described in detail below with reference to the accompanying drawings.

[0067] Example 1

[0068] like Figure 1 — Figure 3As shown, the intelligent suspended microwave drying system of this embodiment includes the following components: a material conveying and collecting device, specifically including a feeding bin 5, a high-temperature resistant glass cavity 6, a suction assembly 9, a conveying pipe 10, and a receiving bin 11 connected in sequence, used to realize the continuous conveying of materials from feeding, drying to finished product collection.

[0069] To facilitate the suspension and distribution of materials during the drying process, a honeycomb parallel suspension airflow generator is installed at the lower air inlet direction of the high-temperature resistant glass cavity 6, including: a honeycomb guide grid 4 and a supporting cavity 3 connected in sequence on the lower side of the high-temperature resistant glass cavity 6, such as... Figure 3 As shown, the grille holes 401 of the honeycomb airflow guide grille 4 are regular polygonal holes; they can evenly and regularly regulate the incoming airflow; the lower side of the supporting cavity 3 is provided with an air duct assembly 2 for even airflow and multiple sets of blower motors 1, and the air duct assembly 2 is connected to the supporting cavity 3; as shown Figure 2 As shown, the air duct assembly includes a combination plate 202 connected to the lower end of the support cavity 3. Several air inlets are evenly distributed on the surface of the combination plate 202. Each air inlet is connected to an air inlet duct 201 and a blower motor 1. The connection section between the air inlet duct 201 and the air inlet is a fish-belly shaped bend. In this embodiment, the air inlets, air inlet ducts 201 and blower motor are all one-to-one, and a total of seven sets are provided. The air inlets are evenly distributed on the surface of the combination plate to achieve uniform air intake. To facilitate comprehensive monitoring of the air intake, multiple sets of wind speed sensors are distributed in a ring at equal angles on the inner wall of the supporting cavity 3. Each set of wind speed sensors corresponds one-to-one with and is electrically connected to the uniform air supply motor 1, allowing the control system to monitor and adjust the air supply motor volume simultaneously. In the above structure, the supporting cavity 3 is used to spray the air supply air from each air duct of the air duct assembly outwards and covers all the grid holes of the honeycomb guide grid. The honeycomb guide grid is used to regulate the airflow, and the wind speed sensors are used to collect wind speed data in the corresponding area in real time to accurately adjust the wind speed of the corresponding air supply motor 1, so as to ensure the stable suspension of the material in the high-temperature resistant glass cavity 6. To facilitate material drying, a... The three-dimensional microwave field generating and control device includes several sets of microwave generating components 7 arranged at equal angles on the outer peripheral wall of a high-temperature resistant glass cavity 6. Each set of microwave generating components 7 includes a microwave generator and a sliding mechanism that drives the microwave generator to reciprocate along the axial direction of the high-temperature resistant glass cavity. To facilitate microwave power adjustment, the microwave generating components 7 are connected to a power adjustment module. To facilitate real-time imaging of the material suspension distribution, three full-view monitoring high-speed cameras 12 are installed on the inner side wall of the high-temperature resistant glass cavity 6. The three high-speed cameras are spaced 120° apart and positioned at the same height from top to bottom to achieve full-view monitoring within the cavity. In addition, a feed inlet is provided on the lower outer periphery of the high-temperature resistant glass cavity 6, and a feed bin 5 is connected to the feed inlet. To facilitate material conveying from the feed inlet into the cavity, the feed bin 5 is shaped like an upward-sloping bucket. A top cover assembly 8 is provided at the upper end of the high-temperature resistant glass cavity 6, and a suction assembly 9 is located below the top cover assembly 8 and is connected to the receiving bin via a conveying pipe.

[0070] To achieve intelligent control of drying process parameters during the drying process, the microwave drying system of this invention also includes a data acquisition module, a data processing module, and an execution control module. The data acquisition module is electrically connected to a wind speed sensor and a high-speed camera 12, respectively, to synchronously receive wind speed data and image data, and transmit them to the data processing module. The data processing module has built-in wind field intensity judgment and control models, wind field uniformity judgment and control models, and drying efficiency judgment and control models. It constructs a hybrid architecture of traditional machine learning + deep learning and integrates model predictive control (MPC). It analyzes and processes the collected multi-source data, identifies key features such as material height, distribution uniformity, color, and aggregation state, and performs analysis and processing. The execution control module is electrically connected to the adjustment modules of the motor-driven air supply component, microwave generator component 7, material suction component 9, and receiving bin 11, respectively. It generates control commands based on the data processing results, drives the actions of each execution component, and achieves coordinated optimization of the wind field and microwave field operating parameters.

[0071] The data processing module in this embodiment includes three control models. The specific functions and control processes of each control model are described in detail below.

[0072] The first control model, the wind field intensity judgment and control model, takes the precise mapping of "pixel coordinates → physical height" as its core, integrates traditional computer vision algorithms and lightweight deep learning detection models, and combines MPC to realize real-time judgment and dynamic control of the height of upper materials.

[0073] The regulation process of the first regulation model mentioned above includes:

[0074] S1.1 Image Preprocessing and Data Acquisition: Images of the material above the cavity captured by each high-speed camera 12 are received at a receiving frequency of 1 frame per second, focusing on the core area of ​​0-10cm below the suction component 9; distortion correction is performed using the Zhang Zhengyou calibration method to eliminate the influence of lens distortion on height measurement; a "pixel coordinate - physical height" mapping matrix is ​​established based on the preset physical calibration points inside the cavity (such as the lower edge of the suction component and the scale on the side wall of the cavity); in image preprocessing, median filtering is used to remove image noise caused by microwave interference, and the material area is extracted by color threshold segmentation (based on the HSV color difference between the material and the cavity background) to remove background interference; the image is grayscaled and histogram equalized to enhance the edge features of the material, and the edge features are enhanced by grayscaled and histogram equalized based on the HSV extraction of the material area;

[0075] S1.2 Height Detection and Threshold Judgment: Perform Canny edge detection on the preprocessed image to extract the upper edge contour of the material. Fit the coordinates of the highest pixel points through the Hough line transformation, and convert them into the actual physical height Hreal(t) in combination with the calibrated mapping matrix. Train a key point detection model for the upper edge of the material using the YOLOv8n model to correct the height value. Pre-calibrate the effective adsorption height range of the material suction component as Htarget±ΔH, where Htarget is the physical height of the lower edge of the material suction component, that is, the core threshold, and ΔH is the allowable deviation, with a value range of 5 - 10 mm, which can be adjusted specifically according to the material category. For example, for grains, ΔH = 5 mm, and for fruit and vegetable particles, ΔH = 8 mm. Dynamically update the threshold based on the historical operation data of the equipment (material height - adsorption success rate): when the adsorption success rate < 95%, automatically fine-tune ΔH (±1 mm) to adapt to the suspension characteristics of different materials. If Hreal(t) ∈ [Htarget−ΔH, Htarget+ΔH], it is judged as "flush", and the judgment result "qualified" is output; if Hreal(t) < Htarget−ΔH, it is judged as "insufficient height", and the judgment result "unqualified" is output, and then the intelligent control wind field intensity jumps upward; if Hreal(t) > Htarget+ΔH, it is judged as "excessive height", and the judgment result "overlimit" is output, and then the intelligent control wind field intensity jumps downward;

[0076] S1.3 Model Predictive Control (MPC) Regulation: Use the lightweight LSTM model to predict the material height in the next 5 control cycles. With the goal of "stabilizing the material height within the flush judgment threshold range", minimize the height deviation and the amplitude of the control action. The lightweight LSTM model is as follows:

[0077] ;

[0078] P is the prediction horizon, with a value of 5 control cycles (to adapt to the dynamic change characteristics of the material height);

[0079] M is the control horizon, with a value of 2 control cycles (to balance the timeliness and stability of regulation);

[0080] ΔU is the change in the control quantity, the adjustment amount of the microwave power ΔP, with a value range of -1.2~1.8 kW), 7 groups of wind speed adjustment amounts Δv1 - Δv7, with a value range of -0.3~0.5 m / s, and the adjustment amount of the negative pressure of the material suction component ΔF with a value range of -5~8 kPa, Hpred is the future material height predicted by the LSTM model, the weight coefficient λ = 0.15, to avoid the instability of material suspension, and n refers to the number of air supply motors and wind speed sensors.

[0081] The second control model is a wind field uniformity judgment and control model based on the flatness of the bottom material layer. With "bottom material flatness characteristics → wind field uniformity mapping" as the core, it integrates traditional computer vision quantitative analysis and lightweight deep learning semantic segmentation. Combined with model predictive control (MPC) and the linkage of data from various wind speed sensors, it realizes real-time judgment and precise control of wind field uniformity. The model balances detection real-time performance (inference delay ≤60ms) and control accuracy (wind speed adjustment error ±0.1m / s). By optimizing the wind field distribution, it ensures stable material suspension and solves the problems of local accumulation or uneven suspension.

[0082] The specific regulation process of the aforementioned second regulation model includes:

[0083] S2.1 Image Preprocessing and Data Acquisition: Images of the material layer at the bottom of the cavity acquired by each group of high-speed cameras are received, along with data from each group of wind speed sensors, forming a dual-source input. Median filtering and bilateral filtering are used for noise reduction, histogram equalization enhances contrast, and HSV threshold segmentation and U-Net lightweight semantic segmentation are used to accurately extract the material region. The specific processing steps are as follows:

[0084] It receives image data of the material layer at the bottom of the cavity from a high-speed camera. The camera's focusing area covers the core area at the bottom of the cavity to ensure that dynamic changes in the material layer are captured. Simultaneously, it receives real-time wind speed data (v1-v7) from each set of wind speed sensors (distributed in a ring at equal angles along the inner wall of the supporting cavity), forming a dual-source input of "image data + sensor data".

[0085] Calibration and distortion correction: Zhang Zhengyou's calibration method is used to eliminate lens distortion. Based on the preset physical calibration points at the bottom of the cavity, a mapping relationship between image pixels and physical space is established to ensure the accuracy of material area size measurement.

[0086] Denoising and Enhancement: A combination of median filtering and bilateral filtering is used to remove image noise caused by microwave interference and airflow fluctuations; histogram equalization is used to enhance the contrast between the material and the background, highlighting the bottom outline of the material; Material Region Segmentation: Based on the combination of HSV color space threshold segmentation and U-Net lightweight semantic segmentation model, the bottom material layer region is accurately extracted, background interference such as cavity walls and flow guide grilles is removed, and a binarized material region image is output;

[0087] S2.2 Flatness Feature Extraction and Quantification: Four core indicators were extracted: ① Particle distribution uniformity index; ② Area ratio of the accumulation region; ③ Edge contour regularity index; ④ Peak height difference index.

[0088] Among them, ① the particle distribution uniformity index extraction method is as follows: calculate the variance σ² of pixel density in the material area - when the wind field is uniform, the material distribution is dense and consistent, the variance value is small, and the preset standard range is σ²≤0.05; when the wind field is uneven, local accumulation leads to large density differences, and the variance value exceeds the standard; ② accumulation area quantification index: detect the protruding or concave areas in the material layer, and define the height difference >3mm as accumulation / concavity, calculate the area ratio of the accumulation area S_accum / S_total as the standard threshold ≤5%, the higher the ratio, the more serious the wind field unevenness; ③ edge contour regularity index: extract the edge contour of the material layer, calculate the smoothness of the contour (contour length / contour enclosed area ratio), when the smoothness is ≤1.2, the contour is smooth and regular when the wind field is uniform, otherwise, the contour is obviously jagged when the wind field is uneven; ④ height difference peak index: based on the pixel-physical mapping relationship, calculate the physical height difference Δh between the highest pixel point and the lowest pixel point in the material layer, the standard threshold Δh≤10mm, if it exceeds, it is judged as local wind speed abnormality;

[0089] Deep learning-assisted feature optimization: MobileNetV3-Small is used as the backbone network. The preprocessed bottom material layer image is input to extract a 256-dimensional deep feature vector, which is used to correct the error of traditional quantification indicators (such as density variance deviation caused by material particle size differences) and improve feature robustness.

[0090] Feature fusion: The four quantified core indicators are concatenated with the deep feature vector of deep learning, and after Min-Max normalization, a 32-dimensional comprehensive flatness feature vector is formed, which serves as the core input for uniformity determination.

[0091] S2.3 Wind Field Uniformity Determination: Input the comprehensive feature vector into the "Random Forest + LightGBM" ensemble classification model, output the preliminary determination result, and output the wind field uniformity determination result: 0 = uniform, 1 = slightly uneven, 2 = severely uneven; Combined with the wind speed standard deviation, when Δv≤0.3m / s, cross-validation is performed to locate abnormal areas, and the determination result and the corresponding abnormal air supply area are output.

[0092] MPC Control Module: Communicatively connected to the uniformity determination module and wind speed sensor, it uses an LSTM model to predict the smoothness index for the next 6 control cycles. The target smoothness vector is σ²=0.03, S_accum / S_total=3%, smoothness=1.0, and Δh=3mm. The output wind speed adjustment Δvn is constrained to -0.3~0.5m / s, with a weighting coefficient λ=0.25. The rate of change of wind speed adjustment between adjacent control cycles is constrained to ≤20%. The LSTM model is as follows:

[0093] ;

[0094] P represents the prediction time domain, with a value of 6 control cycles used to adapt to the dynamic response characteristics of the wind field. After the wind field is adjusted, it takes 3-5 cycles to stabilize.

[0095] M represents the control time domain, and its value is set to two control cycles to balance the timeliness of regulation with the stability of the wind field.

[0096] Δvn is the wind speed adjustment of the nth group of motor air supply components, with a value range of 1-7 and a unit of m / s, which is the change in the control quantity;

[0097] Φtarget is the target smoothness index vector ([σ²=0.03, S_accum / S_total=3%, smoothness=1.0, Δh=3mm]), representing the ideal smoothness state when the wind field is uniform;

[0098] Φpred is the vector of future smoothness indices predicted by the LSTM model;

[0099] λ is a weighting coefficient with a value of 0.25, used to balance the accuracy of the flatness regression and the wind speed adjustment range, and to avoid sudden changes in wind speed.

[0100] Constraints: Δvn∈[−0.3,0.5]m / s to match the speed regulation capability of the motor air supply component; the rate of change of wind speed adjustment between adjacent control cycles ≤20% to avoid drastic fluctuations in the wind field; the standard deviation of n sets of wind speeds Δv≤0.3m / s for overall wind field uniformity constraints.

[0101] The third control model, the drying efficiency judgment and control model, adopts a hybrid architecture of "random forest + gradient boosting tree (XGBoost / LightGBM) + MobileNetV3 + multi-task head". Combined with the model prediction control MPC control module, dynamic control is achieved. The traditional machine learning branch is responsible for processing manually extracted quantitative features to ensure inference speed and interpretability. The deep learning branch is responsible for automatically mining deep image features to improve the generalization ability of complex scenes. After the feature fusion of the two branches, the optimal adjustment instructions for microwave power and wind speed are output through MPC rolling optimization, perfectly adapting to the closed-loop drying link of the equipment of "sensing-analysis-control-feedback".

[0102] The regulation process of the third regulation model includes:

[0103] S3.1 Image Preprocessing and Data Acquisition Module: Receives real-time images from high-speed cameras at a rate of 1 frame per second. Image data collected by multiple high-speed cameras 12 at different locations covers the entire material area within the high-temperature resistant plexiglass cavity 6. Simultaneously, it acquires data from various wind speed sensors (v1-v7), microwave power (P), and measured moisture content data (collected in real-time by a moisture analyzer), forming a multi-source data set. Image preprocessing: A combination of median filtering and Gaussian filtering is used for noise reduction to eliminate image noise caused by microwave interference. RGB images are converted to HSV color space (resistant to light fluctuations). Semantic segmentation (U-Net lightweight version) separates the material area from the cavity and airflow background. Data enhancement is performed on the images through rotation, scaling, and brightness fine-tuning to expand sample diversity.

[0104] Data synchronization and alignment: Based on timestamps, image data (material color, aggregation state), process parameter data (wind speed, microwave power), and measured moisture content data are precisely aligned to form a time-series dataset; S3.2 Feature extraction and fusion: In the traditional machine learning branch, material HSV color features, aggregation state features, and process-related features are extracted to form a 30-35 dimensional handmade feature vector; the deep learning branch uses MobileNetV3 Small to extract 256 dimensional deep features, and configures multi-task heads to output predicted moisture content, aggregation level, and aggregation region coordinates; the dual-branch features are fused into a 512-dimensional comprehensive feature vector through "splicing + attention weighting";

[0105] Specifically, the traditional machine learning branch includes: ① Color features: extracting the mean (H_mean, S_mean, V_mean), variance (H_var, S_var, V_var), peak position of the V channel histogram, and color change rate (ΔV_mean / Δt) of the material region in HSV space, quantifying the moisture content correlation characteristics; ② Aggregation state features: calculating the material stacking density (material pixel ratio), aggregation region contour complexity (contour perimeter / area), centroid distribution variance, and particle spacing mean, simultaneously locating the coordinates of the aggregation region, and quantifying the sufficiency of microwave and airflow contact; ③ Process correlation features: concatenating real-time wind speed (v1-v7) and microwave power (P) with the above features to form a 30-35 dimensional handmade feature vector, which is then normalized using Min-Max to eliminate dimensional differences.

[0106] Deep learning branch: The lightweight MobileNetV3 (Small version) is used as the backbone network. The preprocessed material image (640×480 pixels) is input, and deep features of the image are extracted through depthwise separable convolution, and a 256-dimensional high-dimensional feature vector is output. Multiple task heads are configured: ① Regression head 1: outputs the predicted value of material moisture content; ② Classification head 1: outputs the clustering level (0=uniform, 1=slight, 2=severe); ③ Detection head: outputs the coordinates of the boundary box of the clustering area, providing a location basis for precise control.

[0107] Dual-branch feature fusion: The "spoofing + attention weighting" method is used to fuse the handmade feature vectors with the deep feature vectors output by MobileNetV3 to generate a 512-dimensional comprehensive feature vector, which retains the physical meaning of the quantized features and incorporates deep image correlation information.

[0108] S3.3 Hybrid Prediction Model: The traditional machine learning branch constructs an ensemble model of "Random Forest + XGBoost + LightGBM" to predict drying efficiency Δη / Δt and clustering level; the deep learning branch uses transfer learning for fine-tuning, with a multi-task loss function = 0.5 × moisture content regression loss + 0.3 × clustering level classification loss + 0.2 × clustering region detection loss; the weighted fusion weights of the two branches are: traditional branch weight 0.4, deep learning branch weight 0.6.

[0109] The traditional machine learning training method involves constructing an ensemble model of "Random Forest + XGBoost + LightGBM" using normalized handcrafted feature vectors as input. The output objectives are: ① Regression task: predict drying efficiency (Δη / Δt, unit: % / s); ② Classification task: determine clustering level, with weights of 0.6 and 0.4 respectively. Training optimization employs 5-fold cross-validation and optimizes hyperparameters (number of trees in the Random Forest, learning rate of XGBoost, number of leaf nodes in LightGBM) through grid search, aiming to minimize the MSE loss function.

[0110] Deep learning branch training: Using preprocessed images as input, the MobileNetV3 backbone network employs transfer learning (pre-trained weights fine-tuned based on ImageNet) to reduce training costs; Multi-task loss function: Total loss = 0.5 × water content regression loss (MSE) + 0.3 × clustering level classification loss (cross-entropy) + 0.2 × clustering region detection loss (GIoU); Optimization strategy: AdamW optimizer, cosine annealing learning rate scheduling (initial 1e-3), early stopping strategy (stop if validation set loss does not decrease after 5 epochs) to avoid overfitting.

[0111] Dual-branch result fusion: The drying efficiency prediction results of the two branches are fused using a weighted voting method (traditional machine learning branch weight 0.4, deep learning branch weight 0.6); the aggregation state determination result takes the consistent output of the two branches, and if there is a discrepancy, the output of the detection head of the more accurate deep learning branch shall prevail;

[0112] The MPC control module receives the fused comprehensive feature vector in each control cycle (≤100ms) and outputs a drying efficiency prediction sequence for the next P control cycles through a hybrid model. Its core objective is to minimize the deviation between the future drying efficiency and the target drying efficiency (η˙target, determined by material type, target moisture content, and production cycle preset), while also constraining the intensity of control actions to avoid airflow instability or material damage. It continuously solves the optimization problem and outputs optimal control quantity adjustment commands (microwave power adjustment ΔP, and wind speed adjustment of the 7 sets of motor-driven air supply components Δv1−Δv7), which are then sent to the execution control module. The following optimization problem is solved in each control cycle:

[0113] ;

[0114] P represents the prediction time domain, with a value taken over 6 control cycles to adapt to the dynamic response characteristics of the drying process.

[0115] M represents the control time domain, taking values ​​for three control cycles to balance the timeliness and stability of regulation.

[0116] ΔU is the change in control quantity, including: microwave power adjustment ΔP, unit: kW; 7 sets of wind speed adjustment Δv1−Δv7, unit: m / s;

[0117] ηtarget is the target drying efficiency, with a preset range of 0.3-1.2% / s, which can be adjusted according to the material type;

[0118] ηpred represents the future drying efficiency predicted by the hybrid model;

[0119] λ is a weighting coefficient (with a value of 0.2), used to balance the accuracy of drying efficiency tracking with the smoothness of control actions;

[0120] Constraints: ΔP∈[−1.2,1.8]kW, used to match the power regulation module capability of the microwave generator; Δvi∈[−0.4,0.6]m / s, used to ensure stable suspension of materials in the cavity; the rate of change of control quantity between adjacent control cycles ≤30% to avoid sudden changes.

[0121] The intelligent suspended microwave drying system of this embodiment integrates a continuous material conveying, drying, and collection device, a honeycomb parallel suspended airflow generator, a three-dimensional microwave field generator and control device, a multi-sensor system, and further forms a closed-loop intelligent drying link through an intelligent airflow and microwave field control system module. The intelligent control model judges the airflow intensity by visually identifying whether the height of the upper material layer is level with the collection device, judges the airflow uniformity by the flatness of the bottom material layer, and judges the drying efficiency by the material color and whether the material is aggregated. The three intelligent control models (airflow intensity judgment and control, airflow uniformity judgment and control, and drying efficiency judgment and control) are all based on a hybrid architecture of "traditional machine learning + deep learning" and integrate model predictive control (MPC). They follow a unified logic of "data acquisition - preprocessing - feature engineering - hybrid training - actual calibration - closed-loop control" and are adapted to the closed-loop drying link of intelligent sensing dual-field collaborative suspended microwave drying equipment. Through the above-mentioned multi-sensor linkage, the microwave and suspended air field parameters are monitored and adjusted in real time according to the degree of material drying during the microwave drying process, so as to realize intelligent and continuous drying, improve drying efficiency, and be applicable to intelligent drying of various materials such as grains, fruit and vegetable granules, and Chinese medicinal materials.

[0122] Example 2

[0123] This embodiment provides a material drying method based on the intelligent suspension microwave drying system of Embodiment 1. The specific implementation steps are as follows:

[0124] Step 1, System Start-up and Parameter Setting: Based on the type of material to be dried (such as different materials like grains, fruits and vegetables, or Chinese medicinal herbs), set the initial moisture content, initial wind speed, initial microwave power, target drying efficiency, and effective adsorption height range of the suction component. The control system automatically loads the built-in material distribution-microwave control mapping model to complete the system initialization parameter setting.

[0125] Step 2, Intelligent wind field pre-construction: Start each group of air supply motors 1, the airflow enters the support cavity 3 through the air duct assembly 2, and the airflow is spread outward by jetting so that the incoming airflow covers the honeycomb guide grid 4, and the initial rising airflow is formed by the grid; the wind speed sensor collects data in real time, and the control system judges the wind field uniformity and drives the speed adjustment of each air supply motor according to the wind speed sensor data of each group. The speed adjustment model drives each air supply motor to adjust the speed until the standard deviation of the wind speed of each group is ≤0.3m / s, and the wind field reaches a stable state;

[0126] Step 3, Gradual Feeding and Dynamic Stabilization of the Airflow: After the airflow stabilizes, the feed hopper 5 gradually releases the material. With the help of the upward-sloping bucket-shaped inlet of the feed hopper 5 and the negative pressure effect inside the cavity, the material enters the cavity evenly. The wind speed sensor continuously collects data, and the high-speed camera 12 captures images of the bottom material layer at a receiving frequency of 1 frame per second. The control system uses the airflow uniformity judgment and control model to fine-tune the speed of the corresponding blower motor in real time to maintain the material layer flatness index to meet the following requirements: σ²≤0.05, accumulation area ratio≤5%, smoothness≤1.2, and height difference Δh≤10mm.

[0127] Step 4, Intelligent Microwave Drying Operation: A high-speed camera captures real-time images of the material's suspension distribution, color, and aggregation state. The image data is pre-processed and then transmitted to the control system. The data processing module analyzes the material's moisture content changes and aggregation level through a drying efficiency judgment and control model, generates control commands, and links the microwave generator component to achieve dynamic adjustment of power and position. The execution control module drives the microwave generator component to adjust the axial position and output power of the microwave generator (701), achieving 360° precise microwave irradiation of the material without dead angles. At the same time, the wind field intensity judgment and control model maintains the upper material height stable within the effective adsorption range of the suction component, which is generally 0-10cm according to the preset effective adsorption range.

[0128] Step 5, Drying material collection and continuous discharge: The dried material rises to the height of the suction component 9, the suction component 9 adsorbs the material under negative pressure, and the material is transferred to the receiving bin 11 through the conveying pipe 10. The discharge valve of the receiving bin 11 is dynamically opened to achieve continuous discharge.

[0129] Step 6, Closed-loop control and system shutdown: The control system continuously receives wind speed and image data synchronously, and dynamically adjusts the wind field and microwave parameters through three intelligent control models; when the material in feed hopper 5 is exhausted or the receiving hopper 11 is full, the drying ends and the system shuts down.

[0130] The above embodiments fully present the end-to-end intelligent continuous drying process of the device described in this invention. Relying on the multi-position layout and multi-sensor mode collaborative design of wind speed sensors and high-speed cameras, and the coupled dual-field collaborative control mechanism of wind field and microwave field, real-time and accurate monitoring of material distribution and suspension state is achieved throughout the entire product suspension drying process, providing effective technical support for precise control of product drying processing.

Claims

1. A smart suspended microwave drying system, characterized in that, include: The material conveying and collecting device includes a feed hopper, a high-temperature resistant glass cavity, a suction assembly, a conveying pipe and a receiving hopper connected in sequence, which is used to realize the continuous conveying of materials from feeding, drying to finished product collection; A honeycomb parallel suspended wind field generator includes: a honeycomb flow guide grid and a support cavity connected in sequence on the lower side of a high-temperature resistant glass cavity; an air duct assembly and multiple sets of air supply motors are provided on the lower side of the support cavity; the air duct assembly is connected to the support cavity; multiple sets of wind speed sensors are distributed in a ring at equal angles on the inner wall of the support cavity; the wind speed sensors are connected to the air supply motors one by one. A three-dimensional microwave field generating and controlling device includes several groups of microwave generating components arranged at equal angles on the outer peripheral wall of a high-temperature resistant glass cavity. Each group of microwave generating components includes a microwave generator and a sliding mechanism that drives the microwave generator to reciprocate along the axial direction of the high-temperature resistant glass cavity. The microwave generating components are connected to a power adjustment module. Three high-speed cameras are used to monitor the interior of the high-temperature resistant glass cavity in all directions, and to capture the suspension and distribution of materials in real time. The control system includes a data acquisition module, a data processing module, and an execution control module. The data acquisition module is electrically connected to the wind speed sensor and the high-speed camera, respectively, and is used to synchronously receive wind speed data and image data and transmit them to the data processing module. The data processing module includes a built-in wind field intensity judgment and control model, a wind field uniformity judgment and control model, and a drying efficiency judgment and control model. It constructs a hybrid architecture of traditional machine learning and deep learning and integrates model predictive control (MPC). It analyzes and processes the collected multi-source data, identifies key features such as material height, distribution uniformity, color, and aggregation state, and performs analysis and processing on these features. The execution control module is electrically connected to the adjustment modules of the motor air supply component, microwave generator component, suction component, and receiving bin, respectively. It generates control commands based on data processing results to drive the actions of each execution component and achieve coordinated optimization of the air field and microwave field.

2. The intelligent suspended microwave drying system according to claim 1, characterized in that, The lower outer periphery of the high-temperature resistant glass cavity is provided with a feed inlet, and the feed bin is connected to the feed inlet. The feed bin is in the shape of an upwardly inclined bucket. The upper end of the high-temperature resistant glass cavity is provided with a top cover assembly, and the suction assembly is located on the lower side of the top cover assembly and is connected to the receiving bin via a conveying pipe.

3. The intelligent suspended microwave drying system according to claim 1, characterized in that, The data processing module includes the following control model: The first control model, the wind field intensity judgment and control model, takes the precise mapping of "pixel coordinates → physical height" as its core, integrates traditional computer vision algorithms and lightweight deep learning detection models, and combines MPC to realize real-time judgment and dynamic control of the height of upper materials; The second control model, the wind field uniformity judgment and control model, takes "bottom material flatness characteristics → wind field uniformity mapping" as its core. It integrates traditional computer vision quantitative analysis and lightweight deep learning semantic segmentation, and combines MPC and wind speed sensor data linkage to achieve real-time judgment and precise control of wind field uniformity. The third control model, the drying efficiency judgment and control model, adopts a hybrid architecture of random forest + gradient boosting tree XGBoost / LightGBM + MobileNetV3 + multi-task head. Combined with the model prediction control MPC control module, dynamic control is achieved. The traditional machine learning branch is responsible for processing manually extracted quantitative features to ensure inference speed and interpretability. The deep learning branch is responsible for automatically mining deep image features to improve the generalization ability of complex scenes. After the feature fusion of the two branches, the optimal adjustment instructions for microwave power and wind speed are output through MPC rolling optimization, perfectly adapting to the closed-loop drying link of the equipment "sensing-analysis-control-feedback".

4. The intelligent suspended microwave drying system according to claim 3, characterized in that, The regulation process of the first-tone model includes: S1.1 Image Preprocessing and Data Acquisition: Receive material images of the upper part of the cavity from various high-speed cameras, focusing on a certain core area below the suction component; perform distortion correction using the Zhang Zhengyou calibration method, and establish a "pixel coordinates - physical height" mapping matrix; use median filtering for noise reduction, extract material areas based on HSV color threshold segmentation, and enhance edge features through grayscale conversion and histogram equalization; S1.2 Height Detection and Threshold Determination: Canny edge detection is performed on the preprocessed image to extract the upper edge contour of the material. The coordinates of the highest pixel point are fitted by Hough linear transformation and converted into the actual physical height Hreal(t). The YOLOv8n model is used to train the key point detection model of the upper edge of the material and the height value is corrected. The effective adsorption height range Htarget±ΔH of the suction component is pre-calibrated, where Htarget is the physical height of the lower edge of the suction component, i.e., the core threshold, and ΔH is the allowable deviation, which is 5-10 mm, and is dynamically updated based on the historical data of material height-adsorption success rate. S1.3 Model Predictive Control (MPC): A lightweight LSTM model is used to predict the material height for the next 5 control cycles, aiming to "stabilize the material height within the target range" and minimize height deviation and control action amplitude. The lightweight LSTM model is as follows: ： P represents the prediction time domain, with a value of 5 control cycles (adapting to the dynamic change characteristics of material height). M represents the control time domain, taking values ​​for two control cycles (balancing the timeliness and stability of regulation). ΔU represents the change in control quantity, ΔP represents the microwave power adjustment, ranging from -1.2 to 1.8 kW, Δv1-Δvn represents the n sets of wind speed adjustments, ranging from -0.3 to 0.5 m / s, and ΔF represents the negative pressure adjustment of the suction assembly, ranging from -5 to 8 kPa. Hpred represents the future material height predicted by the LSTM model, with a weighting coefficient λ=0.15 to avoid material suspension and instability. n refers to the number of blower motors and wind speed sensors.

5. The intelligent suspended microwave drying system according to claim 3, characterized in that, The regulation process of the second regulation model includes: S2.1 Image Preprocessing and Data Acquisition: Receives images of the material layer at the bottom of the cavity from a high-speed camera, and simultaneously receives data from various wind speed sensors to form a dual-source input; employs median filtering + bilateral filtering for noise reduction, histogram equalization to enhance contrast, and HSV threshold segmentation and U-Net lightweight semantic segmentation to accurately extract the material region; S2.2 Flatness Feature Extraction and Quantification: Four core indicators were extracted: ① Particle distribution uniformity index; ② Area ratio of the accumulation region; ③ Edge contour regularity index; ④ Peak height difference index. Among them, ① the particle distribution uniformity index extraction method is as follows: calculate the variance σ² of pixel density in the material area - when the wind field is uniform, the material distribution is dense and consistent, the variance value is small, and the preset standard range is σ²≤0.05; when the wind field is uneven, local accumulation leads to large density differences, and the variance value exceeds the standard; ② accumulation area quantification index: detect the protruding or concave areas in the material layer, and define the height difference >3mm as accumulation / concavity, calculate the area ratio of the accumulation area S_accum / S_total as the standard threshold ≤5%, the higher the ratio, the more serious the wind field unevenness; ③ edge contour regularity index: extract the edge contour of the material layer, calculate the smoothness of the contour = contour length / contour enclosed area ratio, when the smoothness ≤1.2, the contour is smooth and regular when the wind field is uniform, otherwise, the contour is obviously jagged when the wind field is uneven; ④ height difference peak index: based on the pixel-physical mapping relationship, calculate the physical height difference Δh between the highest pixel point and the lowest pixel point in the material layer, the standard threshold Δh≤10mm, if it exceeds, it is judged as local wind speed abnormality; Deep learning-assisted feature optimization: MobileNetV3-Small is used as the backbone network. The preprocessed bottom material layer image is input to extract a 256-dimensional deep feature vector, which is used to correct the error of traditional quantification indicators (such as density variance deviation caused by material particle size differences) and improve feature robustness. Feature fusion: The four quantified core indicators are concatenated with the deep feature vector of deep learning, and after Min-Max normalization, a 32-dimensional comprehensive flatness feature vector is formed, which serves as the core input for uniformity determination. S2.3 Wind Field Uniformity Determination: Input the comprehensive feature vector into the "Random Forest + LightGBM" ensemble classification model, output the preliminary determination result, and output the wind field uniformity determination result: 0 = uniform, 1 = slightly uneven, 2 = severely uneven; Combined with the wind speed standard deviation, when Δv≤0.3m / s, cross-validation is performed to locate abnormal areas, and the determination result and the corresponding abnormal air supply area are output. MPC Control Module: Communicatively connected to the uniformity determination module and wind speed sensor, it uses an LSTM model to predict the smoothness index for the next 6 control cycles. The target smoothness vector is σ²=0.03, S_accum / S_total=3%, smoothness=1.0, and Δh=3mm. The output wind speed adjustment amount Δvn (constraints range from -0.3 to 0.5 m / s) is calculated, with a weighting coefficient λ=0.

25. The rate of change of wind speed adjustment amount between adjacent control cycles is constrained to ≤20%. The LSTM model is as follows: P represents the prediction time domain, with a value of 6 control cycles used to adapt to the dynamic response characteristics of the wind field. After the wind field is adjusted, it takes 3-5 cycles to stabilize. M represents the control time domain, and its value is set to two control cycles to balance the timeliness of regulation with the stability of the wind field. Δvn is the wind speed adjustment of the nth group of motor air supply components, with a value range of 1-7 and a unit of m / s, which is the change in the control quantity; Φtarget is the target smoothness index vector ([σ²=0.03, S_accum / S_total=3%, smoothness=1.0, Δh=3mm]), representing the ideal smoothness state when the wind field is uniform; Φpred is the vector of future smoothness indices predicted by the LSTM model; λ is a weighting coefficient with a value of 0.25, used to balance the accuracy of the flatness regression and the wind speed adjustment range, and to avoid sudden changes in wind speed. Constraints: Δvn∈[−0.3,0.5]m / s to match the speed regulation capability of the motor air supply component; the rate of change of wind speed adjustment between adjacent control cycles ≤20% to avoid drastic fluctuations in the wind field; the standard deviation of n sets of wind speeds Δv≤0.3m / s for overall wind field uniformity constraints.

6. The intelligent suspended microwave drying system according to claim 3, characterized in that, The regulation process of the third regulation model includes: S3.1 Image Preprocessing and Data Acquisition Module: Receives real-time images from a high-speed camera, simultaneously acquires wind speed, microwave power, and measured moisture content data, and aligns them precisely with timestamps; employs median filtering + Gaussian filtering for noise reduction, converts RGB to HSV color space, and performs data enhancement processing through rotation and scaling; S3.2 Feature Extraction and Fusion: The traditional machine learning branch extracts HSV color features, aggregation state features, and process-related features of the raw materials to form a 30-35 dimensional handcrafted feature vector; the deep learning branch uses MobileNetV3 Small to extract 256 dimensional deep features, and configures a multi-task head to output the predicted moisture content, aggregation level, and aggregation region coordinates; the dual-branch features are fused into a 512 dimensional comprehensive feature vector through "splicing + attention weighting"; S3.3 Hybrid Prediction Model: The traditional machine learning branch constructs an ensemble model of "Random Forest + XGBoost + LightGBM" to predict drying efficiency Δη / Δt and clustering level; the deep learning branch uses transfer learning for fine-tuning, with a multi-task loss function = 0.5 × moisture content regression loss + 0.3 × clustering level classification loss + 0.2 × clustering region detection loss; the weighted fusion weights of the two branches are: traditional branch weight 0.4, deep learning branch weight 0.

6. MPC Control Module: Predicts the drying efficiency for the next 6 control cycles, with a target drying efficiency ηtarget (0.3-1.2% / s) as the objective. It outputs microwave power adjustment ΔP and wind speed adjustment Δv1-Δvn, with a weighting coefficient λ=0.2, and constrains the rate of change of control quantities between adjacent control cycles to ≤30%. Each control cycle solves the following optimization problem: P represents the prediction time domain, with a value taken over 6 control cycles to adapt to the dynamic response characteristics of the drying process. M represents the control time domain, taking values ​​for three control cycles to balance the timeliness and stability of regulation. ΔU is the change in control quantity, including: microwave power adjustment ΔP, unit: kW; 7 sets of wind speed adjustment Δv1−Δv7, unit: m / s; ηtarget is the target drying efficiency, with a preset range of 0.3-1.2% / s, which can be adjusted according to the material type; ηpred represents the future drying efficiency predicted by the hybrid model; λ is a weighting coefficient (with a value of 0.2), used to balance the accuracy of drying efficiency tracking with the smoothness of control actions; Constraints: ΔP∈[−1.2,1.8]kW, used to match the power regulation module capability of the microwave generator; Δvi∈[−0.4,0.6]m / s, used to ensure stable suspension of materials in the cavity; the rate of change of control quantity between adjacent control cycles ≤30% to avoid sudden changes.

7. A material drying method based on the intelligent suspension microwave drying system according to any one of claims 1-6, characterized in that, Includes the following steps: Step 1, System Start-up and Parameter Setting: Set the initial wind speed, initial microwave power, target drying efficiency, and effective adsorption height range of the suction component according to the type of material to be dried; Step 2, Intelligent wind field pre-construction: Start each group of air supply motors, the airflow enters the support cavity through the air duct assembly, and the airflow is spread outward by jetting so that the incoming airflow covers the honeycomb guide grid, and the grid is regulated to form the initial rising airflow; The wind speed sensor collects data in real time. The control system judges the wind field uniformity based on the data from each group of wind speed sensors and drives the wind turbines to adjust their speeds until the standard deviation of each group of wind speeds is ≤0.3m / s, and the wind field reaches a stable state. Step 3, Gradual Feeding and Dynamic Stabilization of the Airflow: The material is gradually released from the feed hopper. With the help of the upward-sloping bucket-shaped inlet of the feed hopper and the negative pressure effect inside the cavity, the material enters the cavity evenly. The wind speed sensor continuously collects data, and the high-speed camera captures images of the bottom material layer. The control system uses the airflow uniformity judgment and control model to fine-tune the speed of the corresponding blower motor in real time to maintain the flatness index of the material layer to meet the following requirements: σ²≤0.05, the area ratio of the accumulation area≤5%, the smoothness≤1.2, and the height difference Δh≤10mm. Step 4, Intelligent Microwave Drying Operation: A high-speed camera captures real-time images of the material's suspension distribution, color, and aggregation state. The image data is pre-processed and then transmitted to the control system. The data processing module analyzes the material's moisture content changes and aggregation level using a drying efficiency judgment and control model, generating control commands. The execution control module drives the microwave generator component, adjusting the axial position and output power of the microwave generator (701) to achieve precise 360° microwave irradiation of the material without dead angles. Simultaneously, the wind field intensity judgment and control model maintains the upper layer of material at a stable height within the effective adsorption range of the suction component. Step 5, Drying material collection and continuous discharge: The dried material rises to the height of the suction component, the suction component adsorbs the material under negative pressure, and the material is transferred to the receiving hopper through the conveying pipe. The discharge valve of the receiving hopper is dynamically opened to achieve continuous discharge. Step 6, Closed-loop control and system shutdown: The control system continuously receives wind speed and image data synchronously, and dynamically adjusts the wind field and microwave parameters through three intelligent control models; when the material in the feed hopper is exhausted or the receiving hopper is full, the drying ends and the system shuts down.

8. The material drying method according to claim 7, characterized in that, In step 1, the effective adsorption height range of the suction component is 0-10cm.

9. The material drying method according to claim 7, characterized in that, In step 4, during the intelligent microwave drying operation, the image processing algorithm identifies the material's color, aggregation area, and density, and links the microwave generator component to achieve dynamic adjustment of power and position.

10. The material drying method according to claim 7, characterized in that, The high-speed camera receives real-time image data at a frequency of 1 frame per second.

Images