Cloud edge MEMS atomization microclimate regulation and control device and method

Abstract

The invention discloses a cloud edge MEMS atomization microclimate regulation and control device and a cloud edge MEMS atomization microclimate regulation and control method. The device is composed of a cloud platform, an edge computing unit, an AI visual crop growth recognition subsystem, a three-dimensional humidity field sensing and reconstruction subsystem, a reinforcement learning control subsystem and a bionic adaptive atomization subsystem. The cloud end performs periodic training and issues a strategy, the edge side fuses multi-modal data in real time, the MEMS atomization array is driven to generate 2-8 [mu] m directional fog drops, and humidity field closed-loop control with the resolution of 10 cm and the precision of + / -3% is achieved. A field test result shows that compared with traditional high-pressure mist spraying, water is saved by 37%, the disease rate is reduced by 28%, the fresh weight of a single plant is increased by 11%, and the method is suitable for precise microclimate management of facility agricultural crops.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent equipment and precise environmental control technology for facility agriculture, and specifically involves the comprehensive application of cloud-based artificial intelligence models, edge computing platforms, multimodal sensor fusion, MEMS directional ultrasonic atomization, and reinforcement learning control strategies in the microclimate management of crop cultivation, especially a cloud-edge collaborative crop microclimate precise control system and method that can achieve three-dimensional humidity field construction and real-time closed-loop control with millimeter-level spatial resolution and second-level temporal resolution. Background Art

[0002] With the increasing scale of facility agriculture and demand for high-value-added crop cultivation, the refined management of crop growing environments has gradually become a major bottleneck limiting yield and quality. Current humidification or sprinkler irrigation solutions often use high-pressure sprays, rotary atomizers, or ultrasonic humidifiers. Their droplet size, spray direction, and timing all operate with fixed parameters and cannot be dynamically adjusted based on leaf surface conditions and spatial humidity distribution. This leads to leaf condensation, water and fertilizer waste, and a high incidence of diseases. Furthermore, single temperature and humidity point detection cannot reveal significant three-dimensional non-uniformity within the greenhouse, resulting in distorted control strategies. While attempts have been made in recent years to combine edge computing with IoT sensors, there is still a lack of high-temporal and spatial resolution humidity field reconstruction methods and control algorithms capable of self-learning optimization. Furthermore, there is a lack of an integrated solution that can perform environmental perception, crop phenotyping, and closed-loop control of atomization execution on the same platform. There is an urgent need for a new system that combines refined perception, intelligent decision-making, and flexible execution. Therefore, the present invention uses a MEMS acoustic tunnel array and phase difference control to directly synthesize droplet direction vectors within the chip, representing an innovative approach not disclosed in the prior art. Summary of the Invention

[0003] In order to overcome the defects of the existing technology that it is unable to perceive the fine-grained distribution of the humidity field in real time and cannot dynamically control according to the physiological state of crops, and to realize the refinement and intelligence of the microclimate management of facility agriculture, the present invention provides a cloud-edge MEMS atomization microclimate control device and method. By constructing a crop growth-environment coupling model in the cloud and periodically issuing strategies, multimodal sensor data is integrated in real time on the edge side, and the atomization control vector is optimized online through reinforcement learning algorithm, the MEMS atomization array with directional and variable particle size characteristics is driven to quickly establish a humidity distribution that meets the growth requirements inside the target crop canopy with millimeter-level spatial accuracy, thereby achieving water saving, energy saving, disease suppression and increased yield.

[0004] The technical solution adopted by this invention is composed of a cloud platform, an edge computing unit, an AI visual crop growth recognition subsystem, a three-dimensional humidity field perception and reconstruction subsystem, a reinforcement learning control subsystem, and a bionic adaptive atomization subsystem. The cloud platform is responsible for model training, variety knowledge base update, and policy issuance. The edge computing unit is responsible for sensor data fusion, model reasoning, and actuator scheduling. The AI ​​visual subsystem uses multispectral imaging technology to analyze leaf temperature, moisture content, and lesion distribution in real time. The humidity sensing subsystem uses a MEMS humidity array and laser scattering ranging to fuse the extended Kalman filter algorithm to reconstruct a 10 cm 3 The voxel-level humidity field and the reinforcement learning subsystem map the humidity error field and crop growth characteristics into a six-dimensional vector for atomization control based on the PPO framework. The MEMS chip of the bionic atomization subsystem generates 2–8 μm droplets at a resonant frequency of 1.6 MHz and achieves 0.2 s-level direction switching through an acoustic tunnel.

[0005] Compared with traditional ultrasonic humidification solutions, in comparative tests in tomato greenhouses and apple seedling greenhouses, the present invention saved 37% of water, reduced the disease incidence by 28%, and increased the fresh weight of individual plants by 11% within the same humidity fluctuation control range, fully demonstrating the advanced nature and creativity of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 This is the overall architecture of the system of the present invention, including the data flow and control flow relationships among the cloud platform, edge computing unit, AI visual crop growth recognition subsystem, three-dimensional humidity field perception and reconstruction subsystem, reinforcement learning control subsystem and bionic adaptive atomization subsystem.

[0007] Figure 2 This is the cross-sectional structure of the MEMS atomization chip, which includes, from top to bottom, a hydrophobic nanocoating, a micropore array, a 100 μm thick silicon substrate, and a piezoelectric transducer layer.

[0008] Figure 3 This is a simulation diagram of the sound field distribution. The color scale from blue to red represents the sound pressure intensity from low to high, proving that the acoustic impedance matching layer design between the piezoelectric transducer layer and the silicon substrate can effectively focus the sound energy and achieve efficient atomization.

[0009] Figure 4 To illustrate the three-dimensional humidity field reconstruction process, it includes sensor raw data, signal preprocessing, Mie scattering inversion model, extended Kalman filter fusion and humidity field visualization interface.

[0010] Figure 5 It is a multispectral image processing process, from visible light, near infrared, and thermal infrared images to leaf binary segmentation, NDVI index and leaf area temperature distribution.

[0011] Figure 6 To draw the curve of the comprehensive reward value of the reinforcement learning control subsystem under different algorithm strategies as the number of training iterations changes, and highlight the convergence advantage of the PPO strategy. DETAILED DESCRIPTION

[0012] Example 1: System overall architecture and deployment. In this example, at 2000 m 2 Four edge computing units are evenly spaced along the greenhouse columns. The hardware used is the NVIDIA Jetson Orin AGX, running Ubuntu 22.04 and ROS2Humble. They are connected to a cloud-based Kubernetes cluster via a 5G NR Sub-6G base station. The AI ​​vision subsystem, equipped with two multispectral cameras (450–900 nm) and a long-wave infrared thermal imager, operates at a 15 fps frame rate. It uses a modified U-Net network to perform leaf segmentation and NDVI calculation. Leaf temperature, moisture content, and lesion area are then transmitted as 256-dimensional feature vectors to the reinforcement learning control subsystem.

[0013] Example 2: 3D humidity field sensing and reconstruction. A 64-point MEMS humidity array covers a 1 m × 1 m × 2 m volume with a sampling frequency of 5 Hz. A pulsed laser with a wavelength of 532 nm and a repetition rate of 1 kHz is positioned above the array. The droplet concentration distribution is inverted using Mie scattering intensity. The extended Kalman filter algorithm combines the two types of data to discretize the humidity field into 100 × 100 × 200 voxels, achieving an average reconstruction error of less than 3%.

[0014] Example 3: Reinforcement learning control strategy. The state space is formed by concatenating the humidity error field tensor and the crop feature vector at the previous moment. The action space is the droplet size, spray direction pitch angle, azimuth angle and opening and closing time parameters of each atomization chip. The reward function integrates water utilization rate, leaf drying rate and energy consumption. 5 After rounds of simulation training, the comprehensive reward converged to 0.92. After online deployment, the edge experience buffer increments were transmitted back to the cloud every 24 hours through the federated learning mechanism for further fine-tuning.

[0015] Example 4: Bionic Adaptive Atomization Subsystem Structural Design. The MEMS atomization chip has a 120 μm thick silicon substrate, a 5 μm diameter micropore array with a 20 μm pitch, a 3 μm thick PZT-5H film, a hydrophobic trifluoropropylene / SiO2 nanocoating on the inner shell, and an acoustic impedance matching layer designed to be 240 μm thick based on the λ / 4 principle. Maximum atomization efficiency is 78% at 1.6 MHz excitation. By etching a radial acoustic tunnel around the chip outlet and applying a phase shift between the transducer layers, directivity switching within ±60° can be achieved within 0.2 seconds, with a jet divergence angle ≤ 15°.

[0016] In order to ensure that the acoustic energy of the piezoelectric transducer layer is coupled to the liquid interface to the maximum extent, the present invention designs the acoustic impedance matching layer according to the quarter-wavelength impedance matching principle, and its thickness t satisfies the formula , where c is the speed of sound in the matching layer material (epoxy-quartz composite), 2600 m·s -1 , f is the driving resonant frequency 1.6 MHz, based on which the theoretical thickness is calculated to be 0.406 mm. The actual difference in the peak atomization efficiency in the range of 0.38–0.42 mm does not exceed 1.2%, so 0.40 mm is finally selected. On the other hand, 8 radial acoustic tunnels are evenly arranged around the chip and a 0–π phase difference is applied to adjacent transducer electrodes to generate a controllable lateral acoustic pressure gradient. By changing the phase difference The amplitude can continuously adjust the jet deflection angle in the range of 0°–60°. The COMSOL acoustic field simulation and high-speed photography test results verify that when The jet deflection angle is about 59.4°, and the direction switching time constant is 0.18 s.

[0017] Example 5: Field Verification and Performance Evaluation. During the autumn tomato production cycle, the system of the present invention was compared with a conventional high-pressure mist spray system, with the target humidity range set at 80% ± 3%. The water consumption of the system of the present invention was 142 L·d. -1 , while the control group was 225 L·d -1 The cumulative duration of leaf condensation was 1.4 h·d⁻¹ and 4.7 h·d -1 The incidence of fusarium wilt was 5.6% and 7.8%, respectively; and the average per-plant fresh fruit weight was 1.96 kg and 1.76 kg, respectively. This example demonstrates that the present invention can significantly reduce water consumption and increase crop yield while maintaining humidity stability.

[0018] Result Analysis

[0019] The system verification test results show that the humidity control curves achieved by the present invention in 120 test units in a tomato greenhouse and an apple seedling greenhouse are highly consistent with the cloud-based model prediction, with an average deviation of only 2.7%, which is well consistent with the actual crop phenotypic monitoring results in the field. It can also effectively regulate known high-humidity-sensitive and humidity-tolerant vegetable varieties.

[0020] For the humidity field reconstruction error of 3%, a two-tailed t-test was conducted using 10 independent experimental data. Compared with the 8.5% error of the traditional single-point sensing + linear interpolation method, the p-value was 4.6×10 -4The standard deviation of the repeatability test was 0.37%, indicating that the algorithm has a stable and reliable accuracy advantage under different environmental conditions.

[0021] In summary, this invention offers significant innovations in hardware design, sensor fusion algorithms, reinforcement learning control strategies, and biomimetic atomization structures. Through its cloud-edge collaborative architecture and adaptive execution mechanism, it overcomes the shortcomings of existing technologies, such as their inability to perceive the fine-grained distribution of humidity fields in real time and their inability to dynamically adjust humidity based on crop physiological states. This significantly improves microclimate management in facility agriculture.

Claims

1. A cloud edge MEMS atomization microclimate control device, characterized in that The system consists of a cloud platform, an edge computing unit, an AI visual crop growth recognition subsystem, a three-dimensional humidity field perception and reconstruction subsystem, a reinforcement learning control subsystem, and a bionic adaptive atomization subsystem. The cloud platform is responsible for planting model training and remote strategy updates. The edge computing unit interacts with the cloud via 5G or Starlink communication and executes control strategies locally in real time. The AI ​​visual crop growth recognition subsystem uses a multispectral camera to obtain visible light, near-infrared, and thermal infrared information of crops and outputs plant growth status parameters. The three-dimensional humidity field perception and reconstruction subsystem reconstructs 10 cm humidity through MEMS humidity dot matrix and laser scattering inversion algorithm. 3 The spatial humidity field with high resolution is captured by a reinforcement learning control subsystem, which adaptively generates atomization direction, droplet size, and on-off timing control signals based on crop growth state parameters and humidity field reconstruction results. The bionic adaptive atomization subsystem includes several MEMS atomization chips, the inner wall of whose shell substrate is coated with a hydrophobic nano-coating, and the chip micropore array has a diameter of 5 μm. The piezoelectric transducer layer drives the liquid at a resonant frequency of 1.6 MHz to produce 2-8 μm droplets and achieves directional spraying under the control of the acoustic field phase difference.

2. The device according to claim 1, characterized in that The three-dimensional humidity field perception and reconstruction subsystem adopts the extended Kalman filter algorithm to fuse MEMS humidity dot matrix sampling data and laser scanning data inverted based on the Mie scattering model. The spatial average error of the reconstruction result does not exceed 3%.

3. The device according to claim 2, characterized in that The reinforcement learning control subsystem adopts a continuous action space strategy network based on Proximal Policy Optimization. The input is the crop growth feature vector and humidity field error field tensor of the previous moment. The output is the six-degree-of-freedom control vector of the atomization subsystem. The network is in 3×10 5 After rounds of adversarial training, the comprehensive reward value converges to 0.

92.

4. The device according to claim 3, characterized in that An acoustic impedance matching layer is set between the piezoelectric transducer layer and the silicon substrate of the MEMS atomization chip, and a radial acoustic tunnel array is formed in the micropore outlet area. The direction of the droplet jet can be switched within 0.2 seconds, and the jet divergence angle is less than 15°.

5. The device according to claim 4, characterized in that A programmable phase difference modulation module is provided between the acoustic tunnel array and the adjacent driving electrodes of the piezoelectric transducer layer, which is used to achieve mechanical-free directional switching of the droplet jet within the range of ±60°, and the direction switching time does not exceed 0.2 s.

6. The device according to claim 5, characterized in that The AI ​​visual crop growth recognition subsystem utilizes a U-Net-based leaf segmentation network in conjunction with an improved NDVI calculation method to achieve real-time extraction of leaf temperature distribution, moisture content, and lesion area.

7. The cloud edge MEMS atomization microclimate control method is characterized by: The following steps are included: (1) Data acquisition: The MEMS humidity dot matrix sensor and AI visual crop growth recognition subsystem deployed in the planting environment are used to obtain three-dimensional humidity raw data and crop leaf temperature, moisture content, and disease spot information; (2) Humidity field reconstruction: The humidity raw data and the laser scanning data inverted based on the Mie scattering model are input into the extended Kalman filter algorithm to reconstruct the three-dimensional humidity field of the target area in real time; (3) Strategy generation: A reinforcement learning control subsystem based on Proximal Policy Optimization is used to generate an atomization control strategy based on the humidity field reconstruction results and crop growth status; (4) Droplet generation: driving several MEMS atomization chips according to the atomization control strategy to generate directional droplets of 2 μm to 8 μm; (5) Closed-loop control: In the edge computing unit, the atomization control strategy is dynamically updated based on the real-time feedback of step (1) and the execution result of step (4) to achieve a 10 cm 3 Humidity field closed-loop control with spatial resolution and humidity error not exceeding ±3%.

8. The method according to claim 7, characterized in that During the training phase, the reinforcement learning control subsystem uses the cloud platform to perform offline iterative updates on the Proximal Policy Optimization algorithm and performs online fine-tuning at a frequency of 1 Hz within the edge computing unit.

9. The method according to claim 7, characterized in that The switching time of the droplet jet direction of the MEMS atomization chip within the range of ±60° does not exceed 0.2 s, and the jet divergence angle is less than 15°.

10. The method according to claim 8 or 9, characterized in that The closed-loop control process is applicable to continuous 24-h greenhouse operation tests.

Images