A soybean seed radio frequency drying intelligent monitoring system based on computer vision and deep reinforcement learning

Abstract

The application discloses a soybean seed radio frequency drying intelligent monitoring system based on computer vision and deep reinforcement learning, and relates to the technical field of agricultural products processing. The data processing hub is connected with a drying execution module, a visual monitoring module and a reinforcement learning control module, is used for receiving data of the modules and outputting control instructions, and realizes phenotype monitoring and temperature control in a soybean seed radio frequency drying process. The application can accurately identify the embryo cracks of soybean seeds, accurately monitor color characteristics, solves the problem of low precision of traditional phenotype monitoring methods, and provides reliable data support for the regulation and control of the drying process.

Description

Technical Field

[0001] This invention relates to the field of agricultural product processing technology, and more specifically to an intelligent monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning. Background Technology

[0002] Soybeans are a crucial oilseed crop globally. In 2020, Brazil, the United States, and Argentina accounted for 82.9% of global production. China, the world's largest soybean consumer, produces only 8.1% of global production, yet imports 62.01% of total global grain imports. However, increasing yields through expanded planting areas conflicts with China's strategic goal of achieving basic self-sufficiency in other food crops. Therefore, increasing soybean yield per unit area is imperative, and maintaining the germination capacity of soybean seeds is key to achieving this goal. However, during the harvesting and processing of soybean seeds in China, mechanical damage leading to viability loss is significant (approximately 5%-12%), with drying being the primary cause of this loss. This is because industrially, hot air drying is commonly used to extend the shelf life of soybean seeds; during this process, heat transfer and mass transfer occur in opposite directions, causing stress concentration within the seed and resulting in cracks that hinder germination.

[0003] To address these challenges, radio frequency (RF) and microwave technologies exhibit potential advantages because these dielectric heating techniques generate frictional heat by inducing polarization of microscopic particles within the material. RF, with its longer wavelength, enables deeper penetration, allowing for rapid heating of the material volume. However, the high heating efficiency of RF amplifies temperature overshoot caused by thermal inertia during the drying process, making precise temperature control difficult. Deep reinforcement learning can interact with real-time data from the actual environment, making autonomous decisions and training based on real-time temperature data during the interaction. This enhances the dynamic adaptability and robustness of temperature control, thereby improving the temperature control accuracy of the RF drying process.

[0004] To analyze the effectiveness of radio frequency (RF) heating, online monitoring is needed during soybean seed drying to capture changes in seed cracking. Computer vision can be used to detect real-time changes in the appearance quality of dried products, providing rich data support for the control of the drying process. However, due to the small color difference between cracked and non-cracked areas of soybean seeds, color threshold-based segmentation methods cannot achieve accurate monitoring. Semantic segmentation, by combining computer vision and deep learning techniques, classifies each pixel in the image according to its semantic category, thereby achieving pixel-level semantic understanding. The study used semantic segmentation technology to monitor the hot air drying process online, achieving an intersection-union (IUU) of 98.16% and an average precision of 98.96%.

[0005] Therefore, proposing an intelligent monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning to solve the difficulties of existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides an intelligent monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning, which is used to solve the technical problems existing in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A smart monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning includes: a drying execution module, a visual monitoring module, a reinforcement learning control module, and a data processing center; wherein, The data processing center is connected to the drying execution module, the visual monitoring module, and the reinforcement learning control module, respectively, to receive data from each module and output control commands, thereby realizing phenotypic monitoring and temperature control during the radio frequency drying process of soybean seeds.

[0008] Optionally, the drying execution module is an RF drying device, including a metal drying chamber, an RF generator, electrodes, a vacuum pump, a pneumatic solenoid valve, and a weight sensor. The electrode spacing inside the drying chamber is 250 mm. The radio frequency generator operates at a frequency of 27.12 MHz, has a power of 1 kW and a 50 Ω specification, and is connected to the electrodes to transmit radio frequency energy. The weight sensor is installed on the outside of the drying chamber to monitor changes in the weight of soybean seeds.

[0009] Optionally, the visual monitoring module includes an image acquisition unit and an image processing unit; The image acquisition unit consists of an industrial camera and a ring-shaped shadowless lamp installed on the top of the drying chamber. The resolution of the industrial camera is set to 2048×1536 pixels. The image processing unit adopts the FPE-DeepLab-C model, which integrates a content-aware feature reconstruction module, concurrent spatial and channel compression and attention-stimulating mechanism, and discrete wavelet transform, and is equipped with a feature pre-extraction module.

[0010] Optionally, the feature pre-extraction module includes a discrete wavelet transform processing unit, a dense block unit, and a feature fusion unit; The discrete wavelet transform processing unit processes the input image to generate four frequency component sub-bands, which are then processed by a 3×3 convolutional layer, ReLU activation function, and inverse discrete wavelet transform to obtain frequency domain features; the dense block unit has a growth rate of 48, and the output of each layer is connected to the input of all subsequent layers. The feature fusion unit concatenates frequency domain features with spatial domain features, processes them through concurrent spatial and channel compression and attention activation mechanisms, and then achieves fusion through a 1×1 convolutional layer and the ReLU activation function.

[0011] Optionally, the training parameters for the FPE-DeepLab-C model are as follows: the input image is adjusted to 512×512 pixels, and the freeze training method is used. First, the backbone network is frozen and the head network is trained for 50 epochs, and then the backbone network is unfrozen and trained for 120 epochs. The frozen training batch size is 16, and the unfrozen batch size is 8. The initial learning rate is 0.001, and the cosine annealing strategy is used for adjustment. The average crossover ratio (OCR) of the model is verified every 5 epochs until the loss value and the average OCR fluctuate stably.

[0012] Optionally, the reinforcement learning control module includes a fiber optic temperature sensor and a deep Q-network model. The fiber optic temperature sensor is used to collect the real-time temperature of soybean seeds. The state space of the deep Q-network model includes the moisture content, temperature, specific heat capacity, heating time and thermal conductivity of soybean seeds. The action space includes turning the radio frequency generator on and off. Each action lasts for 5 seconds. The reward function is set according to the temperature deviation and temperature threshold.

[0013] Optionally, the reinforcement learning control module constructs a simulation environment based on the dielectric properties, heat transfer, and mass transfer models of soybean seeds. One time step in the simulation environment corresponds to 5 seconds in the real environment. The expressions for the dielectric properties, heat transfer, and mass transfer models are as follows: Dielectric property model:

[0014] in, , Where is the dielectric constant. For temperature, It is the dielectric loss factor; Heat transfer model:

[0015] in, This refers to the voltage across the anode plate. The distance from the material to the anode plate is 0.01 m. The thickness of the material is 0.18 m. Bulk density of the material, in kg / m³ 3 ; , where is the specific heat capacity of soybean, expressed in J / kg·K; The vacuum permittivity is 8.854 × 10⁻⁶.-12 F / m; The conversion power from electromagnetic energy to heat energy, measured in W / m³. 3 ; The frequency is 27.12 MHz (radio frequency). denoted as electric field strength.

[0016] Mass transfer model:

[0017] in, This is the rate of water evaporation per unit time per unit surface area, expressed in kg / m². 2 ·s; The latent heat of vaporization per unit mass is expressed in J / kg. The temperature of the ambient air in the radio frequency heating chamber; It is the normal vector perpendicular to the material surface; The total convective heat transfer coefficient is expressed in W / m³. 2 ·K.

[0018] Optionally, the deep Q-network model includes a policy network and a target network with the same architecture, both of which are fully connected networks with a 4-dimensional input layer, a 128-dimensional hidden layer, and a 2-dimensional output layer. The parameters of the target network are soft-updated through the parameters of the policy network, with an update rate τ of 0.005. An experience replay buffer with a capacity of 50,000 is used to store the state transition vectors. The policy network weights are optimized by the Adam optimizer with a learning rate of 0.0005 and 150,000 training steps.

[0019] Optionally, the data processing hub uses the NVIDIA Jetson XavierNX development board.

[0020] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a smart monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning, the beneficial effects of which are: 1) By setting up a visual monitoring module and adopting the FPE-DeepLab-C model, which integrates a content-aware feature reconstruction module, concurrent space and channel compression and attention-stimulating mechanism and discrete wavelet transform, it can accurately identify germ cracks in soybean seeds and accurately monitor color features, thus solving the problem of low accuracy in traditional phenotypic monitoring methods and providing reliable data support for the regulation of the drying process. 2) The reinforcement learning control module is based on a deep Q-network model and is trained in a simulated environment. The simulated environment is based on the dielectric properties, heat transfer, and mass transfer models of soybean seeds, which can accurately reflect the changes in the thermal properties of the seeds during the drying process. The trained model has good dynamic adaptability and robustness. Compared with traditional PID control, the temperature control accuracy is higher, the maximum temperature deviation and overshoot are significantly reduced, and the seed cracking caused by improper temperature control is effectively avoided. 3) By coordinating the work of the drying execution module, visual monitoring module, and reinforcement learning control module through the data processing center, closed-loop control of the soybean seed drying process is realized. It can monitor seed quality in real time and accurately adjust the temperature according to the monitoring results, ensuring the germination rate of soybean seeds. This provides a reliable solution for high-quality soybean seed drying and promotes the intelligent development of agricultural product processing. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 The present invention provides a structural diagram of an intelligent monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning. Figure 2 This is a structural diagram of the DeepLab V3+ model provided by the present invention; Figure 3 The structure diagram of the FPE-DeepLab-C model provided by this invention; Figure 4 This invention provides a content-aware feature reorganization structure diagram. Figure 5 This is a structural diagram of the channel extrusion and spatial excitation module provided by the present invention; Figure 6 This is a structural diagram of the spatial compression and channel activation module provided by the present invention; Figure 7 A structural diagram of the space and channel synchronous extrusion and excitation module provided by the present invention; Figure 8 The structural diagram of the feature pre-extraction module provided by the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] See Figure 1 As shown, this invention discloses an intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning, comprising: a drying execution module, a visual monitoring module, a reinforcement learning control module, and a data processing center; wherein, The data processing center is connected to the drying execution module, the visual monitoring module, and the reinforcement learning control module, respectively, to receive data from each module and output control commands, thereby realizing phenotypic monitoring and temperature control during the radio frequency drying process of soybean seeds.

[0025] Furthermore, the drying execution module is an radio frequency drying device, including a metal drying chamber, a radio frequency generator, electrodes, a vacuum pump, a pneumatic solenoid valve, and a weight sensor. The electrode spacing inside the drying chamber is 250 mm. The radio frequency generator operates at a frequency of 27.12 MHz, has a power of 1 kW and a 50 Ω specification, and is connected to the electrodes to transmit radio frequency energy. The weight sensor is installed on the outside of the drying chamber to monitor changes in the weight of soybean seeds.

[0026] Furthermore, the visual monitoring module includes an image acquisition unit and an image processing unit; The image acquisition unit consists of an industrial camera and a ring-shaped shadowless lamp installed on the top of the drying chamber. The resolution of the industrial camera is set to 2048×1536 pixels. The image processing unit adopts the FPE-DeepLab-C model, which integrates a content-aware feature reconstruction module, concurrent spatial and channel compression and attention-stimulating mechanism, and discrete wavelet transform, and is equipped with a feature pre-extraction module.

[0027] Specifically, soybean seed images were captured using an industrial camera, supplemented by a ring-shaped shadowless lamp as an auxiliary light source, providing 36,000 lux of illumination. All images were taken at a resolution of 2048×1536 pixels, with the germ facing the camera directly during shooting. A total of 1,679 soybean seed images were acquired, each saved as a JPG file. Data augmentation was performed on the acquired images: 50% random flipping, 50% Gaussian blur application, and 100% random rotation, ultimately generating 5,037 enhanced images. Among these, 3,087 images were of single soybean seeds, and 1,950 images contained two or more soybean seeds (see Table 1). The soybean seed images were divided into training, validation, and test sets in a 6:2:2 ratio (see Table 2). The Labelme tool was used to label the images, categorizing them into "seeds with germ cracks" and "seeds without germ cracks," with the labeling information saved in TXT format.

[0028] Table 1 Distribution of soybean seed images

[0029] Table 2. Division of the soybean seed dataset

[0030] The DeepLab V3+ model employs an encoder-decoder architecture and a spatially hollow pyramid pooling method. This architectural improvement makes the integration of global and local information more efficient. The structure of the DeepLab V3+ model is as follows: Figure 2 As shown.

[0031] Content-aware feature reconstruction technology is used to upsample the DeepLab V3+ model to obtain rich semantic information. To address the low segmentation accuracy of the DeepLab V3+ model when handling small target cracks and edges, a feature pre-extraction module is constructed by integrating parallel spatial and channel compression, channel-excited attention mechanisms, and discrete wavelet transform. This module aims to improve the sensitivity and accuracy of the model in extracting detailed features. The structure of the FPE-DeepLab-C model is as follows: Figure 3 As shown.

[0032] The standard DeepLab V3+ model uses bilinear interpolation for feature map upsampling, which is simple in structure and computationally efficient. However, this method's upsampling process mainly focuses on the sub-pixel domain, leading to blurring of seed gaps and edge features. Compared to bilinear interpolation upsampling, CARAFE upsampling has significant advantages in this regard: its structure is lightweight yet possesses a wide receptive field, and it can dynamically generate adaptive kernel functions based on the input content, thus utilizing contextual information more efficiently. Its network structure is as follows: Figure 4 As shown.

[0033] The kernel prediction module is responsible for generating the reassembled kernel. First, the input feature map is compressed using channel compression, and then a feature map is generated through a 1×1 convolution operation. This step effectively reduces the subsequent computational load. Next, the encoder... An upsampling kernel is used for prediction, followed by spatial dimensionality expansion to generate a feature map of shape [shape omitted]. Finally, normalization is achieved using the softmax function. The content-aware reassembly module is designed to map the position of each output feature onto the input feature. Then, within a rectangular region centered at that position, a dot product operation is performed with the upsampling kernel to generate the output feature map.

[0034] To enhance the key feature extraction capability of semantic segmentation models, this invention introduces the scSE attention mechanism, such as... Figure 7 As shown, the spatial dimension (channel compression and spatial excitation, sSE) and the channel dimension (spatial compression and channel excitation, cSE) are fused together, and the feature outputs of these two dimensions are fused along the channel axis, thereby effectively improving the performance of the feature map in both spatial and channel dimensions.

[0035] like Figure 5 As shown, the sSE module can enhance and integrate spatial dimensional information, extract weight information, and multiply the weight information with the original feature map to achieve attention enhancement. It uses a 1×1 convolutional layer with one output channel to integrate convolutional layer information. Specifically, according to the following formula, the input feature map is first divided into several blocks.

[0036]

[0037]

[0038]

[0039] The cSE module is a channel attention mechanism, and its structure is as follows: Figure 6 As shown. This design aims to reduce the dimensionality of feature maps and enhance attention weights. By using this module, channel information can be effectively integrated, simplifying module complexity, reducing computational load, and improving computational speed. In this module, features... .

[0040]

[0041]

[0042] .

[0043] Furthermore, the feature pre-extraction module includes a discrete wavelet transform processing unit, a dense block unit, and a feature fusion unit; The discrete wavelet transform processing unit processes the input image to generate four frequency component sub-bands, which are then processed by a 3×3 convolutional layer, ReLU activation function, and inverse discrete wavelet transform to obtain frequency domain features; the dense block unit has a growth rate of 48, and the output of each layer is connected to the input of all subsequent layers. The feature fusion unit concatenates frequency domain features with spatial domain features, processes them through concurrent spatial and channel compression and attention activation mechanisms, and then achieves fusion through a 1×1 convolutional layer and the ReLU activation function.

[0044] For details, see Figure 8 As shown, the feature pre-extraction module starts with the frequency domain feature extraction process, which can capture multiple features in the image and minimize the influence of noise. The LWT layer is applied to the input image to generate four sub-bands, representing the approximate (global information), horizontal, vertical, and diagonal frequency components, respectively.

[0045] When extracting subbands from the input image using Discrete Wavelet Transform (DWT), downsampling is performed to expand the receptive field. Taking into account the expanded receptive field, a 3×3 convolutional layer is applied to each subband, as shown in the following formula:

[0046]

[0047]

[0048] in, DWT It is a discrete wavelet transform. IDWT It is the inverse discrete wavelet transform. It is by... DWT Applied to The obtained sub-band, Input image, , It is the output of each sub-band after passing through the convolutional layer and applying the activation function. It is a two-dimensional convolutional layer. This is the activation function.

[0049] To minimize the loss of high-frequency information and preserve fine-grained information during the DWT process, dense blocks are used to enhance identity feature extraction. The design of the dense blocks ensures that the output of each layer is connected to the input of all subsequent layers, thereby achieving continuous feature extraction. As shown in the following formula, the composite function consists of 3×3 convolutional layers and ReLU, and the growth rate of the dense blocks is 48.

[0050]

[0051] This represents the output obtained after applying dense blocks.

[0052] Frequency domain features obtained through DWT and dense blocks are converted into spatial domain features using IDWT. However, these features exhibit different properties compared to features extracted directly from the spatial domain. Therefore, a process is needed to fuse these two types of features. In this process, wavelet features and spatial domain features are first concatenated, and then the scSE attention mechanism is applied to highlight important channels extracted from both domains. As shown in the following formula, the features obtained after the scSE step are fused using a 1×1 convolutional layer and ReLU.

[0053]

[0054]

[0055] It is the output after channel cascading. This represents the final output of the module after applying scSE, 1×1 convolution, and ReLU steps.

[0056] Furthermore, the training parameters of the FPE-DeepLab-C model are as follows: the input image is adjusted to 512×512 pixels, and the freeze training method is adopted. First, the backbone network is frozen and the head network is trained for 50 epochs, and then the backbone network is unfrozen and trained for 120 epochs. The batch size for frozen training is 16, and it is 8 after unfreezing. The initial learning rate is 0.001, and the cosine annealing strategy is used for adjustment. The average crossover ratio (OCR) of the model is verified every 5 epochs until the loss value and the average OCR fluctuate stably.

[0057] Furthermore, the reinforcement learning control module includes a fiber optic temperature sensor and a deep Q-network model. The fiber optic temperature sensor is used to collect the real-time temperature of soybean seeds. The state space of the deep Q-network model includes the moisture content, temperature, specific heat capacity, heating time, and thermal conductivity of the soybean seeds. The action space includes turning the radio frequency generator on and off. Each action lasts for 5 seconds. The reward function is set according to the temperature deviation and temperature threshold.

[0058] Furthermore, the reinforcement learning control module constructs a simulation environment based on the dielectric properties, heat transfer, and mass transfer models of soybean seeds. One time step in the simulation environment corresponds to 5 seconds in the real environment. The expressions for the dielectric properties, heat transfer, and mass transfer models are as follows: Dielectric property model:

[0059]

[0060] in, , Where is the dielectric constant. For temperature, It is the dielectric loss factor; Specifically, in the process of radio frequency heating, dielectric properties are the most critical factor affecting the heating rate and heating effect of materials. Its main characteristics are the dielectric constant and dielectric loss factor. Dielectric properties are affected by a variety of factors, including moisture content, temperature and frequency.

[0061] Heat transfer model:

[0062]

[0063]

[0064] in, This refers to the voltage across the anode plate. The distance from the material to the anode plate is 0.01 m. The thickness of the material is 0.18 m. Bulk density of the material, in kg / m³ 3 ; , where is the specific heat capacity of soybean, expressed in J / kg·K; The vacuum permittivity is 8.854 × 10⁻⁶. -12 F / m; The conversion power from electromagnetic energy to heat energy, measured in W / m³. 3 ; The frequency is 27.12 MHz (radio frequency). denoted as electric field strength.

[0065] Specifically, in the radio frequency (RF) system, the size of the anode plate (280×280cm²) is much smaller than the RF wavelength (11m), therefore the voltage on the anode plate can be approximated as uniformly distributed. Based on the premise that RF energy is uniformly absorbed by the material, a voltage estimation model for the anode plate based on the material's average heating rate is established:

[0066] The distance from the material to the anode plate is 0.01 m; the material thickness is 0.18 m; and the vacuum dielectric constant is 8.854 × 10⁻⁶. -12 F / m; Specific heat capacity of soybean is J / (kg·K); Bulk density of material is kg / m³ 3 The heating time is s.

[0067]

[0068] Left and right electrodes The spacing between the boards is 0.25 meters.

[0069] After solving the electromagnetic equations, the heat generation inside the material can be derived from the electromagnetic equations and the material's characteristic parameters. Therefore, it is necessary to establish the relationship between the electromagnetic equations and the heat conduction equations. The power of electromagnetic energy to thermal energy conversion (P, W / m³) is closely related to the dielectric properties of the material being processed. For a given radio frequency heating system with a constant electric field strength, the following calculation can be used:

[0070] The electric field strength is expressed in V / m.

[0071] Furthermore, in the case of radio frequency heating of solid materials, the transient heat conduction differential equation describing heat transfer in the material is a mathematical relationship established by combining the law of energy conservation with the unsteady-state Fourier equation, as shown in the following equation:

[0072] The thermal diffusivity of the heating material is The unit is m 2 / s; thermal conductivity is The unit is W / (m·K).

[0073] If heat loss between the heating material and the ambient air is ignored, the calculation method for the material heating rate can be rewritten as the following formula: .

[0074] Mass transfer model:

[0075] in, This is the rate of water evaporation per unit time per unit surface area, expressed in kg / m². 2 ·s; The latent heat of vaporization per unit mass is expressed in J / kg. The temperature of the ambient air in the radio frequency heating chamber; It is the normal vector perpendicular to the material surface; The total convective heat transfer coefficient is expressed in W / m³. 2 ·K.

[0076] Solve by the convective heat transfer boundary equation between the material surface and the ambient air: .

[0077] Specifically, during the radio frequency heating of materials, moisture migration always accompanies heat transfer. Changes in the material's moisture content lead to alterations in dielectric properties, thus affecting heating efficiency. Therefore, moisture migration from the material to the surrounding air must be considered. Boundary conditions related to the heat of vaporization effect need to be defined at the interface between the material and the ambient air. The internal heat conduction of the material must be balanced with convective heat transfer from the external environment and heat loss due to moisture evaporation.

[0078] Furthermore, the deep Q-network model includes a policy network and a target network with the same architecture, both of which are fully connected networks with a 4-dimensional input layer, a 128-dimensional hidden layer, and a 2-dimensional output layer. The parameters of the target network are soft-updated through the parameters of the policy network, with an update rate τ of 0.005. An experience replay buffer with a capacity of 50,000 is used to store the state transition vectors. The policy network weights are optimized by the Adam optimizer with a learning rate of 0.0005 and 150,000 training steps.

[0079] Specifically, during training, the parameters of the target network are soft-updated through the parameters of the policy network.

[0080]

[0081] The update rate is defined. The learning process is stabilized by reducing the correlation between action values ​​and target values, with an update rate of 0.005. DQN's training data is generated using a replay buffer.

[0082] The optimizations to the replay buffer and policy network are as follows: (1) Agent architecture The agent's network takes the current state as input and predicts the Q-value for each action. The network consists of a fully connected network with three layers: an input layer (4-dimensional), a hidden layer (128-dimensional), and an output layer (2-dimensional). These layers transform the input features into two outputs, corresponding to the action value (RF switch state) for each possible action.

[0083] (2) Experience playback buffer

[0084] To generate training data for DQN, an empirical replay buffer was used. This replay buffer is a circular buffer with a capacity of 5000, containing state transition vectors. From this buffer, a dataset of size [missing value] is extracted. Mini-batches are used for network training. In the simulation, agents operating according to a probability-based training policy fill the buffer, where the agents use probability vectors. Select a random action from the action space:

[0085]

[0086] Higher values ​​result in all actions having equal probabilities, while lower values ​​lead to a greedy training strategy. This probability-based training strategy has an advantage over the standard greedy strategy because it does not depend on the number of training steps, which typically requires a scheduling scheme that decreases with the number of training steps.

[0087] (3) Optimize the policy network

[0088] The optimal action value function is approximated by minimizing the smooth L1 loss between the target and predicted values. The weights are optimized using the Adam optimizer (learning rate). Optimization of the policy network begins after the replay buffer is filled to 50% and continues for the number of training steps. The optimal weights are selected based on the highest average reward observed within the evaluation period (time step in the simulation) on the validation set.

[0089] .

[0090] Furthermore, the data processing hub utilizes the NVIDIA Jetson Xavier NX development board.

[0091] Specifically, the system's workflow is as follows: (1) Set the target temperature (e.g., 40℃, 55℃, 70℃) through the human-computer interaction software, start the drying execution module, and start the radio frequency generator to perform radio frequency drying on the soybean seeds in the drying chamber; (2) The industrial camera in the visual monitoring module acquires soybean seed images in real time with the assistance of the ring shadowless lamp. The acquired images are transmitted to the image processing unit and processed by the FPE-DeepLab-C model to realize germ crack recognition and color feature monitoring. The monitoring data is transmitted to the data processing center. (3) The fiber optic temperature sensor in the reinforcement learning control module collects the temperature data of soybean seeds in real time, and the weight sensor collects the weight data of seeds. The reinforcement learning control module calculates the moisture content of seeds by combining the weight data, and then calculates the specific heat capacity and thermal conductivity. The state parameters such as temperature, moisture content, specific heat capacity, thermal conductivity and heating time are input into the trained deep Q network model. The model outputs the RF generator switch control signal and transmits it to the data processing center. (4) The data processing center receives the monitoring data from the visual monitoring module and the control signal from the reinforcement learning control module. It judges the seed status based on the monitoring data, adjusts the working status of the radio frequency generator in combination with the control signal, and judges whether the drying endpoint has been reached (the seed moisture content reaches the safe moisture content of 13%). If it is reached, the drying process is terminated; otherwise, the above steps are repeated until the drying endpoint is reached.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0093] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A smart monitoring system for radio frequency drying of soybean seeds based on computer vision and deep reinforcement learning, characterized in that, include: The system includes a drying execution module, a visual monitoring module, a reinforcement learning control module, and a data processing center; among which, The data processing center is connected to the drying execution module, the visual monitoring module, and the reinforcement learning control module, respectively, to receive data from each module and output control commands, thereby realizing phenotypic monitoring and temperature control during the radio frequency drying process of soybean seeds.

2. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 1, characterized in that, The drying execution module is an RF drying device, including a metal drying chamber, an RF generator, electrode plates, a vacuum pump, a pneumatic solenoid valve, and a weight sensor. The electrode spacing inside the drying chamber is 250 mm. The radio frequency generator operates at a frequency of 27.12 MHz, has a power of 1 kW, and is 50Ω. It is connected to the electrodes to transmit radio frequency energy. The weight sensor is installed on the outside of the drying chamber to monitor changes in the weight of soybean seeds.

3. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 1, characterized in that, The visual monitoring module includes an image acquisition unit and an image processing unit; The image acquisition unit consists of an industrial camera and a ring-shaped shadowless lamp installed on the top of the drying chamber. The resolution of the industrial camera is set to 2048×1536 pixels. The image processing unit adopts the FPE-DeepLab-C model, which integrates a content-aware feature reconstruction module, concurrent spatial and channel compression and attention-stimulating mechanism, and discrete wavelet transform, and is equipped with a feature pre-extraction module.

4. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 3, characterized in that, The feature pre-extraction module includes a discrete wavelet transform processing unit, a dense block unit, and a feature fusion unit; The discrete wavelet transform processing unit processes the input image to generate four frequency component sub-bands, which are then processed by a 3×3 convolutional layer, ReLU activation function, and inverse discrete wavelet transform to obtain frequency domain features; the dense block unit has a growth rate of 48, and the output of each layer is connected to the input of all subsequent layers. The feature fusion unit concatenates frequency domain features with spatial domain features, processes them through concurrent spatial and channel compression and attention activation mechanisms, and then achieves fusion through a 1×1 convolutional layer and the ReLU activation function.

5. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 3, characterized in that, The training parameters for the FPE-DeepLab-C model are as follows: the input image is adjusted to 512×512 pixels, and the frozen training method is used. First, the backbone network is frozen and the head network is trained for 50 epochs, and then the backbone network is unfrozen and trained for 120 epochs. The frozen training batch size is 16, and the unfrozen batch size is 8. The initial learning rate is 0.001, and the cosine annealing strategy is used for adjustment. The average crossover ratio (OCR) of the model is verified every 5 epochs until the loss value and the average OCR fluctuate stably.

6. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 1, characterized in that, The reinforcement learning control module includes a fiber optic temperature sensor and a deep Q-network model. The fiber optic temperature sensor is used to collect the real-time temperature of soybean seeds. The state space of the deep Q-network model includes the moisture content, temperature, specific heat capacity, heating time and thermal conductivity of soybean seeds. The action space includes turning the radio frequency generator on and off. Each action lasts for 5 seconds. The reward function is set according to the temperature deviation and temperature threshold.

7. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 6, characterized in that, The reinforcement learning control module constructs a simulation environment based on the dielectric properties, heat transfer, and mass transfer models of soybean seeds. One time step in the simulation environment corresponds to 5 seconds in the real environment. The expressions for the dielectric properties, heat transfer, and mass transfer models are as follows: Dielectric property model: in, , For dielectric constant and dielectric loss factor, For temperature, This refers to the moisture content of soybean seeds; Heat transfer model: in, This refers to the voltage across the anode plate. The distance from the material to the anode plate is 0.01 m. The thickness of the material is 0.18m; Bulk density of the material, in kg / m³ 3 ; , where is the specific heat capacity of soybean, expressed in J / kg·K; The vacuum permittivity is 8.854 × 10⁻⁶. -12 F / m; The conversion power from electromagnetic energy to heat energy, measured in W / m³. 3 ; The frequency is 27.12 MHz (radio frequency). Electric field strength; Mass transfer model: in, This is the rate of water evaporation per unit time per unit surface area, expressed in kg / m². 2 ·s; The latent heat of vaporization per unit mass is expressed in J / kg. The temperature of the ambient air in the radio frequency heating chamber; It is the normal vector perpendicular to the material surface; The total convective heat transfer coefficient is expressed in W / m³. 2 ·K.

8. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 6, characterized in that, The deep Q-network model includes a policy network and a target network with the same architecture, both of which are fully connected networks with a 4-dimensional input layer, a 128-dimensional hidden layer, and a 2-dimensional output layer. The parameters of the target network are soft-updated through the parameters of the policy network, with an update rate τ of 0.

005. An experience replay buffer with a capacity of 50,000 is used to store the state transition vectors. The weights of the policy network are optimized by the Adam optimizer with a learning rate of 0.0005 and 150,000 training steps.

9. The intelligent monitoring system for soybean seed radio frequency drying based on computer vision and deep reinforcement learning according to claim 1, characterized in that, The data processing hub uses the NVIDIA Jetson Xavier NX development board.

Images