An adaptive control device and method for rice seed drying based on LF-NMR and XGBoost

Abstract

This invention discloses an adaptive control device and method for rice seed drying based on LF-NMR and XGBoost, relating to the field of post-harvest processing and intelligent control technology for agricultural products. The device includes a gas-solid fluidized bed drying module, an in-situ nuclear magnetic resonance (NMR) detection module, and a control module. The gas-solid fluidized bed drying module includes a drying chamber, a hot air blower, and a cold air blower. The in-situ NMR detection module acquires spin echo train attenuation signals. The control module incorporates the SIRT algorithm, which inverts and decouples the physical features of the spin echo train attenuation signals to obtain core feature variables. These variables are then input into an XGBoost model for mapping calculation, followed by threshold determination. Based on the threshold determination result, the hot air blower and cold air blower are physically closed or opened. This device can solve the problem of decreased rice seed vigor in existing drying processes, featuring an in-situ online feedback mechanism and a non-destructive online feedback mechanism. The algorithm's accuracy and robustness are significantly improved, thus enhancing the quality of rice seed drying.

Description

Technical Field

[0001] This invention relates to the field of post-harvest processing and intelligent control technology for agricultural products, specifically to an adaptive control device and method for drying rice seeds based on LF-NMR (low-field nuclear magnetic resonance) and XGBoost. Background Technology

[0002] Rice is a cereal crop of the genus Oryza, and post-harvest drying of rice seeds is a crucial step in preventing mold growth and maintaining their seed value. Current technologies for controlling rice seed drying have the following main shortcomings: 1. Single and Lagging Control Indicators: Existing drying processes primarily rely on macroscopic moisture content for endpoint control. However, macroscopic moisture content cannot reflect the microstructural damage to rice seeds caused by thermal effects. Studies have shown that at the same endpoint moisture content, different drying processes result in significant differences in rice seed germination rates. During the dehydration stage, as free water is gradually removed, bound water binds more tightly to the substrate; when overheating causes damage, it is essentially due to the collapse of the cellular microstructure, leading to a large loss of macromolecular binding sites. Traditional methods relying on macroscopic moisture content control are completely unable to capture these mutations in subcellular structural binding sites.

[0003] 2. Lack of a true in-situ online feedback mechanism: Most existing drying equipment with NMR detection adopts bypass sampling or offline sampling methods, which not only interferes with the continuity of drying, but also the sudden change in the environment at the moment of sampling can easily introduce errors, making it impossible to achieve real-time in-situ feedback under the fluidized state inside the drying chamber.

[0004] 3. Lack of non-destructive online feedback mechanism: Traditional rice seed vigor testing relies on standard germination tests, which take more than 7 days and are destructive tests, making them unsuitable for real-time feedback and adjustment during the drying process.

[0005] 4. Prediction models are highly susceptible to noise interference: Most existing technologies directly input full-spectrum NMR data into linear models (such as PLSR), failing to eliminate noise interference caused by drastic fluctuations in free water during the later stages of drying, resulting in low prediction accuracy (R0). 2 Typically below 0.6, it fails to meet the robustness requirements of industrial control. Summary of the Invention

[0006] The purpose of this invention is to overcome the aforementioned problems and provide an adaptive control device for rice seed drying based on LF-NMR and XGBoost. This device can solve the problem of decreased rice seed viability caused by the inability to detect microstructural collapse and loss of macromolecular binding sites in the existing drying process. It has an in-situ online feedback mechanism and a non-destructive online feedback mechanism, significantly improving the accuracy and robustness of the algorithm and improving the drying quality of rice seeds.

[0007] Another objective of this invention is to provide an adaptive control method for rice seed drying based on LF-NMR and XGBoost.

[0008] The objective of this invention is achieved through the following technical solution: An adaptive control device for rice seed drying based on LF-NMR and XGBoost includes a gas-solid fluidized bed drying module, an in-situ nuclear magnetic resonance detection module, and a control module; wherein, The gas-solid fluidized bed drying module includes a drying chamber, a hot air blower, and a cold air blower. The drying chamber is provided with a feeding port and a discharging port. The drying chamber includes an exhaust end at the top, a detection section in the middle, and an air inlet at the bottom. The hot air blower and the cold air blower are respectively connected to the air inlet. The in-situ nuclear magnetic resonance detection module includes an RF coil and a nuclear magnetic resonance control system, wherein the nuclear magnetic resonance control system is connected to the RF coil; the nuclear magnetic resonance control system controls the RF coil to emit a CPMG pulse sequence to collect the spin echo train attenuation signal of hydrogen protons in rice seeds; The air inlet and air outlet of the drying chamber are both equipped with temperature control probes; the control module is connected to the temperature control probes, the nuclear magnetic resonance control system, the hot air blower and the cold air blower respectively. The control module incorporates the SIRT algorithm. After inverting the spin echo train attenuation signal and decoupling its physical features, the SIRT algorithm obtains microscopic moisture characteristic parameters. Based on the feature contribution rule, it selects target features and uses the peak area value of the target features as the core feature variable. The core feature variable is input into the XGBoost model for mapping calculation, thereby outputting the predicted vitality index. After the XGBoost model calculation is completed, the predicted vitality index and the normalized value of the target features are respectively thresholded. Based on the threshold determination result, the hot air blower and the cold air blower are physically closed or opened.

[0009] The working principle of the above-mentioned adaptive control device for drying rice seeds is as follows: Rice seeds to be dried are fed into the drying chamber through the feeding port. The control module controls the hot air fan to run at full load, and hot air rushes into the drying chamber from the air inlet. The rice seeds tumble in a gas-solid fluid state within the drying chamber, rapidly removing moisture from their interior. During this process, the RF coil penetrates the cavity wall of the drying chamber and emits a CPMG pulse sequence, causing the magnetization vector of hydrogen protons inside the rice seeds to precess and refocus. The RF coil collects the spin echo train attenuation signal of hydrogen protons in the rice seeds and transmits the spin echo train attenuation signal to the control module. The SIRT algorithm performs inversion operations on the spin echo train attenuation signal, then decouples the physical features to obtain the microscopic physical phase of the rice seed. A preset integral algorithm is used to decouple the microscopic moisture characteristic parameters corresponding to the microscopic physical phase. The control module, based on a pre-determined target feature rule based on SHAP feature contribution, targets and locks the peak area of ​​strongly bound water and the peak area of ​​bound water as target features. It extracts the peak area values ​​of the strongly bound water peak area and the peak area of ​​bound water as core feature variables. These core feature variables are input into the XGBoost model, which performs mapping calculations, mapping the core feature variables to a macroscopic predicted vigor index in real time. The control module performs threshold monitoring and determination, with thresholds including the safe vigor threshold and the peak area of ​​strongly bound water (A...). 21 ) and the area of ​​the bound water peak (A 22 The critical control thresholds for microscopic physical damage are characterized by comparing the predicted vigor index with the safe vigor threshold, and comparing the peak area of ​​strongly bound water and the normalized value of the peak area of ​​bound water with their corresponding critical control thresholds, thereby analyzing the trend of changes in the cellular microstructure state of rice seeds (i.e., predicting the current state of rice seeds). Based on the comparison results, the control module performs physical closure or opening of the hot air blower and the cold air blower, thereby achieving precise regulation and optimal balance between dehydration efficiency and seed vigor during continuous drying.

[0010] In a preferred embodiment of the present invention, a metal shield is provided around the detection section, and the RF coil is disposed between the metal shield and the periphery of the detection section. By providing a metal shield to enclose the RF coil, it acts as a Faraday cage for electromagnetic shielding.

[0011] Preferably, an interlayer is provided between the RF coil and the periphery of the detection section. By providing the interlayer, a completely independent physical isolation layer is formed without interfering with the magnetic field penetration, preventing the hot air in the drying chamber from causing the permanent magnet magnetic field of the RF coil to drift.

[0012] Preferably, the metal shielding body is equipped with a circulating cooling medium. By providing a circulating cooling medium, rapid heat dissipation is achieved, ensuring the RF coil operates at a constant temperature and eliminating thermal drift.

[0013] Preferably, the in-situ nuclear magnetic resonance detection module further includes a temperature control unit, a circulating pump, and a cooling tank. A cooling channel is provided within the metal shield, and the circulating cooling medium is located within the cooling channel. One end of the circulating pump is connected to the inlet of the cooling channel, and the other end is connected to the cooling tank. The outlet of the cooling channel is connected to the cooling tank. The temperature control unit is used to monitor the temperature of the RF coil in real time. The cooling tank can store the circulating cooling medium and also cool it.

[0014] Preferably, the exhaust end is provided with an exhaust outlet, and the air inlet end is provided with an air inlet. There are multiple air inlets arranged in an array, and both the hot air blower and the cold air blower are connected to the air inlets. Through the underlying control signal, the device can realize rapid physical switching of the air source. During the hot air drying stage, the hot air blower remains fully loaded. When one of the thresholds is met and the thermal inertia of the device needs to be overcome, the hot air blower is rapidly shut down and the cold air blower is switched to full load operation to perform forced physical cooling and tempering. The air enters the drying chamber from the air inlet end. Specifically, hot air or cold air enters the drying chamber from the air inlet array at the bottom of the drying chamber.

[0015] Preferably, the gas-solid fluidized bed drying module further includes a cyclone separator connected to the outlet. Hot air from the hot air blower and cold air from the cold air blower pass over the rice seeds, forming dusty exhaust gas which is discharged from the outlet and enters the cyclone separator for gas-solid separation. Dust is discharged from the bottom of the cyclone separator, and gas is discharged from the top.

[0016] An adaptive control method for rice seed drying based on LF-NMR and XGBoost includes the following steps: (1) Material loading and initial baseline adaptive calibration: Before the formal drying process begins, fresh, untreated rice seeds are fed into the drying chamber through the feeding port. The RF coil performs in-situ LF-NMR baseline scanning on the rice seeds in the drying chamber. The control module extracts the initial strong bound water peak area and the initial bound water peak area of ​​the rice seeds, and substitutes them into the XGBoost model to calculate the initial vigor index and automatically generate a safe vigor threshold. (2) In-situ fluidized bed hot air drying and multidimensional signal acquisition: The control module controls the closure of the high-voltage circuit of the hot air blower, and the hot air blower runs at full load; the rice seeds form a gas-solid fluidized rolling state in the drying chamber, and during the fluidized rolling process, the in-situ nuclear magnetic resonance detection module performs in-situ non-destructive testing according to a set period; the RF radio frequency coil collects the spin echo train attenuation signal. (3) Decoupling of signal inversion and physical features based on SIRT algorithm: After receiving the spin echo train attenuation signal, the control module calls the SIRT algorithm to perform inversion calculation, reconstructs the transverse relaxation time with high signal-to-noise ratio, and obtains the T2 relaxation spectrum. After extracting the T2 relaxation spectrum, the control module divides it into four microphysical phase intervals representing different microphysical phase states through a preset integration algorithm. The four microphysical phase intervals are the strong bound water region, the bound water region, the weak free water region, and the free water region. The signal waveforms in the four microphysical phase intervals are integrated to obtain four corresponding microphysical water characteristic parameters, which are the peak area of ​​strong bound water, the peak area of ​​bound water, the peak area of ​​weak free water, and the peak area of ​​free water. (4) Targeted feature screening: The control module targets and locks the peak area of ​​strongly bound water and the peak area of ​​bound water as target features based on the target feature rules determined in advance based on the contribution of SHAP features. It also extracts the peak area values ​​of strongly bound water and bound water as core feature variables to eliminate redundant noise interference caused by the rapid removal of free water. (5) Nonlinear prediction of XGBoost model: The core feature variables are input into the XGBoost model, which then maps them to a macroscopic predictive vitality index in real time. (6) Threshold determination and physical motion feedback: The control module performs threshold determination, which includes a safety activity threshold, a first critical control threshold representing microscopic physical damage based on the peak area of ​​strongly bound water, and a second critical control threshold representing microscopic physical damage based on the peak area of ​​bound water. When the control module detects a sharp decrease in the normalized value of the strongly bound water peak area and reaches the first critical control threshold, or when the bound water peak area (A) 22 When the normalized value of the indicator drops sharply and reaches the second critical control threshold, or when the macroscopic predicted vitality index falls below the safe vitality threshold, an early warning is triggered, and a forced heat release type slow recovery process is executed. The control module physically disconnects the hot air blower, stopping its operation, and simultaneously physically closes the cold air blower, causing it to run at full load. (7) Self-healing condition: When the temperature control probe detects that the temperature inside the drying chamber has dropped to a safe range, and the control module confirms that the peak area of ​​strong bound water and the normalized value of the peak area of ​​bound water have stopped decaying and have rebounded, exceeding the first critical control threshold and the second critical control threshold respectively, and the predicted vitality index exceeds the safe vitality threshold, the control module physically disconnects the cold air blower and physically closes the hot air blower, entering the next drying cycle until the safe moisture content range is reached.

[0017] Preferably, the control module is a field-programmable gate array (FPGA) or an industrial computer equipped with a high-performance microprocessor. After completing the XGBoost model calculation and threshold determination, the industrial computer outputs low-level digital output commands through the I / O interface. The solid-state relay (SSR) or PLC control cabinet electrically connected to the I / O interface receives the commands and directly performs physical closing or opening of the high-voltage circuits of the hot air blower and cold air blower.

[0018] Preferably, in steps (2) and (3), the spin echo train attenuation signal is converted into a digital signal by A / D conversion and transmitted to an industrial computer. The field programmable gate array in the industrial computer or the high-performance microprocessor receives the digital signal and calls the SIRT algorithm to perform inversion calculation.

[0019] Preferably, in step (3), the relaxation time interval of the strongly bound water region is 0.01~0.3ms, the relaxation time interval of the bound water region is 0.3~8ms, the relaxation time interval of the weakly free water region is 8~76ms, and the relaxation time interval of the free water region is >76ms.

[0020] Compared with the prior art, the present invention has the following advantages: The adaptive control device and method for rice seed drying in this invention acquires low-field nuclear magnetic resonance (LF-NMR) signals through an in-situ nuclear magnetic resonance detection module, obtains core feature variables through the SIRT algorithm, performs high-precision nonlinear prediction of rice seed vigor through the XGBoost model, and implements physical closure or opening of the hot air blower and cold air blower by establishing threshold judgment. In the continuous drying process, it achieves precise control and optimal balance between dehydration efficiency and seed vigor, solving the problem of decreased rice seed vigor caused by the inability to detect microstructure collapse and loss of macromolecular binding sites in existing drying processes. It has an in-situ online feedback mechanism and a non-destructive online feedback mechanism, significantly improving the accuracy and robustness of the algorithm and improving the quality of rice seed drying. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of an adaptive control device for rice seed drying based on LF-NMR and XGBoost, as described in this invention.

[0022] Figure 2 This is a schematic diagram of the drying chamber in this invention. Detailed Implementation

[0023] To enable those skilled in the art to fully understand the technical solutions of the present invention, the present invention will be further described below in conjunction with embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0024] Example 1 See Figures 1-2 This embodiment discloses an adaptive control device for rice seed drying based on LF-NMR and XGBoost, including a gas-solid fluidized bed drying module, an in-situ nuclear magnetic resonance detection module, and a control module 1.

[0025] See Figures 1-2 The gas-solid fluidized bed drying module includes a drying chamber 2, a hot air blower 3, and a cold air blower 4. The drying chamber 2 is provided with a feeding port 5 and a discharging port 6. The drying chamber 2 includes an exhaust end 2-1 located at the top, a detection section 2-3 located in the middle, and an air inlet end 2-2 located at the bottom. The hot air blower 3 and the cold air blower 4 are respectively connected to the air inlet end 2-2.

[0026] See Figures 1-2 The in-situ nuclear magnetic resonance (NMR) detection module includes an RF coil 7 and an NMR control system 8, which is connected to the RF coil 7. The RF coil transmits a CPMG pulse sequence through the cavity wall of the drying chamber 2. The RF coil collects the spin echo train attenuation (FID) signal of hydrogen protons in rice seeds, and outputs a digital signal through a high-frequency analog-to-digital (A / D) converter circuit. The NMR control system 8 is used to regulate the RF coil 7.

[0027] See Figures 1-2 Temperature control probes 9 are installed at both the air inlet 2-2 and the air outlet 2-1 of the drying chamber 2. The control module 1 is connected to the temperature control probes 9, the nuclear magnetic resonance control system 8, the hot air blower 3, and the cold air blower 4. The digital signals output by the temperature control probes 9 and the RF coil 7 are fed into a high-speed bus and sent to the control module 1. The two temperature control probes 9 are located at the air inlet 2-2 and the air outlet 2-1 respectively, in order to more accurately control the temperature inside the drying chamber 2.

[0028] See Figures 1-2 The control module 1 incorporates the SIRT algorithm. After inverting the spin echo train attenuation signal (FID) and decoupling the physical features, the SIRT algorithm obtains the microscopic physical phase range of the rice seed, thereby obtaining the microscopic moisture characteristic parameters corresponding to the microscopic physical phase range. Based on the feature contribution rule, the target feature is selected from the microscopic moisture characteristic parameters, and the peak area value of the target feature is used as the core feature variable. The core feature variable is input into the XGBoost model for mapping calculation, thereby outputting the predicted vigor index. After the XGBoost model calculation is completed, the predicted vigor index and the normalized value of the target feature are respectively thresholded to predict the current state of the rice seed. According to the threshold determination result, the hot air fan 3 and the cold air fan 4 are physically closed or opened.

[0029] See Figures 1-2 A metal shield 10 is provided around the detection section 2-3, and the RF coil is disposed between the metal shield 10 and the periphery of the detection section 2-3. By providing the metal shield 10 to enclose the RF coil, it acts as a Faraday cage for electromagnetic shielding.

[0030] See Figures 1-2 An interlayer can be provided between the RF coil and the periphery of the detection section 2-3. By providing an interlayer, a completely independent physical isolation layer is formed without interfering with the magnetic field penetration, preventing the hot air in the drying chamber 2 from causing the permanent magnet magnetic field of the RF coil 7 to drift.

[0031] See Figures 1-2 The metal shield 10 is equipped with a circulating cooling medium. By providing a circulating cooling medium, rapid heat dissipation is achieved, ensuring that the RF coil operates at a constant temperature and eliminating thermal drift.

[0032] See Figures 1-2 The circulating cooling medium is fluorinated oil.

[0033] See Figures 1-2 The in-situ nuclear magnetic resonance detection module also includes a temperature control unit 11, a circulating pump 12, and a cooling tank 13. The metal shield 10 has a cooling channel within it, and the circulating cooling medium is located in the cooling channel. One end of the circulating pump 12 is connected to the inlet of the cooling channel, and the other end is connected to the cooling tank 13. The outlet of the cooling channel is connected to the cooling tank 13. The temperature control unit 11 is used to monitor the temperature of the RF coil in real time. The cooling tank 13 can store the circulating cooling medium and also cool it.

[0034] See Figures 1-2 The exhaust end 2-1 is provided with an exhaust outlet 2-11, and the air inlet end 2-2 is provided with an air inlet 2-21. There are multiple air inlets 2-21, which are arranged in an array. The hot air blower 3 and the cold air blower 4 are both connected to the air inlets 2-21. Through the underlying control signal, the equipment can realize the rapid physical switching of the air source. During the hot air drying stage, the hot air blower 3 is kept running at full load. When one of the thresholds is met and it is necessary to overcome the thermal inertia of the equipment, the hot air blower 3 is rapidly shut down and the cold air blower 4 is switched to full load operation to perform forced physical cooling and slow cooling. Specifically, wind or cold air enters the interior of the drying chamber 2 from the array of air inlets 2-21 at the bottom of the drying chamber 2.

[0035] See Figures 1-2The gas-solid fluidized bed drying module also includes a cyclone separator connected to the outlet 2-11. Hot air from the hot air blower 3 and cold air from the cold air blower 4 pass over the rice seeds, forming dusty exhaust gas which is discharged from outlet 2-11 and enters the cyclone separator for gas-solid separation. Dust is discharged from the bottom of the cyclone separator, and gas is discharged from the top.

[0036] See Figures 1-2 The working principle of the above-mentioned adaptive control device for drying rice seeds is as follows: Rice seeds to be dried are fed into the drying chamber 2 through the feeding port 5. The control module 1 controls the hot air blower 3 to run at full load, and hot air flows into the drying chamber 2 from the air inlet 2-2. The rice seeds tumble in a gas-solid fluid state within the drying chamber 2, rapidly removing moisture from their interior. During this process, the RF coil penetrates the cavity wall of the drying chamber 2 and emits a CPMG (Carr-Purcell-Meiboom-Gill) pulse sequence, causing the magnetization vector of hydrogen protons inside the rice seeds to precess and refocus. A frequency coil collects the spin echo train attenuation (FID) signal of hydrogen protons in rice seeds and transmits it to control module 1. The SIRT algorithm performs inversion operations on the FID signal, followed by physical feature decoupling to obtain the microscopic physical phase of the rice seed. A preset integration algorithm is used to decouple the microscopic moisture characteristic parameters corresponding to the microscopic physical phase. The control module, based on a pre-determined targeting feature rule based on SHAP feature contribution, targets and locks the area of ​​the strongly bound water peak (A). 21 ) and the area of ​​the bound water peak (A 22 ) as the target feature, and extract the peak area of ​​strongly bound water (A) 21 ) and the area of ​​the bound water peak (A 22 The peak area of ​​the strong bound water is used as the core feature variable. This core feature variable is input into the XGBoost model, which performs mapping calculations to map the core feature variable to a macroscopic predicted vitality index (VI) in real time. Control module 1 monitors and determines thresholds, including the safe vitality threshold and the peak area of ​​the strong bound water (A). 21 ) and the area of ​​the bound water peak (A 22 The critical control thresholds for microscopic physical damage are characterized by comparing the predicted viability index (VI) with the safe viability threshold, and the peak area of ​​strongly bound water (A) is used to characterize the critical control thresholds for microscopic physical damage. 21 ) and combined water peak area (A 22The normalized values ​​of the rice seeds are compared with their corresponding critical control thresholds to analyze the trend of changes in the cellular microstructure of the rice seeds (i.e., to predict the current state of the rice seeds). Based on the comparison results, the control module 1 physically closes or opens the hot air blower 3 and the cold air blower 4, thereby achieving precise control and optimal balance between dehydration efficiency and seed vitality during continuous drying. After drying, the rice seeds are discharged from the outlet 6 of the drying chamber 2.

[0037] See Figures 1-2 This embodiment also discloses an adaptive control method for rice seed drying based on LF-NMR and XGBoost, including the following steps: (1) Material loading and initial baseline adaptive calibration (eliminating physiological differences in rice seeds): Before the formal drying process begins, fresh, untreated rice seeds are fed into the drying chamber 2 through the feeding port 5. The nuclear magnetic resonance control system 8 controls the RF coil to emit a CPMG pulse sequence. The CPMG pulse sequence penetrates the chamber wall of the drying chamber 2, performing an in-situ LF-NMR baseline scan on the rice seeds inside the drying chamber 2. The control module 1 extracts the initial strong bound water peak area (A) of the rice seeds. 21 ) and the initial bound water peak area (A 22 The XGBoost model is used to calculate the initial vigor index (VI), and a safe vigor threshold suitable for the current batch of rice seeds is automatically generated, so that the control method can be compatible with different varieties of rice seeds (such as japonica rice and indica rice) and physiological baseline differences caused by different initial moisture contents. (2) In-situ fluidized bed hot air drying and multidimensional signal acquisition: The control module 1 controls the closure of the high-voltage circuit of the hot air blower 3, and the hot air blower 3 runs at full load to dry the rice seeds. The hot air generated by the hot air blower 3 enters the drying chamber 2 from the air inlet 2-2. The rice seeds form a gas-solid fluidized tumbling state in the drying chamber 2, which quickly removes the moisture inside the rice seeds. During the fluidized tumbling (drying) process, the in-situ nuclear magnetic resonance detection module performs in-situ non-destructive testing according to a set period. During the in-situ non-destructive testing, the RF coil penetrates the cavity wall of the drying chamber 2 to emit a CPMG pulse sequence, causing the hydrogen protons in the rice seeds to undergo spin-spin relaxation attenuation. Then, the spin echo train attenuation signal (FID) of the hydrogen protons in the rice seeds is collected.

[0038] (3) Decoupling of signal inversion and physical features based on SIRT algorithm: After receiving the spin echo train attenuation signal (FID) (converted to a digital signal), the control module 1, due to the presence of fan vibration and electromagnetic interference in the industrial environment, calls the SIRT (Synchronous Iterative Reconstruction) algorithm for inversion. The SIRT algorithm effectively suppresses the background noise of the industrial environment through residual iteration, reconstructing the spin echo train attenuation signal (FID) with background noise, i.e., the original noisy time-domain signal, into a transverse relaxation time with a high signal-to-noise ratio and specific physical meaning, thus obtaining the T2 relaxation spectrum. Based on the physical characteristics of the hydrogen proton's degrees of freedom, the control module extracts the T2 relaxation spectrum and divides it into four microscopic physical phase regions representing different microscopic physical phase states using a preset integration algorithm. The four microscopic physical phase state regions are: the strongly bound water region T... 21 Combined water area T 22 Weak free water region T 23 and free water zone T 24 In terms of physical mechanisms, strongly bound water regions represent interfacial water that is tightly bound to biological macromolecules, while bound water regions represent water that is constrained by the integrity of the cell wall's multi-layered alternating structure; among them, The relaxation time intervals for the strongly bound water region and the weakly free water region are 0.01–0.3 ms, 0.3–8 ms, 8–76 ms, and >76 ms respectively. The signal waveforms within these four microscopic physical phase intervals are integrated to obtain the corresponding microscopic water characteristic parameters. The microscopic water characteristic parameter obtained for the strongly bound water region is the area of ​​the strongly bound water peak (A0). 21 The microscopic water characteristic parameters obtained from the bound water region are the bound water peak area (A). 22 The microscopic water characteristic parameter obtained from the weak free water region is the weak free water peak area (A). 23 The microscopic water characteristic parameter obtained from the free water region is the free water peak area (A). 24 ); (4) Targeted feature screening: To eliminate redundant noise caused by drastic fluctuations in free water during the drying process and ensure the robustness of industrial control, control module 1, based on SHAP (SHapley Additive exPlanations) feature contribution weights and pre-determined targeted feature rules, automatically eliminates weak free water regions and their corresponding microscopic moisture characteristic parameters that have little impact on vitality, and targets and locks the peak area of ​​strongly bound water (A). 21 ) and the area of ​​the bound water peak (A 22 ) as the target feature, and extract the peak area of ​​strongly bound water (A) 21 ) and the area of ​​the bound water peak (A 22 The peak area value is used as the core feature variable; (5) Nonlinear prediction of XGBoost model: The core feature variables are input into the XGBoost model, which maps the core feature variables to the macroscopic predictive vitality index (VI) in real time. (6) Threshold determination and physical motion feedback: Control module 1 determines the threshold based on the contribution of SHAP features. The threshold includes the safety activity threshold and the peak area of ​​strongly bound water (A). 21 The first critical control threshold for characterizing microscopic physical damage and the bound water peak area (A) 22 The second critical control threshold characterizing microscopic physical damage; The determination process is as follows: When control module 1 detects the area of ​​the strongly bound water peak (A) 21 When the normalized value of ) drops sharply and reaches the first critical control threshold (e.g., the first critical control threshold - 0.5), or when combined with the water peak area (A 22 When the normalized value of ) drops sharply and reaches the second critical control threshold, the SHAP model shows that the peak area of ​​strongly bound water (A) 21 ) or combined with the peak water area (A 22 When the contribution of the rice seed to the vitality index drops sharply (a serious negative impact occurs), or when the macroscopic predicted vitality index (VI) falls below the safe vitality threshold (i.e., falls below the safety bottom line), it indicates that the continuous 55℃ hot air has caused the cell microstructure of the rice seed to begin to collapse, and the macromolecular binding sites are about to be irreversibly lost. After triggering one of the above thresholds, an early warning is triggered, and a forced heat release type slow tempering process is executed: the control module 1 physically disconnects the hot air blower 3, the hot air blower 3 stops running, and simultaneously physically closes the cold air blower 4, the cold air blower 4 runs at full load; cold air rushes in from the air inlet 2-2 of the drying chamber 2, forcibly releasing the residual heat accumulated in the drying chamber 2 and the rice seed, and executing the "slow tempering" process; at this stage, the hot air stops, surface evaporation is suppressed, and under the water potential gradient driven by the high water potential inside and the low water potential on the surface, the bound water of the multi-layer alternating structure of the cell wall diffuses in the opposite direction to the surface, effectively dissipating the internal humid heat stress and preventing mechanical damage to the cell membrane caused by glass transition; (7) Self-healing condition: When the temperature control probe 9 detects that the temperature inside the drying chamber 2 has dropped to a safe range, such as below 25 degrees Celsius, and the control module 1 confirms the area of ​​the strong bound water peak (A... 21 ) and the area of ​​the bound water peak (A 22 The normalized value of ) stops decaying and begins to rise again (e.g., the area of ​​the strongly bound water peak (A) 21 ) and the area of ​​the bound water peak (A 22 The normalized values ​​of the predicted vitality index (VI) rise back to their respective critical control thresholds, and the predicted vitality index (VI) exceeds the safe vitality threshold. Specifically, the area of ​​the strong bound water peak (A) rises. 21When the temperature rises to a high level greater than 0.5, it has a significant positive contribution to the vitality of rice seeds. It is determined that the microstructure stress has been relieved. The control module 1 physically disconnects the cold air blower 4 and physically closes the hot air blower 3. The hot air blower 3 is then turned on again at a low power (e.g., switched to a mild parameter of 45℃). The next drying cycle is entered (the above steps (3)-(7) are repeated, and threshold judgment and physical action feedback are performed in each cycle) until the rice seeds reach a safe moisture content range and the drying is completed.

[0039] See Figures 1-2 The control module 1 is a field-programmable gate array (FPGA) or an industrial computer (IPC) equipped with a high-performance microprocessor (CPU). After completing the XGBoost model calculation and threshold determination, the industrial computer (IPC) outputs low-level digital switching instructions through the I / O interface. The solid-state relay (SSR) or PLC control cabinet electrically connected to the I / O interface receives the instructions and directly performs physical closing or opening of the high-voltage circuits of the hot air blower 3 and the cold air blower 4.

[0040] See Figures 1-2 In step (2), the hot air blower 3 is running at full load, the set temperature is 55℃, and the rice seeds in the drying chamber 2 are subjected to in-situ non-destructive testing according to the set cycle (such as every 0.5h), thereby collecting the spin echo train attenuation signal (FID). The CPMG pulse sequence emitted by the RF coil has a main frequency of 21 kHz, and the specific parameters are: 90° pulse time 19.00μs, 180° pulse time 36.00μs, echo time 0.15 ms, number of echoes 3000, and repeated sampling time 1000 ms.

[0041] See Figures 1-2 In steps (2) and (3), the spin echo train attenuation signal (FID) is converted into a digital signal by A / D conversion and transmitted to the industrial computer (IPC). The field programmable gate array in the industrial computer (IPC) or equipped with a high-performance microprocessor receives the digital signal and calls the SIRT algorithm to perform inversion calculation.

[0042] See Figures 1-2 In step (5), the specific steps for nonlinear prediction of the XGBoost model are as follows: extracting the area of ​​the strongly bound water peak (A 21 ) and the area of ​​the bound water peak (A 22The peak area value is input into the pre-trained XGBoost model in the memory of the control module 1. This XGBoost model uses an additive model to train multiple weak classifiers to fit the residuals, mapping the nonlinear characteristics of microscopic moisture parameters to macroscopic physical representations. After 5-fold cross-validation, the XGBoost model exhibits extremely high robustness on the test set. The test set determination coefficient (R²) calculated by the XGBoost model is [value missing]. 2 p) up to 0.986; Root Mean Square Error (RMSE) P The value is 6.7844; the XGBoost model outputs the current macroscopic predicted vigor index (VI) of rice seeds in real time.

[0043] In step (6), the mathematical model for threshold determination is: Among them, S control These are low-level switch physical action commands used to control the physical opening or closing of hot air blowers and cold air blowers; VI safe It is the safety vitality threshold (safety baseline) dynamically calibrated in step (1); A 21norm It is the area of ​​strongly bound water peaks (A 21 The normalized value at the current moment; A 22norm It is the combined water peak area (A) 22 The normalized value at the current moment; λ SHAP1 It is the area of ​​strongly bound water peaks (A 21 The first critical control threshold characterizing microscopic physical damage, λ SHAP2 It is the combined water peak area (A) 22 The second critical control threshold characterizing microscopic physical damage is derived from the model's SHAP dependency plot analysis, selecting the inflection point where the SHAP contribution value rapidly declines from positive to negative. During the drying (dehydration) process, when A... 21norm Reduced to λ SHAP1 When this happens, it means that the interfacial water that is tightly bound to biological macromolecules is disrupted; when A 22norm Reduced to λ SHAP2 When this happens, it means that the cell wall water has been excessively stripped away, which foreshadows the impending formation of micro-cracks and irreversible thermal damage.

[0044] Physical motion feedback mechanism: When S is calculated controlWhen the value is 0, the industrial computer (IPC) outputs a hold command to the solid-state relay or PLC control cabinet, keeping the hot air blower 3 running. The drying chamber 2 maintains a high temperature (e.g., 55°C) hot air heating and blowing physical state to ensure the effective moisture diffusion coefficient (Deff up to 2.03 × 10⁻⁶) is 2.03 × 10⁻⁶. -11 m 2 The extremely high dehydration efficiency is achieved by ( / s).

[0045] When S is calculated control When =1, (i.e., the area of ​​the strongly bound water peak (A) 21 The normalized value of ) drops sharply and reaches the first critical control threshold or the combined water peak area (A 22 When the normalized value of the cell wall bound water drops sharply and reaches the second critical control threshold, or when the macroscopic predicted vitality index (VI) falls below the safe vitality threshold (predicted vitality index (VI) less than or equal to the safe vitality threshold, a precursor to irreversible detachment of cell wall bound water), the Industrial Computer (IPC) initiates a forced heat release tempering process. ① Cut-off and forced discharge: The solid-state relay or PLC control cabinet cuts off the operation of the hot air blower 3, and simultaneously turns on the cold air blower 4, keeping the cold air blower 4 running at full load; this process quickly removes the accumulated residual heat in the drying chamber 2. This action quickly removes the accumulated residual heat in the drying chamber, overcomes the thermal inertia of the large metal body of the drying chamber, and prevents the temperature from "surging" at the moment of shutdown, which could cause secondary serious thermal damage.

[0046] ② Physical settling and moisture equalization: After the temperature inside the drying chamber drops sharply to a safe range (safe room temperature, 25℃), reduce the speed of the cooling fan and enter the physical settling period. Utilizing the water potential difference between the high water potential core and the low water potential epidermis of the rice seed, the bound water inside is driven to permeate in reverse to the surface, physically dissolving the internal moisture gradient and hygrothermal stress.

[0047] ③ Self-healing and recovery: After the physical settling and humidification process is completed, the in-situ nuclear magnetic resonance detection module reaches the next cycle and collects the spin echo train decay signal (FID) of hydrogen protons in rice seeds. When the area of ​​the strong bound water peak (A) is detected... 21 The normalized value of ) and the bound water peak area (A) 22 When the normalized value of the indicator stops declining or steadily recovers, and the predicted vitality index (VI) exceeds the safe vitality threshold, it proves that the microstructural stress has been relieved. The solid-state relay or PLC control cabinet recloses the high-voltage circuit of the hot air blower and adjusts it to a mild power (such as 45°C) to continue dehydration until the safe termination moisture content is reached.

[0048] The standard germination test (GB / T3543.4-1995) verification results show that although the traditional 55℃ constant temperature drying dehydrates quickly, it causes severe heat damage, and the seed germination rate drops to 41.58%. However, this embodiment adopts an adaptive control strategy (i.e., automatic switching to forced heat release type slow tempering process), which successfully maintains the integrity of the cell membrane system and microscopic water phase while ensuring high drying efficiency. It significantly improves the germination rate and maintains it at 92.33%, and the vigor index (VI) reaches 156.69, achieving the optimal solution for rice seed drying energy efficiency and physiological quality in industrial continuous production.

[0049] See Figures 1-2 The adaptive control device and method for rice seed drying in this embodiment overcomes the limitations of traditional macroscopic moisture content control, which cannot characterize the water distribution and thermal damage of subcellular structures. It acquires low-field nuclear magnetic resonance (LF-NMR) signals through an in-situ nuclear magnetic resonance detection module, utilizes the SHAP interpretability mechanism to target and eliminate interference from easily escaping free water and weakly free water, and obtains the area of ​​strongly bound water peaks (A) that plays a decisive role in cell wall structure and macromolecular integrity through the SIRT algorithm. 21 ) and the area of ​​the bound water peak (A 22 The peak area of ​​the peak area is used as the core feature variable input to the XGBoost model. The XGBoost model is used to make high-precision nonlinear predictions of rice seed vigor. By establishing a threshold judgment based on "the macroscopic predicted vigor index (VI) falling below the safety line" or "the microscopic moisture characteristic parameter touching the critical inflection point of damage", the hot air fan 3 and the cold air fan 4 are physically closed or opened. The hot air fan 3 is automatically controlled to forcibly cut off the heat source and release heat, and enter a slow-release state to dissipate the internal moisture gradient and hygrothermal stress. In the continuous drying process, the dehydration efficiency and seed vigor are precisely controlled and optimally balanced. This solves the problem of rice seed vigor decline caused by the inability to detect the collapse of microstructure and loss of macromolecular binding sites in the existing drying process. It has an in-situ online feedback mechanism and a non-destructive online feedback mechanism. The algorithm accuracy and robustness are significantly improved, and the quality of rice seed drying is improved.

[0050] The adaptive control device and method for rice seed drying in this embodiment significantly improves algorithm accuracy and robustness. The XGBoost strategy used improves the prediction accuracy to R0. 2 p =0.986, and the root mean square error (RMSEP) decreased to 6.7844. This is thanks to the SHAP feature selection, which eliminated the interference of free water noise.

[0051] The adaptive control device and method for rice seed drying in this embodiment effectively resolves the contradiction between drying efficiency and quality. Experimental data shows that traditional 55℃ constant temperature drying causes the germination rate to drop to 41.58%. The adaptive control device and method for rice seed drying in this embodiment automatically introduces a tempering process when a risk of microscopic damage is detected, maintaining a high drying efficiency (effective diffusion coefficient Deff is better than natural drying) while keeping the seed germination rate above 92.33%.

[0052] The adaptive control device and method for rice seed drying in this embodiment are based on advanced control of microscopic mechanisms. SHAP analysis clarifies that "strongly bound water" and "bound water" are key physical quantities that determine viability. Compared with the lagging germination experiment, monitoring the inflection points of these two physical quantities can detect microscopic changes in cell structure in advance, thus achieving advanced control.

[0053] Example 2 The other structures in this embodiment are the same as those in Embodiment 1. The difference is that the LF-NMR detection coil adopts an isolated bypass sampling, that is, the RF coil 7 is independently set in a shockproof constant temperature and humidity detection control cabinet; the side wall of the drying chamber 2 is connected to the automatic sampling bypass, and a small amount of fluidized rice seed material is guided into the detection chamber (such as a 12mm diameter test tube) according to a preset cycle to complete signal extraction.

[0054] Example 3 The other structures in this embodiment are the same as those in Embodiment 1, except that the circulating cooling medium is a deeply dried, constant-temperature cold air stream, or other insulating coolant that does not contain free hydrogen protons. As long as the circulating cooling medium can physically cool and insulate the RF coil and the LF-NMR detection coil, and does not interfere with the penetration of the RF magnetic field of the RF coil or the acquisition of hydrogen proton signals inside the rice seed, it is acceptable.

[0055] Example 4 The other structures in this embodiment are the same as those in Embodiment 1. The difference is that the cold air blower 4 is replaced by an electric dehumidification damper. When one of the thresholds is triggered, the hot air blower 3 is turned off, the large-diameter electric dehumidification damper is fully opened, and the blower of the circulating electric dehumidification damper is kept running at full load to forcibly draw outside room temperature air into the drying chamber 2. The above structure omits the independent cold air blower 4 compressor component. The purpose of forcibly releasing residual heat and preventing the collapse of cell microstructure can also be achieved by relying on a large volume of room temperature air.

[0056] The above are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above content. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. An adaptive control device for rice seed drying based on LF-NMR and XGBoost, characterized in that, It includes a gas-solid fluidized bed drying module, an in-situ nuclear magnetic resonance detection module, and a control module; among which, The gas-solid fluidized bed drying module includes a drying chamber, a hot air blower, and a cold air blower. The drying chamber is provided with a feeding port and a discharging port. The drying chamber includes an exhaust end at the top, a detection section in the middle, and an air inlet at the bottom. The hot air blower and the cold air blower are respectively connected to the air inlet. The in-situ nuclear magnetic resonance detection module includes an RF coil and a nuclear magnetic resonance control system, wherein the nuclear magnetic resonance control system is connected to the RF coil; the nuclear magnetic resonance control system controls the RF coil to emit a CPMG pulse sequence to collect the spin echo train attenuation signal of hydrogen protons in rice seeds; The air inlet and air outlet of the drying chamber are both equipped with temperature control probes; the control module is connected to the temperature control probes, the nuclear magnetic resonance control system, the hot air blower and the cold air blower respectively. The control module incorporates the SIRT algorithm. After inverting the spin echo train attenuation signal and decoupling its physical features, the SIRT algorithm obtains microscopic moisture characteristic parameters. Based on the feature contribution rule, it selects target features and uses the peak area value of the target features as the core feature variable. The core feature variable is input into the XGBoost model for mapping calculation, thereby outputting the predicted vitality index. After the XGBoost model calculation is completed, the predicted vitality index and the normalized value of the target features are respectively thresholded. Based on the threshold determination result, the hot air blower and the cold air blower are physically closed or opened.

2. The adaptive control device for rice seed drying according to claim 1, characterized in that, A metal shield is provided around the detection section, and the RF coil is located between the metal shield and the periphery of the detection section.

3. The adaptive control device for rice seed drying according to claim 2, characterized in that, The metal shield is equipped with a circulating cooling medium.

4. The adaptive control device for rice seed drying according to claim 3, characterized in that, The in-situ nuclear magnetic resonance detection module also includes a temperature control unit, a circulating pump, and a cooling tank. The metal shield is provided with a cooling channel, and the circulating cooling medium is located in the cooling channel. One end of the circulating pump is connected to the inlet of the cooling channel, and the other end is connected to the cooling tank. The outlet of the cooling channel is connected to the cooling tank. The temperature control unit is used to monitor the temperature of the RF coil in real time.

5. The adaptive control device for rice seed drying according to claim 1, characterized in that, The exhaust end is provided with an exhaust outlet, and the air inlet end is provided with an air inlet. There are multiple air inlets, which are arranged in an array. The hot air blower and the cold air blower are both connected to the air inlets.

6. The adaptive control device for rice seed drying according to claim 5, characterized in that, The gas-solid fluidized bed drying module also includes a cyclone separator, which is connected to the outlet.

7. An adaptive control method for rice seed drying based on LF-NMR and XGBoost, characterized in that, The adaptive control method for rice seed drying is applied to the adaptive control device for rice seed drying as described in any one of claims 1-6, and the adaptive control method for rice seed drying includes the following steps: (1) Material loading and initial baseline adaptive calibration: Before the formal drying process begins, fresh, untreated rice seeds are fed into the drying chamber through the feeding port. The RF coil performs in-situ LF-NMR baseline scanning on the rice seeds in the drying chamber. The control module extracts the initial strong bound water peak area and the initial bound water peak area of ​​the rice seeds, and substitutes them into the XGBoost model to calculate the initial vigor index and automatically generate a safe vigor threshold. (2) In-situ fluidized bed hot air drying and multidimensional signal acquisition: The control module controls the closure of the high-voltage circuit of the hot air blower, and the hot air blower runs at full load; the rice seeds form a gas-solid fluidized rolling state in the drying chamber, and during the fluidized rolling process, the in-situ nuclear magnetic resonance detection module performs in-situ non-destructive testing according to a set period; the RF radio frequency coil collects the spin echo train attenuation signal. (3) Decoupling of signal inversion and physical features based on SIRT algorithm: After receiving the spin echo train attenuation signal, the control module calls the SIRT algorithm to perform inversion calculation, reconstructs the transverse relaxation time with high signal-to-noise ratio, and obtains the T2 relaxation spectrum. After extracting the T2 relaxation spectrum, the control module divides it into four microphysical phase intervals representing different microphysical phase states through a preset integration algorithm. The four microphysical phase intervals are the strong bound water region, the bound water region, the weak free water region, and the free water region. The signal waveforms in the four microphysical phase intervals are integrated to obtain four corresponding microphysical water characteristic parameters, which are the peak area of ​​strong bound water, the peak area of ​​bound water, the peak area of ​​weak free water, and the peak area of ​​free water. (4) Targeted feature screening: The control module targets and locks the peak area of ​​strongly bound water and the peak area of ​​bound water as target features based on the target feature rules determined in advance based on the contribution of SHAP features, and extracts the peak area values ​​of strongly bound water and the peak area of ​​bound water as core feature variables. (5) Nonlinear prediction of XGBoost model: The core feature variables are input into the XGBoost model, which then maps them to a macroscopic predictive vitality index in real time. (6) Threshold determination and physical motion feedback: The control module performs threshold determination, which includes a safety activity threshold, a first critical control threshold representing microscopic physical damage based on the peak area of ​​strongly bound water, and a second critical control threshold representing microscopic physical damage based on the peak area of ​​bound water. When the control module detects a sharp drop in the normalized value of the strong bound water peak area and reaches the first critical control threshold, or a sharp drop in the normalized value of the bound water peak area and reaches the second critical control threshold, or when the macroscopic predicted vitality index falls below the safe vitality threshold, an early warning is triggered, and a forced heat release type slow-release process is executed. The control module physically disconnects the hot air blower, stopping its operation, and simultaneously physically closes the cold air blower, causing it to run at full load. (7) Self-healing condition: When the temperature control probe detects that the temperature inside the drying chamber has dropped to a safe range, and the control module confirms that the peak area of ​​strong bound water and the normalized value of the peak area of ​​bound water have stopped decaying and have rebounded, exceeding the first critical control threshold and the second critical control threshold respectively, and the predicted vitality index exceeds the safe vitality threshold, the control module physically disconnects the cold air blower and physically closes the hot air blower, entering the next drying cycle until the safe moisture content range is reached.

8. The adaptive control method for rice seed drying according to claim 7, characterized in that, The control module is a field-programmable gate array or an industrial computer equipped with a high-performance microprocessor.

9. The adaptive control method for rice seed drying according to claim 8, characterized in that, In steps (2) and (3), the spin echo train attenuation signal is converted into a digital signal by A / D conversion and transmitted to the industrial computer. The field programmable gate array in the industrial computer or the high-performance microprocessor receives the digital signal and calls the SIRT algorithm to perform inversion calculation.

10. The adaptive control method for rice seed drying according to claim 7, characterized in that, In step (3), the relaxation time interval of the strongly bound water region is 0.01~0.3ms, the relaxation time interval of the bound water region is 0.3~8ms, the relaxation time interval of the weakly free water region is 8~76ms, and the relaxation time interval of the free water region is >76ms.

Images