A flame soil disinfection self-adaptive temperature control system and method

Abstract

The application provides a kind of flame soil disinfection adaptive temperature control system and method, system obtains basic gas flow by fusing latent heat of vaporization multiphase flow physical feedforward model, to overcome large time lag response deviation, utilize the digital odometer alignment algorithm based on integral dead zone tracking, accurately align the peak value of soil surface and decay temperature in the same area, extract space-time heat decay tensor, then combine with each sensor acquisition data, use the asymmetric loss function based on the boundary of agricultural lethal temperature to train LSTM network to ensure that "prevent underburning" is the absolute priority, finally, the dynamic correction coefficient is obtained by LSTM network inversion reasoning;The expected gas flow command is obtained by adaptively weighting the dynamic correction coefficient and the basic gas flow, and the gas flow is adjusted by the control actuator accordingly, to realize adaptive temperature control.The application realizes accurate servo control of invisible deep layer temperature under complex and harsh field conditions, significantly improves the reliability and energy utilization rate of soil disinfection.

Description

Technical Field

[0001] This invention belongs to the field of smart agricultural equipment and automatic control technology, and in particular relates to an adaptive temperature control system and method for flame soil disinfection. Background Technology

[0002] In modern ecological agricultural production, flame soil disinfection, as an environmentally friendly physical technology for weeding and sterilization, has attracted much attention due to its advantages such as no chemical residues and rapid effectiveness. This technology uses high-temperature flames to instantly act on the soil, effectively killing weed seeds, pathogens, and insect eggs in the shallow soil layer.

[0003] However, existing flame-based soil disinfection equipment faces significant technical bottlenecks in practical operations. First, soil differs from homogeneous industrial materials; it contains a large amount of moisture. When a multi-nozzle flame array acts on the surface, shallow moisture rapidly vaporizes, absorbing a large amount of latent heat. This causes traditional energy feedforward models based on a single specific heat capacity to be severely distorted in moist soil. Second, when flame-based disinfection equipment operates in unstructured fields, the vehicle speed and nozzle height fluctuate frequently due to surface undulations. Traditional fixed-time-step sampling algorithms or methods using instantaneous vehicle speed feature calculations cannot accurately depict the real spatial heat conduction relaxation process. Finally, traditional deep learning typically aims to minimize mean squared error, indiscriminately penalizing both "positive errors" (overburning) and "negative errors" (underburning). However, in agricultural disinfection scenarios, failure to reach the critical point for pest and disease mortality in the deep soil can lead to complete operational failure, while overburning only causes partial energy loss. Therefore, using general models and judgment mechanisms easily leads to frequent underburning.

[0004] Existing thermal control fields (such as industrial furnaces and metal heat treatment) have widely adopted composite control architectures of feedforward + feedback or AI + PID. However, those skilled in the art face insurmountable physical barriers when attempting to transplant these general architectures to flame and soil sterilization operations.

[0005] Industrial heating objects are mostly homogeneous solids (such as metals and plastics) with relatively constant specific heat capacity. However, farmland soil is a porous medium with solid, liquid, and gas phases. When subjected to the instantaneous impact of a high-temperature flame, shallow water rapidly boils and vaporizes, absorbing a large amount of latent heat of vaporization, resulting in a strong nonlinear cutoff effect in heat conduction. General industrial feedforward models lack a thermal compensation mechanism for this water phase change, inevitably leading to severe prediction distortion. Furthermore, industrial heat sources are typically in closed or semi-closed steady-state environments, while agricultural machinery operations occur in open environments with splashing mud and violent shaking. Degradation protection in general architectures usually relies solely on electrical signal diagnosis such as sensor disconnection or power loss, failing to recognize physical failures such as sensors being obscured by mud but electrically functioning normally. This easily leads to the continuous execution of erroneous commands.

[0006] Therefore, developing a temperature control system capable of monitoring soil thermal state and self-optimizing is of vital importance for improving operational quality and achieving energy conservation and emission reduction. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an adaptive temperature control system and method for flame soil disinfection, overcoming the deficiencies of existing flame disinfection equipment, such as being greatly affected by soil thermal characteristics and lacking real-time closed-loop control.

[0008] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0009] An adaptive temperature control system for flame soil disinfection includes an edge computing control unit and a front-end infrared detector, a flame monitoring unit, an infrared sensor array, a flow sensor, a speed sensor, and an electromagnetic proportional control valve that are signal-connected to the edge computing control unit.

[0010] The front-mounted infrared detector is installed in the undisturbed area at the front of the flame soil disinfection equipment to obtain the initial surface temperature and soil volumetric moisture content; the flame monitoring unit includes a high-temperature armored thermocouple array installed inside the core area of ​​each burner inside the flame hood of the flame soil disinfection equipment; the infrared sensor array includes a first linear infrared array arranged close to the rear edge of the flame hood and along the direction of travel, and a sensor array arranged at a distance behind the first linear infrared array. A second infrared linear array at the location; used to collect the instantaneous peak temperature immediately after leaving the flame's area and the temperature spanning the spatial distance. The decay temperature after combustion; the flow sensor is installed on the gas pipeline to monitor the real-time gas flow; the speed sensor is installed on the flame soil disinfection equipment vehicle to detect the real-time vehicle speed.

[0011] An adaptive temperature control method for flame-soil disinfection in the above-mentioned flame-soil disinfection adaptive temperature control system includes the following steps:

[0012] Step 1: Construct a multiphase flow physical feedforward model that incorporates latent heat of vaporization within the edge computing control unit to obtain the basic gas flow rate that maintains the target disinfection depth and reaches the target temperature;

[0013] Step 2: Extract the spatiotemporal thermal decay tensor of the multi-nozzle system to cope with nonlinear vehicle speeds;

[0014] Step 3: Construct a stacked LSTM network, build an asymmetric loss function based on the agronomic lethal temperature of the target organism, train the LSTM network, and use the trained LSTM network to infer and generate dynamic correction coefficients.

[0015] Step 4: The edge computing control unit performs weighted fusion of the basic gas flow rate and dynamic correction coefficient to generate the final control command, and controls the gas flow rate through the electromagnetic proportional regulating valve.

[0016] Furthermore, in step 1, the multiphase flow physics feedforward model incorporating latent heat of vaporization is as follows:

[0017]

[0018] in, Soil density; The overall specific heat capacity of the soil; Real-time vehicle speed; For multi-nozzle operation width; The latent heat of vaporization of water; This refers to the soil volumetric water content. For the combustion efficiency of the equipment; This refers to the calorific value of the gas. The non-target heat dissipation coefficient of porous media; The initial surface temperature, To achieve the target disinfection depth, The target temperature; Based on the basic gas flow rate.

[0019] Furthermore, step 2 is specifically as follows:

[0020] First, a digital odometry alignment algorithm based on integral dead-zone tracking is introduced, and the time when the first infrared linear array scans to the physical coordinate point A is set as... Edge computing control unit Start of instantaneous vehicle speed Perform high-frequency calculus accumulation until the accumulated displacement equals the physical distance between the first and second infrared linear arrays. : ;

[0021] When the equation holds true, determine that the second infrared linear array has reached coordinate point A, and record this moment. Measured decay temperature Based on this, the spatiotemporal thermal decay tensor is calculated. :

[0022]

[0023] in, The instantaneous peak temperature immediately after leaving the flame's area of ​​influence. To bridge spatial distance The decay temperature afterwards.

[0024] Furthermore, in step 3, the asymmetric loss function for the offline training phase before LSTM network deployment is as follows:

[0025]

[0026] in, This represents the total number of samples participating in the loss calculation for the current batch during the offline training phase. For sample index; For LSTM networks targeting the first Predicted temperature of deep soil output for each sample; The first temperature sensor pre-embedded at the target disinfection depth was used to obtain the first temperature data. The actual soil temperature after flame sterilization for each sample; The agronomic lethal temperature for the target organism; It is an indicator function;

[0027] and These are all hyperparameters pre-set before training the LSTM network, representing the underburn risk penalty weight and the overburn risk penalty weight, respectively. Both are determined simultaneously before training begins and remain fixed throughout the training process. The value range is [10, 50]. The value range of is [1, 5], and satisfies If the current training sample < This indicates a state of underburn risk, activated by the indicator function. The corresponding high penalty item, at this time The preset values ​​are retained, but the corresponding terms are not included in the loss calculation for this sample; if the current training sample's ≥ This indicates that the system is in an overheated and redundant state, activated by the indicator function. The corresponding low penalty item, at this time The preset values ​​are retained, but the corresponding items are not included in the loss calculation for this sample.

[0028] Furthermore, the LSTM network inverts the actual deep soil temperature, which cannot be directly measured in real time, based on the multi-source time-series feature sequence formed by continuous time steps during the mobile operation of the flame soil disinfection equipment; that is, the predicted deep soil temperature. The multi-source time-series feature sequence includes: the instantaneous peak temperature immediately after leaving the flame's action zone, and the spatial distance traversed. The data includes the following parameters: decay temperature, spatiotemporal thermal decay tensor, initial surface temperature, soil volumetric water content, real-time vehicle speed, flame core temperature obtained by the flame monitoring unit, and real-time or basic gas flow rate. The LSTM network performs state memory and nonlinear mapping on the multi-source temporal feature sequences to finally output the predicted temperature of deep soil.

[0029] Furthermore, in step 3, during the on-site operation phase of the flame-based soil disinfection equipment, the edge computing control unit no longer calculates the asymmetric loss function, but instead uses the spatiotemporal feature tensor acquired in real time. The input is fed into a pre-trained and fixed LSTM network. After state propagation through the hidden layers, the dynamically corrected coefficients are generated directly by forward inference from the fully connected output layer using the Sigmoid activation function. :

[0030]

[0031] In the formula, This is the output state tensor of the last hidden layer of the LSTM network. and After the offline training phase The weight and bias matrix fixed after optimization; It is the multidimensional spatiotemporal feature tensor constructed by the edge computing control unit at each time step. , Indicates real-time vehicle speed. Indicates soil volumetric moisture content. This represents the initial surface temperature. Represents the spacetime thermal decay tensor. This indicates the basic gas flow rate.

[0032] Furthermore, step 4 specifically includes:

[0033] Step 4.1: Adaptive weighting and flow command generation;

[0034] Based on dynamic correction coefficient and basic gas flow Calculate the desired gas flow command :

[0035]

[0036] in, The preset nonlinear adaptive gain coefficient;

[0037] Step 4.2: Gas flow control based on fuzzy PID;

[0038] To eliminate steady-state errors caused by pipeline pressure fluctuations, the desired gas flow command is... The real-time gas flow rate measured by the flow sensor will be used as the system target value. As a feedback quantity; calculate the deviation between the two. = - , will deviation The input is fed into the underlying fuzzy PID controller; the PID controller adjusts the proportional, integral, and derivative parameters online in real time, and outputs a high-frequency PWM duty cycle signal to directly drive the valve core displacement of the electromagnetic proportional regulating valve, thereby achieving servo-following control of the gas flow.

[0039] Furthermore, to address extreme conditions such as mud and water splashing or strong crosswinds causing infrared sensor array failure during farmland operations, a physical state safety check is introduced. The edge computing control unit verifies the spatiotemporal thermal attenuation tensor in real time. Detected When this occurs, degradation protection is triggered, forcing the setting to... =0.5, the system smoothly degrades to a pure physical feedforward control mode to avoid gas supply runaway caused by erroneous commands.

[0040] The present invention has the following beneficial effects:

[0041] A breakthrough solution has been found for the nonlinear heat transfer problem in porous agricultural media: for the first time, the latent heat of water vaporization has been introduced into the feedforward control of flame sterilization, which effectively corrects the prediction distortion caused by the large amount of heat consumption in moist soil and greatly improves the adaptability to soils with different moisture contents.

[0042] Strong resistance to interference from harsh field conditions: By using integral displacement to reconstruct the spatiotemporal thermal attenuation tensor, the data noise caused by track / tire bumps and sudden speed changes during field operations is completely decoupled, enabling accurate feature extraction under multi-nozzle, wide-area operations.

[0043] A "preventing under-burning" mechanism that is highly compatible with agronomic needs: The original asymmetric constraint deep learning algorithm based on the process lethal temperature enables the control system to have the agronomic thinking of "ensuring that the sterilization and weeding effect is better than saving energy consumption per cycle" from the bottom up, fundamentally eliminating ineffective operations. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the layout of the flame-soil disinfection adaptive temperature control system described in this invention;

[0045] Figure 2 Flowchart of an adaptive temperature control method for flame-induced soil disinfection;

[0046] Figure 3 This is a control flow diagram based on LSTM.

[0047] In the diagram: 1-Front-mounted infrared detector; 2-Flame hood; 3-Flame monitoring unit; 4-Flow meter; 5-Edge computing control unit; 6-Speed ​​sensor; 7-Electromagnetic proportional control valve; 8-Second infrared linear array; 9-First infrared linear array. Detailed Implementation

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0049] like Figure 1 As shown, the flame soil disinfection adaptive temperature control system of the present invention includes a multi-source sensing module, an edge computing control unit 5, and an actuator.

[0050] like Figure 1 As shown, the multi-source sensing module is used to perceive the working environment and thermodynamic state in all directions, including a front infrared detector 1, a flame monitoring unit 3, an infrared sensor array, a flow sensor, an edge computing control unit 5, and a speed sensor 6.

[0051] like Figure 1 As shown, the front infrared detector 1 is installed at the very front of the flame soil disinfection equipment (in the undisturbed area by fire) to obtain the initial surface temperature. and soil volumetric water content .

[0052] like Figure 1 As shown, for multi-burner coordinated operation, the flame monitoring unit 3 includes a high-temperature armored thermocouple array installed inside the core area of ​​each burner inside the flame hood 2. This array collects the flame temperature at multiple points in real time at high frequency to accurately capture local flameout or heat deflection caused by crosswinds or surface undulations in the field, ensuring the authenticity of the effective heat source input.

[0053] like Figure 1 As shown, the infrared sensor array is installed behind the flamethrower device, including a first infrared linear array 9 (arranged close to the rear edge of the flamethrower 2) and a second infrared linear array 8 (arranged behind the first infrared linear array 9 at a distance of...) arranged along the direction of travel. (Location), used to collect the instantaneous peak temperature immediately after the flame jet ends (just after leaving the flame's area of ​​influence). And a certain time delay after the assignment (across spatial distance) The decay temperature after ) .

[0054] like Figure 1 As shown, the flow sensor includes a flow meter 4 installed on the gas pipeline for monitoring real-time gas flow. .

[0055] like Figure 1As shown, speed sensor 6 is installed on the vehicle body of the flame soil disinfection equipment to detect real-time vehicle speed.

[0056] like Figure 1 As shown, the edge computing control unit 5 is used to realize intelligent control of adaptive temperature control, including a built-in two-layer composite control model based on "phase change compensation physical feedforward + agronomic constraint LSTM inversion".

[0057] like Figure 1 As shown, the actuator includes a high-response electromagnetic proportional control valve 7, which adjusts the gas flow rate of the main gas circuit and each nozzle branch circuit according to the fusion command of the edge computing control unit 5.

[0058] like Figure 2 , 3 As shown, a flame-soil disinfection adaptive temperature control method based on the above-mentioned flame-soil disinfection adaptive temperature control system specifically includes the following process:

[0059] Step 1: Construct a multiphase flow physics feedforward model incorporating latent heat of vaporization within the edge computing control unit 5;

[0060] To address the control distortion caused by the rapid boiling and absorption of large amounts of heat by water in moist soil, this paper breaks through the conventional single specific heat capacity algorithm. Based on the thermodynamic law of conservation of energy, a heat balance equation is constructed that includes compensation for sensible heat (soil warming) and latent heat (pore water phase change). A latent heat compensation term for water vaporization is introduced, based on the initial surface temperature. Calculate the target disinfection depth Reach the target temperature Basic gas flow :

[0061]

[0062] in, Soil density; The overall specific heat capacity of the soil; Real-time vehicle speed; For multi-nozzle operation width; The latent heat of vaporization of water; This refers to the soil volumetric water content. The combustion efficiency of the equipment was obtained through preliminary experiments; This represents the calorific value of the gas. The non-target heat dissipation coefficient of porous media (typically ranging from 1.05 to 1.15) has the physical meaning of compensating for the heat loss due to soil three-dimensional conduction to the target disinfection depth, in addition to the effective sensible heat rise and latent heat phase change of moisture. The diffusion of heat from the deep subsoil and the loss of heat from the surface to the air.

[0063] Traditional closed-loop control (including conventional PID or AI prediction) heavily relies on correcting errors after they occur in the system. When dealing with media like soil, which have enormous thermal inertia and moisture phase change damping, this inevitably leads to severe hysteresis, resulting in large-scale under-firing or energy waste. This feedforward model eliminates the large time-delay response deviation caused by moisture phase change in the system.

[0064] The above The formula essentially constructs a feedforward compensator for a moving unsteady multiphase flow (i.e., a multiphase flow physical feedforward model incorporating latent heat of vaporization), which not only encompasses the geometric space variables of agricultural machinery operations. Furthermore, it specifically introduces energy singularity compensation for water phase change in an open environment; this controller structure, which deeply couples the kinematic parameters of agricultural machinery with the thermodynamic parameters of porous media, has not been revealed in other industrial fields such as heat treatment, and has outstanding substantive features and significant progress.

[0065] Step 2: Extract the spatiotemporal thermal decay tensor of the multi-nozzle system to cope with nonlinear vehicle speeds;

[0066] To overcome the feature misalignment problem caused by track / tire slippage when agricultural machinery operates on muddy or uneven terrain, traditional temperature measurement relies on a fixed time step. When the vehicle speed fluctuates non-linearly, the front and rear sensors may not measure the soil at the same physical coordinate point. Therefore, this invention introduces a digital odometer alignment algorithm based on integral dead-zone tracking: The time when the first infrared linear array 9 scans to physical coordinate point A is set as... Edge computing control unit 5 Start of instantaneous vehicle speed Perform high-frequency calculus accumulation until the accumulated displacement equals the physical distance between the sensors (first infrared linear array 9 and second infrared linear array 8). : .

[0067] When the equation holds true, it is determined that the second infrared linear array 8 has precisely reached coordinate point A, and this moment is recorded. Measured decay temperature Based on this precisely aligned coordinate, the spatiotemporal thermal decay tensor is calculated. :

[0068]

[0069] in, The instantaneous peak temperature measured by the infrared sensor array immediately after the object leaves the flame's influence zone. To bridge spatial distance The decay temperature was measured by the infrared sensor array. This spatiotemporal thermal decay tensor addresses the complex working conditions of disinfection equipment slipping on tracks and fluctuating speeds in the field. By integrating and reconstructing the time step, the accuracy of the extracted temperature decay characteristics was ensured, and the deep heat conduction gradient in the overlapping area of ​​multiple nozzles under complex working conditions was successfully characterized.

[0070] Step 3: Asymmetric inversion and training of a Long Short-Term Memory (LSTM) network based on agronomic lethality boundary; the specific process is as follows:

[0071] Step 3.1: Offline training phase;

[0072] A stacked LSTM network (i.e., an LSTM model) is constructed to invert the actual deep soil temperature, which cannot be directly measured in real time, based on multi-source temporal features during mobile operations. The input to the LSTM network is a feature sequence composed of continuous time steps, which includes at least: the instantaneous peak temperature collected by the first infrared linear array 9, the decaying temperature collected by the second infrared linear array 8, the spatiotemporal thermal decay tensor constructed by a digital odometry alignment algorithm based on integral dead-zone tracking, the initial surface temperature obtained by the front infrared detector 1, the soil volumetric water content obtained by the front infrared detector 1, the real-time vehicle speed or integral displacement information obtained by the speed sensor 6, the flame core temperature obtained by the flame monitoring unit 3, and the real-time gas flow rate obtained by the flow sensor or the basic gas flow rate calculated by the physical feedforward model. The LSTM network performs state memory and nonlinear mapping on the above continuous time step feature sequence to finally output the predicted deep soil temperature.

[0073] In the offline training phase prior to the deployment of the aforementioned LSTM network, the actual deep soil temperature is obtained using a deep temperature sensor pre-embedded at the target disinfection depth as a supervision label. An asymmetric physical constraint penalty function (i.e., an asymmetric loss function) is constructed based on the agronomic lethal temperature of the target organisms (pests, diseases, and weeds). This asymmetric loss function measures the error between the predicted deep soil temperature output by the LSTM network and the actual temperature, and applies different penalty weights depending on whether the predicted temperature is below the lethal temperature threshold.

[0074]

[0075] in, This represents the total number of samples participating in the loss calculation for the current batch during the offline training phase. For sample index; For LSTM networks targeting the first Continuous values ​​of predicted deep soil temperature output for each sample; The first temperature sensor pre-embedded at the target disinfection depth was used to obtain the first temperature data. The actual soil temperature after flame sterilization for each sample; The agronomic lethal temperature for the target organism; It is an indicator function;

[0076] and These are all hyperparameters pre-set before training the LSTM network, representing the underburn risk penalty weight and the overburn risk penalty weight, respectively. Both are determined simultaneously before training begins and remain fixed throughout the training process. The value range is [10, 50]. The value range of is [1, 5], and satisfies Preferred If the current training sample < This indicates a state of underburn risk, activated by the indicator function. The corresponding high penalty item, at this time Although the preset values ​​are retained, the corresponding terms are not included in the loss calculation for that sample; if the current training sample's ≥ This indicates that the system is in an overheated and redundant state, activated by the indicator function. The corresponding low penalty item, at this time Although the preset values ​​are still maintained, the corresponding terms are not included in the loss calculation of the sample; this setting makes the LSTM network bear a higher cost for the risk of underburning during the training process, and establishes an asymmetric safety fault tolerance mechanism of "preventing underburning is better than preventing overburning" from the bottom of the algorithm.

[0077] During training, the Adam optimizer calculates the gradient based on the aforementioned asymmetric loss function and iteratively updates the weight matrix inside the LSTM network. With bias term The process continues until the loss function converges. Through this process, the LSTM network internalizes the agronomic physical constraint of "preventing under-burning is better than preventing over-burning" at the parameter level. After the LSTM model (network) is trained and deployed to the edge computing control unit 5, the asymmetric loss function no longer participates in online calculation. The system only calls the pre-defined LSTM network weights to perform forward inference on the real-time input feature sequence, outputting the predicted temperature value of deep soil or dynamic correction coefficient for subsequent gas flow control.

[0078] Step 3.2: Online Inference Phase;

[0079] During the field operation phase, the system no longer performs calculations. The edge computing control unit 5 will collect spatiotemporal feature tensors in real time. The input is fed into a pre-trained and fixed LSTM network. After state propagation through the hidden layers, the dynamically corrected coefficients are generated directly by forward inference from the fully connected output layer using the Sigmoid activation function. :

[0080]

[0081] In the formula, This is the output state tensor of the last hidden layer of the LSTM network. and After the offline training phase The optimized and fixed weight and bias matrix. Output. As a dimensionless coefficient, it directly characterizes the degree of nonlinear deviation of the current soil heat conduction damping relative to the standard operating conditions. It is the multidimensional spatiotemporal feature tensor constructed by the edge computing control unit 5 at each time step. .

[0082] Step 4: The edge computing control unit 5 uses dynamic fusion control based on safety boundary constraints and the underlying closed loop, and calculates the basic gas flow rate using a multiphase flow physical feedforward model that incorporates the latent heat of vaporization. Dynamic correction coefficients of the output of the LSTM network inversion Weighted fusion is performed to generate the final control command for gas flow control, as follows:

[0083] Step 4.1: Adaptive weighting and flow command generation;

[0084] Dimensionless dynamic correction coefficients based on LSTM network output The basic gas flow rate calculated by combining the physical feedforward model Calculate the desired gas flow command :

[0085]

[0086] in, The preset nonlinear adaptive gain coefficient is derived from the maximum safe flow physical limit of the vehicle-mounted electromagnetic proportional control valve and the flameout limit of the stable burner. It is a dynamic, real-time output variable, belonging to the real-time dynamic control quantity calculated in posterior inference, and originating from the continuous mapping value of the Sigmoid activation function in the output layer of the LSTM network; when When the value is greater than 0.5, it indicates that the actual deep heat conduction is hindered (such as encountering locally moist clay), and the system determines that there is a risk of "underburning" and automatically increases the firepower; when When the value is <0.5, the thermal permeability is good, and the system is determined to be in the "overheating" redundancy zone.

[0087] Meanwhile, to address extreme conditions such as mud and water splashing or strong crosswinds causing infrared sensor array failure during farmland operations, a physical state safety check is introduced: real-time verification of the spatiotemporal thermal decay tensor. If detected (Violating the thermodynamic law of natural cooling decay), the edge computing control unit 5 automatically determines that the confidence level of the AI ​​sensing link is too low. At this point, degradation protection is triggered, forcibly setting... =0.5 (i.e., stripping the LSTM compensation weights), the system smoothly degrades to a pure physical feedforward control mode to ensure that the equipment will not cause the gas supply to run out of control due to erroneous commands.

[0088] Step 4.2: Gas flow control based on fuzzy PID;

[0089] To eliminate steady-state errors caused by pipeline pressure fluctuations, the desired gas flow command is... The real-time gas flow rate measured by flow meter 4 on the main gas pipeline will be used as the system target value. As a feedback quantity; calculate the deviation between the two. = - , will deviation The input is fed into the underlying fuzzy PID controller; the PID controller adjusts the proportional, integral, and derivative parameters online in real time, and outputs a high-frequency PWM duty cycle signal to directly drive the valve core displacement of the electromagnetic proportional regulating valve 7, thereby achieving millisecond-level precise servo following control of the gas flow.

[0090] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for adaptive temperature control of flame-based soil disinfection using a flame-based soil disinfection adaptive temperature control system, characterized in that, The flame soil disinfection adaptive temperature control system includes an edge computing control unit (5) and a front infrared detector (1), a flame monitoring unit (3), an infrared sensor array, a flow sensor, a speed sensor (6), and an electromagnetic proportional control valve (7) connected to the edge computing control unit (5). The front infrared detector (1) is installed in the undisturbed area at the front of the flame soil disinfection equipment to obtain the initial surface temperature and soil volumetric moisture content; the flame monitoring unit (3) includes a high-temperature armored thermocouple array installed inside the core area of ​​each burner inside the flame hood (2) of the flame soil disinfection equipment; the infrared sensor array includes a first infrared linear array (9) arranged close to the rear edge of the flame hood (2) and along the direction of travel, and a distance behind the first infrared linear array (9). The second infrared linear array (8) at the location is used to collect the instantaneous peak temperature and the spatial distance just after leaving the flame's action zone. The decay temperature after the gas flow is measured; the flow sensor is installed on the gas pipeline to monitor the real-time gas flow; the speed sensor (6) is installed on the flame soil disinfection equipment vehicle to detect the real-time vehicle speed. The adaptive temperature control method for flame-based soil sterilization includes the following processes: Step 1: Construct a multiphase flow physical feedforward model that incorporates latent heat of vaporization within the edge computing control unit (5) to obtain the basic gas flow rate that maintains the target disinfection depth and reaches the target temperature; Step 2: Extract the spatiotemporal thermal decay tensor of the multi-nozzle system to cope with nonlinear vehicle speeds; Step 3: Construct a stacked LSTM network, build an asymmetric loss function based on the agronomic lethal temperature of the target organism, train the LSTM network, and use the trained LSTM network to infer and generate dynamic correction coefficients. Step 4: The edge computing control unit (5) performs weighted fusion of the basic gas flow rate and dynamic correction coefficient to generate the final control command, and controls the gas flow rate through the electromagnetic proportional regulating valve (7); In step 1, the multiphase flow physics feedforward model incorporating latent heat of vaporization is as follows: ； in, Soil density; The overall specific heat capacity of the soil; Real-time vehicle speed; For multi-nozzle operation width; The latent heat of vaporization of water; This refers to the soil volumetric water content. For the combustion efficiency of the equipment; This refers to the calorific value of the gas. The non-target heat dissipation coefficient of porous media; The initial surface temperature, To achieve the target disinfection depth, The target temperature; Based on the basic gas flow rate.

2. The adaptive temperature control method for flame soil disinfection according to claim 1, characterized in that, Step 2 is described in detail below: First, a digital odometry alignment algorithm based on integral dead-zone tracking is introduced, and the time when the first infrared linear array (9) scans to the physical coordinate point A is set as... The edge computing control unit (5) is self- Start of instantaneous vehicle speed Perform high-frequency calculus accumulation until the accumulated displacement equals the physical distance between the first infrared linear array (9) and the second infrared linear array (8). : ; When the equation holds true, it is determined that the second infrared linear array (8) has reached coordinate point A, and this moment is recorded. Measured decay temperature Based on this, the spatiotemporal thermal decay tensor is calculated. : ； in, The instantaneous peak temperature immediately after leaving the flame's area of ​​influence. To bridge spatial distance The decay temperature afterwards.

3. The adaptive temperature control method for flame soil disinfection according to claim 1, characterized in that, In step 3, the asymmetric loss function for the offline training phase before LSTM network deployment is as follows: ； in, This represents the total number of samples participating in the loss calculation for the current batch during the offline training phase. For sample index; For LSTM networks targeting the first Predicted temperature of deep soil output for each sample; The first temperature sensor pre-embedded at the target disinfection depth was used to obtain the first temperature data. The actual soil temperature after flame sterilization for each sample; The agronomic lethal temperature for the target organism; It is an indicator function; and These are all hyperparameters pre-set before training the LSTM network, representing the underburn risk penalty weight and the overburn risk penalty weight, respectively. Both are determined simultaneously before training begins and remain fixed throughout the training process. The value range is [10, 50]. The value range of is [1, 5], and satisfies If the current training sample < This indicates a state of underburn risk, activated by the indicator function. The corresponding high penalty item, at this time The preset values ​​are retained, but the corresponding terms are not included in the loss calculation for this sample; if the current training sample's ≥ This indicates that the system is in an overheated and redundant state, activated by the indicator function. The corresponding low penalty item, at this time The preset values ​​are retained, but the corresponding items are not included in the loss calculation for this sample.

4. The adaptive temperature control method for flame soil disinfection according to claim 3, characterized in that, The LSTM network inverts the actual deep soil temperature, which cannot be directly measured in real time, based on the multi-source time-series feature sequence formed by continuous time steps during the mobile operation of the flame soil disinfection equipment; that is, the predicted deep soil temperature. The multi-source time-series feature sequence includes: the instantaneous peak temperature immediately after leaving the flame's action zone, and the spatial distance traversed. The decay temperature, spatiotemporal thermal decay tensor, initial surface temperature, soil volumetric water content, real-time vehicle speed, flame core temperature obtained by the flame monitoring unit (3), real-time gas flow rate or basic gas flow rate; the LSTM network outputs the deep soil predicted temperature by performing state memory and nonlinear mapping on the multi-source time series feature sequences.

5. The adaptive temperature control method for flame soil disinfection according to claim 1, characterized in that, In step 3, during the on-site operation of the flame soil disinfection equipment, the edge computing control unit (5) no longer calculates the asymmetric loss function, and instead uses the spatiotemporal feature tensor collected in real time. The input is fed into a pre-trained and fixed LSTM network. After state propagation through the hidden layers, the dynamically corrected coefficients are generated directly by forward inference from the fully connected output layer using the Sigmoid activation function. : ； In the formula, This is the output state tensor of the last hidden layer of the LSTM network. and After the offline training phase The weight and bias matrix fixed after optimization; It is the multidimensional spatiotemporal feature tensor constructed by the edge computing control unit (5) at each time step. , Indicates real-time vehicle speed. Indicates soil volumetric moisture content. This represents the initial surface temperature. Represents the spacetime thermal decay tensor. This indicates the basic gas flow rate.

6. The adaptive temperature control method for flame soil disinfection according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1: Based on dynamic correction coefficient and basic gas flow Calculate the desired gas flow command : ； in, The preset nonlinear adaptive gain coefficient; Step 4.2: Send the desired gas flow command The real-time gas flow rate measured by the flow sensor serves as the system's target value. As a feedback quantity, the deviation between the two is calculated. , will deviation The input is fed into the underlying fuzzy PID controller; the PID controller adjusts the proportional, integral, and derivative parameters online in real time, and outputs a high-frequency PWM duty cycle signal to directly drive the valve core displacement of the electromagnetic proportional regulating valve (7) to achieve servo-following control of the gas flow.

7. The adaptive temperature control method for flame soil disinfection according to claim 6, characterized in that, To address extreme conditions such as mud splashing or strong crosswinds causing infrared sensor array failure during farmland operations, a physical state safety check is introduced. The edge computing control unit (5) verifies the spatiotemporal thermal attenuation tensor in real time. Detected When this occurs, degradation protection is triggered, forcing the setting to... =0.5, the system smoothly degrades to a pure physical feedforward control mode to avoid gas supply runaway caused by erroneous commands.

Images