Self-adaptive variable-force soil compacting method and device based on image recognition

Abstract

The invention discloses a self-adaptive variable-force soil compacting method and device based on image recognition. According to the method, a camera is utilized to collect a soil surface image of a suppression area, a deep learning target detection model is adopted to identify soil blocks with different particle sizes, the area proportion, the number and the maximum single block area are counted, a soil block comprehensive index CI is constructed, and the CI is mapped into a target suppression force P through a monotone increasing piecewise function; target displacement is calculated by combining the calibration relation between the press force P and the displacement L of the electric push rod, the electric push rod is controlled to adjust the acting force of the press wheel on the ground, and a press force instruction can be smoothed in cooperation with a time sequence prediction model. Aiming at the problems of insufficient pressing force adjusting precision, insensitivity to soil block state response and frequent action of an execution mechanism in the existing pressing technology, the continuous adjustment of the pressing force and the adaptive matching to different soil conditions are realized, the pressure fluctuation is reduced, the energy consumption is reduced, and the pressing quality is obviously improved.

Description

Technical Field

[0001] This invention belongs to the field of agricultural machinery and intelligent control technology, specifically relating to an adaptive variable force soil compaction method and device based on image recognition, which is applicable to scenarios where agricultural machinery such as seeders perform intelligent and adjustable soil compaction during operation. Background Technology

[0002] Soil compaction is a crucial step in agricultural machinery operations such as sowing and planting, ensuring stable sowing depth, high-quality seed covering, and uniform seedling emergence. Existing agricultural machinery widely employs several soil compaction methods, including mechanical fixed-force compaction relying on the weight of the compaction wheel or additional load; spring-suspended compaction using spring deformation for cushioning; and hydraulic / pneumatic variable-force compaction utilizing hydraulic cylinders or airbags for active adjustment. Mechanical fixed-force compaction is simple in structure and low in cost, but the compaction pressure is essentially constant, making it difficult to adapt to different soil textures, moisture contents, and surface roughness, and prone to localized under-compaction or over-compaction. While spring-type structures can cushion the impact of surface undulations to a certain extent, the pressure adjustment precision and range are limited, making it difficult to achieve fine-tuning for different operating conditions. Hydraulic / pneumatic adjustable compaction relies on actuators to actively change the compaction pressure, offering better adaptability, but its system structure is complex, costly, and energy-intensive, requiring high levels of maintenance and control system stability.

[0003] With the development of sensor and electronic control technologies, some agricultural machinery has begun to incorporate pressure sensors, downforce sensors, and displacement sensors to achieve real-time detection of compaction pressure or sowing depth, and to perform closed-loop adjustment of the actuators accordingly. For example, Chinese patent document CN119717915A discloses a method and control device for controlling the sowing depth and compaction of individual seeding units. This method uses a compaction pressure sensor to collect the contact pressure between the compaction wheel and the ground, and combines this with downforce information collected by a downforce sensor to adjust the extension and retraction of the compaction hydraulic cylinder and the downforce hydraulic cylinder in real time. This achieves variable force compaction and sowing depth control for individual seeding components, improving the adaptability and quality of compaction operations to a certain extent. The aforementioned Chinese patent document CN119717915A may also be referred to as "Document 1" below.

[0004] However, existing compaction control schemes based on pressure sensor feedback (including the technical solution disclosed in Reference 1) generally rely on local contact stress or structural force signals at the interface between the compaction wheel and the soil, reflecting comprehensive mechanical results. This makes it difficult to directly characterize key soil condition information such as clod size, clod distribution, and surface roughness. Furthermore, their adjustment strategies are mostly based on a single pressure threshold or simple interval judgment, lacking comprehensive evaluation and quantitative characterization of multiple indicators such as the number of clods of different particle sizes, area proportion, and maximum single-clod size. In areas with significant spatial variations in soil conditions or complex clod size distribution, problems such as delayed compaction pressure adjustment, insufficiently precise response, and large pressure fluctuations easily occur. Insufficient compaction, over-compaction, and high energy consumption remain. Especially in operational scenarios where the quality of crushed soil is significantly affected by clod size distribution, existing technologies lack a technical solution that combines soil surface image recognition results, comprehensively assesses clod conditions, and precisely drives the actuator to achieve adaptive variable force compaction, requiring further improvement and refinement. Summary of the Invention

[0005] The purpose of this invention is to address the problems of insufficient pressure adjustment precision, inadequate response to soil block size and distribution, over-reliance on local pressure sensing information, and high energy consumption due to frequent actuator movements in existing soil compaction technologies. This invention provides an adaptive variable force soil compaction method and apparatus based on image recognition. By integrating a camera, embedded controller, and electric actuator into the compaction device, the particle size and distribution of soil blocks on the soil surface in front of the compaction area are visually perceived and comprehensively quantified. A comprehensive soil block index CI characterizing the surface roughness is constructed, and this index is mapped to the target compaction pressure P. Combined with the monotonic calibration relationship between the compaction pressure and the electric actuator displacement L, the compaction pressure of the compaction wheel is continuously and precisely adjustable, thereby achieving adaptive matching of compaction operations to different soil conditions, improving compaction quality and operational intelligence.

[0006] To achieve the above objectives, this invention provides an adaptive variable force soil compaction method based on image recognition. This method is applied to a soil compaction device including a camera, an electric push rod, and a compaction wheel mechanism. The method includes the following steps: S1. A camera and an electric push rod are installed on the frame of the soil compaction device. The electric push rod is connected to the compaction wheel mechanism. During field operations, the camera acquires images of the soil surface containing soil clods and uploads them to an embedded controller; S2. The acquired soil surface images are input into a soil clod image recognition model deployed on the embedded controller to detect and classify the soil clods in the images, obtaining the number, area, and distribution information of soil clods of different particle sizes; S3. Based on the distribution information of soil clods of different particle sizes, a soil clod comprehensive index CI, representing the roughness of the soil surface, is calculated according to preset rules. CI is a dimensionless comprehensive index representing the current state of soil clods on the soil surface. Based on the preset mapping relationship between CI and compaction pressure, the corresponding target compaction pressure P is obtained, where P is the target compaction pressure exerted by the compaction wheel on the ground; S4. Based on the pre-calibrated compaction pressure P and the electric push rod displacement L... The monotonic correspondence between the two is used to calculate the target push rod displacement L using the target pressing pressure P, where L is the extension and retraction displacement of the electric push rod. The embedded controller sends a control command to the electric push rod controller to drive the electric push rod to extend and retract to the target push rod displacement L, thereby adjusting the pressing pressure of the pressing wheel on the ground so that the actual pressing pressure is close to the target pressing pressure P.

[0007] Preferably, the soil block image recognition model is a deep learning object detection model, and the model construction and training process is included before step S2: using the diameter of the soil block as the grading basis, let the diameter of a single soil block be d, and pre-define K particle size intervals. where K≥3, and satisfies that for any k, When the diameter d of a certain soil block satisfies When the soil block is classified into the k-th class, a multi-class classification of soil blocks based on particle size intervals is obtained. Soil surface images containing various types of soil blocks are collected, and image annotation tools are used to annotate the category index k and corresponding bounding box position of each soil block in the image, forming a sample dataset with category labels and bounding box information. The sample dataset is divided into training, validation, and test sets to train and validate the deep learning object detection model. The target model is determined based on performance indicators such as detection accuracy, false negative rate, and inference speed, and the target model is converted into a model format suitable for running on an embedded controller for deployment. Preferably, the object detection model can adopt a lightweight network structure based on YOLO to balance recognition accuracy and embedded inference speed.

[0008] Preferably, in order to comprehensively reflect the impact of soil blocks of different particle sizes on compaction requirements, the calculation of the soil block comprehensive index CI in this invention includes at least the following steps: calculating the area percentage for each type of soil block (the kth type, k = 1, 2, ..., K). Quantitative indicators and the percentage of the largest single area Among them, the area proportion This represents the proportion of the sum of the areas of all detection frames of this type to the total area of ​​the target region; it is a quantitative indicator. This represents the normalized value of the number of detection boxes of this type relative to the preset reference number, and the percentage of the largest single block area. This indicates the proportion of the total area of ​​the target region to the area of ​​the largest single detection bounding box of this type. Based on preset weighting coefficients. The basic score for the k-th type of soil block is obtained by weighted summation of the above indicators. ,in The base score for the k-th type of soil block is as follows: The This is used to adjust the relative importance of area proportion, quantity indicators, and the largest single block area proportion in the base score. Subsequently, different importance weights are assigned to various types of soil blocks based on the differences in the impact of soil blocks of different particle sizes on the compaction operation. Basic scores for various types of soil blocks Weighted summation is performed to obtain the unnormalized comprehensive measurement of soil blocks. and will Normalize to preset interval The soil block comprehensive index CI is obtained, which is a dimensionless quantity and serves as the sole driving force for the target town pressure P.

[0009] Preferably, the mapping relationship between the soil block comprehensive index CI and the target town pressure P is a monotonically increasing piecewise linear function. Specifically, the normalized CI satisfies... There is a threshold ,satisfy and positive numbers and constant This ensures that the mapping of the target town pressure with respect to CI satisfies:

[0010]

[0011] in, Corresponding to the light repression level, Corresponding to a medium level of repression, Corresponding to the heavy compaction level, and within each level range, the target compaction pressure P is limited to the corresponding compaction pressure level range, so as to achieve differentiated compaction of different soil blocks and ensure that the overall mapping relationship is monotonically increasing.

[0012] Preferably, after obtaining the soil block composite index CI and calculating the target pressing pressure P in step S3, the present invention further performs dynamic smoothing and correction on the soil block composite index CI and the target pressing pressure P using a pre-trained deep learning time series prediction model. The input to the deep learning time series prediction model includes one or more of the following: soil block composite index CI, target pressing pressure P, pressing wheel speed, electric push rod displacement L, and agricultural machinery travel speed at multiple consecutive sampling times. The output is the smoothed and corrected target pressing pressure and the corresponding target push rod displacement. The deep learning time series prediction model can be one or more of the following: recurrent neural network (RNN), long short-term memory network (LSTM), and gated recurrent unit network (GRU), and is obtained through offline training based on historical operation data to reduce drastic fluctuations in pressing pressure commands over time and mitigate the adverse effects of frequent electric push rod movements on system stability and energy consumption.

[0013] Preferably, the monotonic correspondence between the tamping pressure P and the electric actuator displacement L is obtained through experimental calibration. During calibration, tamping pressure P corresponding to different actuator displacements L is collected, and a monotonic function P = g(L) is fitted, where g(·) is a monotonic function. In actual operation, based on the calculated target tamping pressure P, the corresponding target actuator displacement L is obtained through numerical back-calculation or table lookup, i.e. The target push rod displacement L is then sent to the electric push rod controller. The embedded controller sends the target push rod displacement L and one or more of the following signals—running speed, acceleration, and start / stop control signals—to the electric push rod controller via bus communication. The preferred bus communication method is Modbus communication, thereby achieving precise position control and pressure adjustment of the electric push rod.

[0014] In another aspect, the present invention provides a variable force soil compaction device for implementing the above-described method. The device includes a frame, a compaction wheel, a compaction wheel axle connected to the compaction wheel, a fixed bearing device, an electric push rod, a fixed bracket or support rod for connecting the electric push rod to the frame, and a camera mounted on the frame or fixed bracket. One end of the electric push rod is hinged to the frame or fixed bracket, and the other end is hinged to the compaction wheel axle or a rocker arm mechanism connected to the compaction wheel axle. The position of the compaction wheel relative to the frame is changed by extending or shortening the electric push rod. When the electric push rod extends, the compaction pressure of the compaction wheel on the ground increases; when the electric push rod shortens, the compaction pressure of the compaction wheel on the ground decreases. The camera is electrically connected to an embedded controller for acquiring images of the soil surface in front of the compaction area and sending the images to the embedded controller. The embedded controller calculates the soil block composite index (CI) according to the above-described image recognition-based adaptive variable force soil compaction method. The target pressing pressure P is smoothed over time and the displacement is reversed to output the target push rod displacement L, which is then used to control the electric push rod to perform the action, thereby achieving adaptive variable force adjustment of the pressing wheel pressing pressure.

[0015] Compared with existing technologies, this invention directly obtains the particle size and distribution characteristics of soil clods on the soil surface through a camera and soil clod image recognition model, constructs a comprehensive soil clod index (CI) driven by area proportion, quantity index, and the largest single clod area proportion, and combines the importance weights of different particle size categories. This allows the adjustment of compaction pressure to be based on comprehensive quantitative indicators that better reflect the quality of land preparation and the state of soil fragmentation. Compared with the scheme that simply relies on local pressure sensors, it can more accurately reflect the degree of soil surface roughness and the risk of large soil clods, and improve the responsiveness and adaptability of compaction operations to different soil conditions.

[0016] Furthermore, this invention employs a monotonically increasing piecewise linear function to map the soil block comprehensive index CI to the target compaction pressure P. A deep learning time series prediction model is introduced to dynamically smooth and correct CI and P. Simultaneously, experimental calibration establishes a monotonically corresponding relationship between compaction pressure P and electric actuator displacement L, and the electric actuator displacement is precisely controlled via a bus. This ensures that the compaction pressure command exhibits good time smoothness and controllability while maintaining responsiveness, significantly reducing frequent electric actuator movements and system energy consumption. It also improves the control accuracy and stability of the compaction wheel's force on the ground. This provides a simple and engineerable overall technical solution for adaptive variable force soil compaction in complex field environments, thereby significantly improving the quality of compaction operations and the level of intelligence and precision in agricultural machinery operations. Attached Figure Description

[0017] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart of an image recognition-based adaptive variable force soil compaction method;

[0019] Figure 2 This is a schematic diagram of a variable force soil compaction device;

[0020] Figure 3 This is a flowchart of the target town pressure smoothing and correction process based on a deep learning time series prediction model. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 As shown, the overall process of the adaptive variable force soil compaction method based on image recognition provided by this invention includes steps S1 to S3. Step S1 is "construction of soil block dataset and model training", step S2 is "image recognition and calculation of target compaction pressure", and step S3 is "push rod displacement control and variable force compaction". This method takes the soil surface image in front of the compaction area as input and the compaction pressure exerted by the compaction wheel on the ground as the adjustment object. Through soil block image recognition, calculation of soil block comprehensive index CI, and control of electric push rod displacement L, the compaction pressure is adaptively adjusted to different soil and soil block conditions.

[0023] Step S1, Soil clod dataset construction and model training (offline stage). Based on the distribution of soil clods of different sizes in the field, soil surface images containing soil clods of multiple sizes are collected. K particle size intervals are pre-defined based on the soil clod diameter *d*. K≥3, and for any k satisfies When the diameter d of a certain soil block satisfies When the soil block is classified into the k-th class, the collected soil surface images are labeled using an image annotation tool to provide the class index k and corresponding bounding box position for each soil block target, forming a soil block image dataset with class labels and location information. This dataset is divided into a training set, a validation set, and a test set. The soil block image recognition model (preferably a deep learning object detection model) is trained based on the training set. The model structure and hyperparameters are fine-tuned using the validation set, and the detection accuracy and inference speed are evaluated on the test set, resulting in a soil block detection model capable of distinguishing soil blocks of different sizes and meeting the real-time requirements of embedded systems. Preferably, the trained soil block detection model is converted into a model format suitable for running on an embedded controller and deployed in the embedded controller, providing a foundation for subsequent online recognition.

[0024] Step S2: Image Recognition and Target Compressive Pressure Calculation. As the soil compaction device moves forward with the agricultural machinery, a camera mounted on the frame acquires images of the soil surface in front of the area to be compacted at preset intervals. The acquired images are then input into the soil block detection model deployed on the embedded controller in Step S1. The model detects and classifies the soil blocks in the images, obtaining the category index k and bounding box information for each soil block target. Based on this, the embedded controller counts the number of soil blocks of each category, the area of ​​the bounding box, and the spatial distribution information within the field of view. Within the current sampling period, it calculates at least the area percentage for each type of soil block (the kth category, k = 1, 2, ..., K). Quantitative indicators and the percentage of the largest single area One or more of the indicators.

[0025] Among them, the area proportion This represents the proportion of the sum of the areas of all detection frames of this type to the total area of ​​the target region; it is a quantitative indicator. This represents the normalized value of the number of detection boxes of this type relative to the preset reference number, and the percentage of the largest single block area. This indicates the proportion of the total area of ​​the target region to the area of ​​the largest single detection bounding box of this type. Based on preset weighting coefficients. The basic score for the k-th type of soil block is obtained by weighted summation of the above indicators. :

[0026] ……(1)

[0027] In formula (1), This is the base score for the k-th type of soil block. This is used to adjust the relative importance of area proportion, quantity indicators, and the largest single block area proportion in the base score. Furthermore, to reflect the differences in the impact of soil blocks of different sizes on compaction operations, importance weights are assigned to various types of soil blocks. The basic scores of various types of soil blocks are weighted and summed to obtain the unnormalized comprehensive soil block metric. :

[0028] ……(2)

[0029] The summation range is k = 1, 2, ..., K. Then, Normalize to preset interval Within the system, a dimensionless soil block composite index (CI) is obtained, which characterizes the current soil surface roughness and soil fragmentation, and serves as the sole driving force for the target compaction pressure (P). The embedded controller, based on a pre-established monotonically increasing piecewise linear mapping relationship, converts the soil block composite index CI into the target compaction pressure (P), thereby obtaining the target compaction pressure exerted by the compaction wheel on the ground under the current operating conditions.

[0030] Preferably, after obtaining the soil block composite index CI and calculating the target pressing pressure P in step S2, the present invention further performs dynamic smoothing and correction on CI and P using a pre-trained deep learning time series prediction model. The input to the deep learning time series prediction model includes one or more of the following: soil block composite index CI, target pressing pressure P, pressing wheel speed, electric push rod displacement L, and / or agricultural machinery travel speed at multiple consecutive sampling times. The output is the smoothed and corrected target pressing pressure and / or the corresponding target push rod displacement. The deep learning time series prediction model can be one or more of the following: recurrent neural network (RNN), long short-term memory network (LSTM), and gated recurrent unit network (GRU), and is obtained through offline training based on historical operation data to reduce drastic fluctuations in pressing pressure commands over time, suppress high-frequency ineffective adjustments, and reduce the adverse effects of frequent electric push rod movements on system stability and energy consumption.

[0031] Step S3, push rod displacement control and variable force pressing. To achieve a quantitative correspondence between the target pressing force P and the electric push rod displacement L, several sets of different electric push rod displacements L were experimentally measured during the system calibration phase. i Corresponding town pressure P i The data was used to fit a monotonic function:

[0032] ……(3)

[0033] In formula (3), g(·) is a monotonic function characterizing the relationship between the electric push rod displacement and the pressing wheel pressure. During actual field operations, the embedded controller calculates the target push rod displacement L using the inverse function relationship of formula (3) based on the target pressing pressure P (or the smoothed target pressing pressure) obtained from step S2 and time series smoothing. Alternatively, the target displacement value closest to P can be found in a pre-established lookup table, and the target push rod displacement, running speed, acceleration, and start / stop signals can be sent to the electric push rod controller via bus communication.

[0034] The electric actuator controller, based on the target actuator displacement L and related motion parameters issued by the embedded controller, drives the electric actuator to extend or shorten, changing the position of the pressing wheel relative to the frame. This adjusts the force exerted by the pressing wheel on the ground, gradually bringing the actual pressing pressure closer to the target pressing pressure, thus achieving variable force and adaptive pressing under different soil and clod conditions. Preferably, the embedded controller can directly use the smoothed target actuator displacement output by the time series prediction model as a control reference to further improve the smoothness of pressing pressure adjustment.

[0035] Preferably, to further reduce ineffective actuator movements and lower system energy consumption, the embedded controller can also be set with a minimum displacement change threshold and a maximum adjustment frequency: when the target push rod displacement change is less than the preset threshold within a continuous control cycle, and the actual pressure is within the allowable error range of the target pressure, the current push rod position remains unchanged; when the pressure is stably maintained within the target range for a preset time, the electric push rod controller enters a low-power holding state. Through these measures, while ensuring the accuracy of pressure control, the energy consumption and mechanical wear caused by frequent starts and stops of the electric push rod can be effectively reduced, further improving the system's reliability and economy.

[0036] In one specific embodiment, the number of soil clod size classifications is K = 3, corresponding to three categories: small soil clods, medium soil clods, and large soil clods. The classification interval of soil clod diameter d is set, for example, as follows: small soil clods medium soil block Large clods of earth [30,+ In this embodiment, the total area of ​​the target region is determined by the projected area of ​​the camera's field of view onto the ground surface, and the embedded controller calculates the area ratio of the three types of soil blocks in each sampling period. Quantitative indicators and the percentage of the largest single area The scores for various types of soil block foundations are calculated using formula (1). To simplify the calculation, weighting coefficients can be selected. , , The area percentage has the greatest impact on the base score, followed by the quantity index and the largest single area percentage.

[0037] Continuing with the above embodiments, to reflect the greater adverse impact of large soil clods on compaction operations, the importance weight is... Preferred setting: large soil blocks medium soil block small clods of earth The embedded controller calculates according to formula (2). Then, by using historical operation data to obtain statistics... Typical minimum value and maximum value and in a linear manner Normalization to The interval is used to obtain the comprehensive index of soil blocks. In practical engineering applications, the soil type and land preparation conditions of different regions can be considered when making adjustments. Adjustments were made to adapt to different operating conditions.

[0038] In the above embodiments, to facilitate graded management of rolling pressure, the CI interval [0,1] can be divided into three levels: light rolling, medium rolling, and heavy rolling. For example, a threshold value can be selected. , When CI ∈ [0, 0.3), it corresponds to light pressing; when CI ∈ [0.3, 0.7), it corresponds to medium pressing; and when CI ∈ [0.7, 1], it corresponds to heavy pressing. Assume the pressing pressure adjustment range of a certain type of pressing device is... Then, linear mappings can be set in three intervals respectively: in the lightly suppressed interval, CI is mapped from [0, 0.3) to... In the medium-suppression interval, CI is mapped from [0.3, 0.7) to P∈[5, 8] kN; in the heavy-suppression interval, CI is mapped from [0.7, 1] to P∈[8, 10] kN, thereby determining the slope of each segment in formula (2). and intercept .

[0039] In an exemplary calibration experiment, the pressing device is fixed on a test bench. The displacement L of the electric push rod is controlled to vary in stages within the effective stroke range, such as L taking several discrete points like 0, 5, 10, 15, and 20 mm. The pressing force P of the pressing wheel on the ground is measured at each position to obtain a set of calibration data points. An approximate linear relationship is obtained using least squares fitting. Where a>0 indicates that the ballast pressure increases with displacement, and b is a constant term. In practical applications, when the target ballast pressure P is calculated according to steps S2-S3, the embedded controller can... The target push rod displacement L is calculated in reverse and sent to the electric push rod controller for execution.

[0040] It is understood that the particle size range and weighting coefficients in the above embodiments... and threshold , Town pressure range and calibration function The specific form can be appropriately adjusted according to the soil type, crop type, and agronomic requirements of different regions. As long as image recognition based on soil particle size classification is still used, a comprehensive soil index CI is constructed, and adaptive adjustment of pressing pressure and electric actuator displacement is achieved through monotonic mapping relationship, it is considered to fall within the protection scope of this invention.

[0041] like Figure 2 As shown, in this embodiment, the variable force soil compaction device is installed on the frame 7 of the seeder or compactor, including a compaction wheel 1 arranged along the rear of the frame 7, a compaction wheel axle 2 for supporting the compaction wheel 1, a fixed bearing device 3 mounted on the frame 7, an electric push rod 4, a fixed bracket 5 and a support rod 6 for connecting the electric push rod 4 and the frame 7, and the frame 7 body. The compaction wheel 1 is located at the rear of the frame 7 or behind the sowing row, and is used to compact the seedbed area covered with soil. The compaction wheel axle 2 is mounted on the frame 7 through the fixed bearing device 3 and can rotate relative to the frame 7. One end of the electric push rod 4 is hinged to the fixed bracket 5 or the support rod 6, and the other end is hinged to the compaction wheel axle 2 or a rocker arm mechanism connected to the compaction wheel axle. By extending or shortening the electric push rod 4, the spatial position of the compaction wheel 1 relative to the frame 7 is changed, thereby adjusting the force exerted by the compaction wheel 1 on the ground.

[0042] The camera is preferably fixed on frame 7 (in Figure 2 (Not separately labeled), its shooting direction is basically perpendicular to the ground, so that the image captured by the camera covers the surface area of ​​the soil to be compacted. The camera is electrically connected to the embedded controller via wired or wireless means. The embedded controller preferably uses an embedded platform with image processing and deep learning model inference capabilities, such as a small industrial control computer running Linux or an embedded processor with a dedicated GPU acceleration module, and is connected to the electric linear actuator controller via a bus communication interface (such as RS485, CAN, Ethernet, etc.). The electric linear actuator controller is used to receive target displacement commands, as well as speed, acceleration, and start / stop signals from the embedded controller, and drive the electric linear actuator 4 to complete the corresponding extension and retraction actions.

[0043] After the device is installed, the soil block image recognition model corresponding to step S1 is first trained and deployed. Specifically, K particle size intervals are pre-defined based on the soil block diameter d. K≥3 and for any k satisfies When the diameter d of a certain soil block satisfies When the soil clod is classified into the k-th soil clod category, field soil surface images containing soil clods of various particle sizes are collected. Image annotation tools are used to assign category index k and bounding box annotations to the soil clod targets in the images, forming a sample dataset with category labels and bounding box information. The sample dataset is divided into training, validation, and test sets. A deep learning object detection model is trained on the training set, the model structure and hyperparameters are adjusted using the validation set, and the detection accuracy and inference speed are evaluated on the test set to determine the target model whose performance meets the requirements of field operations. To balance accuracy and real-time performance, a lightweight object detection network structure based on YOLO is preferred. Finally, the trained object detection model is converted into a model format suitable for running on an embedded controller and deployed to the embedded controller.

[0044] During operation, the agricultural machinery moves along the designated route, with the compaction device moving synchronously. An embedded controller triggers the camera to acquire soil surface images at a preset sampling period and inputs the acquired images into the aforementioned soil clod image recognition model. The model detects and classifies the soil clods in the images, outputting the position, area, and particle size category (k) of each detection frame. Based on the output results, the embedded controller calculates the area percentage of each type of soil clod. Quantitative indicators and the percentage of the largest single area ,in, This represents the proportion of the sum of the areas of all detection frames for the k-th type of soil block to the total area of ​​the target region. This represents the normalized value of the number of detection frames for the k-th type of soil block relative to a preset reference number. This represents the proportion of the largest detection box in the k-th soil type to the total area of ​​the target area.

[0045] Then, the embedded controller performs the operation according to preset weighting coefficients. Calculate the basic score for each type of soil block. Specifically, as shown in the formula: In the formula, This is the base score for the k-th type of soil block. This is used to balance the relative importance of area proportion, quantity indicators, and the largest single block area proportion in the base score. To reflect the differences in the impact of soil blocks of different sizes on compaction operations, importance weights are assigned to each type of soil block. The basic scores of various types of soil blocks are weighted and summed to obtain the unnormalized comprehensive soil block metric. And calculate according to the formula: .Will Normalize to preset interval Within this, the dimensionless soil block composite index CI is obtained. CI is used to characterize the current soil surface roughness and soil fragmentation, and serves as the sole driving force for the target soil pressure P.

[0046] To convert the soil compaction index (CI) into an agronomically understandable and executable compaction pressure command, this embodiment employs a monotonically increasing piecewise linear function to construct the mapping relationship between CI and the target compaction pressure P. Preferably, the soil compaction index (CI) can be... The system is divided into three sub-intervals, corresponding to light, medium, and heavy compaction levels, respectively. Within each sub-interval, a linear function with different slopes and intercepts is selected, ensuring that the overall P(CI) monotonically increases with respect to CI. When CI is low, the system assigns a lower target compaction pressure P; when CI is high, the system assigns a higher target compaction pressure P, thus achieving differentiated compaction for different soil clod conditions. Specific thresholds and compaction pressure ranges for each level can be determined through experimental calibration based on equipment type, soil type, and crop requirements.

[0047] To reduce the drastic fluctuations in the target tamping pressure P over time, this embodiment preferably introduces a deep learning time series prediction model to dynamically smooth and correct the soil composite index CI and the target tamping pressure P. The time series prediction model can be one or more of a recurrent neural network (RNN), a long short-term memory network (LSTM), or a gated recurrent unit network (GRU). Its inputs include CI, P, tamping wheel speed, electric actuator displacement L, and / or agricultural machinery travel speed at multiple consecutive sampling times. The output is the smoothed and corrected target tamping pressure and / or the corresponding target actuator displacement. This model is trained offline using historical operational data, enabling it to learn the temporal correlation between soil state changes and reasonable tamping pressure adjustments. This allows for appropriate smoothing and correction of the instantaneously calculated target tamping pressure P in actual control, effectively suppressing high-frequency jitter, reducing frequent starts and stops and repeated extensions and retractions of the electric actuator 4, improving system stability, and reducing energy consumption.

[0048] After obtaining the target pressure P after smoothing correction, the embedded controller calculates the target push rod displacement L based on the pre-calibrated monotonic correspondence between the pressure P and the electric push rod displacement L. To obtain this correspondence, different target electric push rod displacements L are applied on the test bench during the system calibration phase. i And measure the corresponding tamping pressure P i Based on a set of collected data Data fitting yields monotonic functions During field operations, the target push rod displacement is determined by numerical inverse calculation or table lookup based on the current target pressure P. The embedded controller sends the target push rod displacement L, as well as the running speed, acceleration, and start / stop signals, to the electric push rod controller via bus communication (preferably Modbus communication). The electric push rod controller then drives the electric push rod 4 to extend or retract, causing the pressing wheel 1 to move up and down relative to the frame 7, thus changing the force exerted by the pressing wheel 1 on the ground and gradually bringing the actual pressing pressure closer to the target pressing pressure. Simultaneously, the embedded controller can set a minimum displacement change threshold and a maximum adjustment frequency. When the target push rod displacement change is below the threshold and the actual pressing pressure is within the allowable error range, the current position remains unchanged. After the pressing pressure is stabilized for a preset time, the electric push rod controller enters a low-power holding mode to further reduce energy consumption and extend the actuator's lifespan.

[0049] In one specific embodiment, the particle size classification number for soil clods is K = 3, corresponding to three categories: small soil clods, medium soil clods, and large soil clods. The equivalent diameter d of the soil clod is used as the classification basis; for example, small soil clods are defined as... The middle soil block is defined as Large clods of earth are defined as By calibrating the camera's field of view relative to the ground, the total area of ​​the target region corresponding to a single image is determined as A_tot. The embedded controller then calculates the area percentage of the three types of soil blocks within each sampling period. Quantitative indicators and the percentage of the largest single area And calculate the score S for each type of soil block foundation according to formula (1). k To highlight the importance of the area factor, in this embodiment, α1 = 0.5, α2 = 0.3, and α3 = 0.2 can be used.

[0050] To highlight the more significant adverse effects of large soil clods on crop emergence and compaction quality, the importance weight β is used in this embodiment. k The preferred settings are: β(large) = 1.0 for large soil clods, β(medium) = 0.6 for medium soil clods, and β(small) = 0.3 for small soil clods. The embedded controller obtains the value according to formula (2). And by using historical operation data to obtain statistics Typical minimum value and maximum value In a linear manner Normalization to The interval is used to obtain the comprehensive soil block index CI∈[0,1]. In a certain operation, if the calculated foundation score for large soil blocks S(large) = 0.6, the foundation score for medium soil blocks S(medium) = 0.3, and the foundation score for small soil blocks S(small) = 0.1, then... = 1.0×0.6 +0.6×0.3 + 0.3×0.1 = 0.81, after normalization we get CI≈0.75, which corresponds to a relatively high degree of soil roughness.

[0051] In the above embodiment, the CI interval [0,1] is divided into three pressing level intervals: [0,0.3) corresponds to light pressing, [0.3,0.7) corresponds to medium pressing, and [0.7,1] corresponds to heavy pressing. The pressing pressure adjustment range of a certain type of pressing device is P∈[3,10] kN. Therefore, linear mappings can be set in the three intervals respectively: when CI∈[0,0.3), CI is mapped to P∈[3,5] kN; when CI∈[0.3,0.7), CI is mapped to P∈[5,8] kN; and when CI∈[0.7,1], CI is mapped to P∈[8,10] kN. This determines the slope of the piecewise linear function P(CI). and intercept Taking CI≈0.75 as an example, this value falls within the heavy suppression range, and the corresponding target suppression pressure P is approximately 8.5kN.

[0052] In an exemplary calibration experiment, the pressing device was fixed on a test bench. The displacement L of the electric push rod was controlled to vary in 5 mm increments within the range of 0–20 mm (i.e., L = 0, 5, 10, 15, 20 mm, etc.). The pressing force P of the pressing wheel 1 on the ground was measured at each position, and a set of calibration data points were obtained. An approximate linear relationship is obtained by fitting using the least squares method. For example, a = 0.25 kN / mm, b = 3 kN. During field operations, if the target compressive pressure P = 8.5 kN is calculated based on CI, the embedded controller can... The calculation is performed in reverse, and the amplitude is limited to 20 mm without exceeding the upper limit of the electric actuator stroke. This is then sent to the electric actuator controller as the target actuator displacement to achieve approximate control of the actual pressure.

[0053] Those skilled in the art will understand that the particle size range and weighting coefficients in the above embodiments... The range of pressing pressure and the specific parameters of the calibration function g(L) can be adjusted according to the soil conditions, equipment structure, and crop agronomic requirements of different regions. As long as image recognition based on soil particle size classification is still used, a comprehensive soil index CI is constructed, and the monotonic mapping relationship is used to achieve adaptive adjustment of pressing pressure and electric actuator displacement, it is considered to fall within the protection scope of this invention.

[0054] like Figure 3 As shown, the target pressure smoothing and pushrod control process based on deep learning time series prediction of the present invention includes steps S301 to S305. This process is based on... Figure 1 The method uses the soil block comprehensive index CI and its mapping to obtain the instantaneous target tamping pressure P as the basic signal. It dynamically predicts and smooths the tamping pressure command in the time dimension, and further converts it into the control quantity of the electric actuator, thereby realizing the stable and adaptive adjustment of the tamping pressure of the tamping wheel.

[0055] Step S301 involves raw data acquisition. The embedded controller acquires multi-source time-series data from the soil block identification and compaction pressure calculation module and field sensors at a preset sampling period k. This data includes at least the soil block comprehensive index CIk at the current sampling time, the instantaneous target compaction pressure Pk mapped from CIk, the electric push rod displacement Lk, the compaction wheel speed, and the agricultural machinery travel speed vk. These variables, along with time k, form multi-dimensional raw time-series data and are stored in a buffer in chronological order for constructing subsequent time window sequences.

[0056] Step S302 involves constructing the time window sequence. The embedded controller extracts multidimensional data from the most recent T sampling times from the buffer, forming a time window sequence of length T, for example... The data from each channel is processed through preprocessing operations such as normalization, outlier removal, and optional missing value imputation to obtain a standardized time series that meets the input requirements of a deep learning model.

[0057] Step S303 involves deep learning time series prediction. The multivariate time window sequence obtained in step S302 is used as input and fed into a pre-trained offline time series prediction model. The model can be one or more of a recurrent neural network (RNN), a long short-term memory network (LSTM), or a gated recurrent unit network (GRU). The deep learning model extracts features from historical sequences and models temporal correlations, outputting the predicted pressure value P1(k) for the current time and / or the next sampling time, as well as the corresponding predicted pusher displacement value L1(k), to characterize a smoother and more reasonable trend of pressure and pusher displacement changes under the current soil and operational conditions.

[0058] Step S304 involves smoothing the target ballast pressure. The embedded controller then applies the predicted ballast pressure obtained in step S303. Compared with the comprehensive index of soil blocks (CI) k The instantaneous target pressure P obtained through the mapping function k By merging, the smoothed target pressure is obtained. Preferably, a weighted fusion method can be used: In the formula, 0 ≤ ω ≤ 1 represents the weighting coefficients, used to adjust the proportion of predicted values ​​and instantaneous values ​​in the smoothing result. When updating P2(k), further limits on the maximum step size and saturation constraints are set, for example: , ,in, The target stagnation pressure for the previous control cycle is defined as ΔPmax, which is the maximum allowable single change. Pmin and Pmax are the lower and upper limits of the stagnation pressure, respectively. Optionally, a hysteresis threshold εhys can also be set. The current command is kept unchanged to suppress high-frequency jitter and frequent small adjustments. Through the above constraints, a target pressure sequence that is continuous, smooth, and within a safe range on the time axis is obtained.

[0059] Step S305 involves push rod control and execution. The embedded controller uses the monotonic function relationship between the pressure P and the push rod displacement L obtained during the calibration phase: By using numerical inverse calculation or table lookup methods, the smoothing target pressure is obtained. Calculate the target push rod displacement for the current control cycle. Subsequently, the embedded controller sends control commands, including the target displacement Lc(k) and operating speed, acceleration, and start / stop control signals, to the electric actuator controller via bus communication (preferably Modbus communication). This drives the electric actuator 4 to perform a telescopic action, changing the position of the press wheel 1 relative to the frame 7, ultimately making the actual pressing pressure follow the smooth target pressing pressure. The changes enable smooth adjustment of soil compaction pressure and control of variable force soil compaction based on deep learning time series prediction.

[0060] Optionally, when the predicted push rod displacement L1(k) output by the time series prediction model in step S303 satisfies the preset constraint conditions, the embedded controller can also use L1(k) as a direct estimate of the target push rod displacement, or perform weighted fusion with Lc(k) calculated by formula (8) to obtain the corrected target push rod displacement. and will The command is then sent to the electric actuator controller. This method allows for smooth adjustment of the pressure, while further enhancing the continuity and coordination of the actuator control commands in the time dimension, thus better implementing the control concept of "smoothly corrected target pressure and / or corresponding target actuator displacement" in this invention.

Claims

1. An image recognition based self-adaptive variable force soil compacting method, characterized in that, The method comprises the following steps: S1, installing a camera and an electric push rod on a soil compacting device, the electric push rod being connected with a compacting wheel mechanism; during operation, the camera collects soil surface images containing soil clumps; S2, inputting the soil surface images into a soil clump image recognition model deployed on an embedded controller to detect and classify the soil clumps in the images, and obtaining the number, area and distribution information of soil clumps of different particle sizes; S3, calculating a soil clump comprehensive index CI representing the roughness of the soil surface according to the distribution information of the soil clumps of different particle sizes and a preset rule, and obtaining a corresponding target compacting force P according to a preset mapping relationship between CI and the compacting force; S4, calculating a target push rod displacement L using the target compacting force P according to a monotonic corresponding relationship between the compacting force and the push rod displacement L obtained through pre-calibration, and issuing a control instruction to the electric push rod controller by the embedded controller to drive the electric push rod to extend or retract to the target push rod displacement L, so as to adjust the compacting force of the compacting wheel on the ground, so that the actual compacting force approaches the target compacting force P.

2. The method of claim 1, wherein, The soil clump image recognition model is a deep learning target detection model, and the step S2 further comprises: (1) taking the diameter of the soil block as the grading basis, assuming that the diameter of a single soil block is d, and K particle size intervals are pre-set , wherein K≥3 and satisfies < When the diameter d of a certain soil block satisfies , the soil block is divided into the kth type of soil block, thereby obtaining the multi-class soil block grading based on the particle size interval; (2) collecting soil surface images containing various types of soil clumps, and using an image labeling tool to label the class index k and the corresponding bounding box position of each soil clump in the image to form a sample data set with class labels and bounding box information; (3) dividing the sample data set into a training set, a validation set and a test set, training and validating the deep learning target detection model, determining a target model according to the performance indicators, and converting the target model into a model format suitable for running on an embedded controller for deployment.

3. The method according to claim 1 or 2, characterized in that, The calculation of the soil clump comprehensive index CI comprises: For each type of soil block (the kth type, k = 1, 2, …, K), at least one or more indicators including area ratio, quantity index and maximum single block area ratio are calculated, wherein the area ratio of the kth type of soil block is denoted as A k , the quantity index is denoted as N k , and the maximum single block area ratio is denoted as M k . The indicators are weighted and summed according to the preset weights α1, α2, α3 to obtain the basic score of the kth type of soil block According to the difference in the influence degree of different particle size soil blocks on the rolling operation, different importance weights β are set for each type of soil block k The basic scores of each type of soil block are weighted and summarized to obtain , and is normalized to the preset interval , to obtain the soil block comprehensive index CI, which is the only driving quantity of the target rolling pressure P.

4. The method according to any one of claims 1 to 3, characterized in that, The mapping relationship between the soil clod comprehensive index CI and the target ground pressure P is a monotonically increasing piecewise linear function, and specifically, the soil clod comprehensive index CI satisfies The threshold values CI1 and CI2 satisfy and a positive number and a constant such that the mapping of the target ground pressure with respect to CI satisfies: wherein, corresponding to a light compaction level, corresponding to a medium compaction level, corresponding to a heavy compaction level, and within each level interval, the target compaction force P is limited within a corresponding compaction force level range to achieve differentiated compaction to different soil clumps.

5. The method according to any one of claims 1 to 4, characterized in that, After obtaining the soil clump comprehensive index CI and calculating the target compacting force P in step S3, a pre-trained deep learning time series prediction model is further used to dynamically smooth and correct the soil clump comprehensive index CI and the target compacting force P, the input of the deep learning time series prediction model including one or more of the soil clump comprehensive index CI, the target compacting force P, the compacting wheel speed, the electric push rod displacement L and the agricultural machinery speed at consecutive sampling times, and the output being the smoothed and corrected target compacting force and the corresponding target push rod displacement; the deep learning time series prediction model is one or more of a recurrent neural network RNN, a long short-term memory network LSTM and a gated recurrent unit network GRU, and is obtained through offline training based on historical operation data to reduce the sharp fluctuations of the compacting force instruction in time and reduce the adverse effects of frequent actions of the electric push rod on the system stability and energy consumption.

6. The method according to any one of claims 1 to 5, characterized in that, The monotonic correspondence between the compacting force P and the electric push rod displacement L is obtained by calibration through experiments, and the corresponding compacting force P is collected for different push rod displacements L during the calibration process, and a monotonic function is obtained by fitting During the operation, the target push rod displacement L is obtained from the target compacting force P through numerical backstepping or table lookup method; the embedded controller sends the target push rod displacement L and one or more of the running speed, acceleration and start-stop control signals to the electric push rod controller through bus communication method, and the bus communication method is preferably Modbus communication.

7. A variable force soil packer device for implementing the method of any one of claims 1 to 6, characterized in that, The method comprises: The frame, the roller, the roller shaft connected with the roller, the fixed bearing device, the electric push rod, the fixed support or support rod for connecting the electric push rod with the frame, and the camera installed on the frame or the fixed support; one end of the electric push rod is hinged with the frame or the fixed support, and the other end is hinged with the roller shaft or a rocker arm mechanism connected with the roller shaft, the position of the roller relative to the frame is changed by the elongation or shortening of the electric push rod, the ground pressure of the roller is increased when the electric push rod is elongated, and the ground pressure of the roller is reduced when the electric push rod is shortened; the camera is electrically connected with the embedded controller, used for collecting the image of the soil surface in front of the roller, and sending the image to the embedded controller, and the electric push rod is controlled by the embedded controller according to the adaptive variable force soil roller method based on image recognition in any one of claims 1 to 6.

Images