Machine learning based intelligent control system for cyperus esculentus harvesting machine

Abstract

The application relates to the field of artificial intelligence and discloses an intelligent control system of a cyperus esculentus bean harvesting machine based on machine learning, which comprises the following steps: collecting soil moisture, canopy spectrum and historical data, predicting a tuber maturity thermal map through a three-dimensional convolution-LSTM hybrid model, fusing ground penetrating radar and vibration sensing data during operation, constructing an underground tuber point cloud in real time and dynamically adjusting the digging depth, speed and swing, instantaneously lifting a shovel to avoid obstacles when encountering sudden resistance, comparing the coverage rate after operation to generate a supplementary harvesting path, and incrementally updating model parameters after each season. The application realizes centimeter-level spatial positioning and hour-level harvesting window prediction, increases the complete harvesting rate, reduces the damage rate, improves the coverage rate and economic benefits.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an intelligent control system for a tiger nut harvester based on machine learning. Background Technology

[0002] Tiger nuts, a high-quality economic crop characterized by high oil, high protein, and high starch content, have enormous application potential in edible oil processing, feed production, and biomass energy development. In recent years, their large-scale planting area has rapidly expanded in arid and semi-arid regions of North and Northwest my country. However, the contradiction between the growth characteristics of their underground tubers and traditional mechanized harvesting techniques has become a core bottleneck restricting the large-scale development of the industry. The specific challenges are concentrated in the following three aspects: Traditional tiger nut harvesting relies entirely on manual experience, primarily by observing the above-ground parts of the plant or randomly digging up a small number of tubers to determine maturity. This presents two major problems: manual sampling can only cover a small area of ​​the plot and cannot reflect the spatial heterogeneity of maturity caused by different soil moisture, micro-topography, and uneven historical fertilization, easily leading to over-harvesting or premature harvesting of unripe tubers; the optimal harvesting period for tiger nut tubers after maturity is only 72-96 hours, and traditional manual judgment often results in delayed or early harvesting due to the lack of quantitative indicators. According to actual measurement data from planting bases in North China, manual judgment can easily lead to economic losses.

[0003] Existing tiger nut harvesters mostly use fixed operating parameters with a single cut, which cannot adapt to the dynamic distribution characteristics of underground tubers, resulting in a high damage rate: if the digging shovel penetrates the soil to a depth lower than the median burial depth of the tubers, it is easy to cause breakage damage; if the depth is too high, although it can reduce breakage, it will increase soil resistance, causing tubers to be squeezed and rubbed against the soil, increasing the squeezing damage rate; at the same time, the fixed travel speed is prone to conveying congestion in high-density tuber areas, further increasing the tuber collision damage rate; low spatial coverage: traditional machinery relies on pre-defined row spacing paths for operation, without considering the cross-row distribution of tubers, resulting in a high rate of omission in edge areas; and it lacks real-time underground imaging capabilities, making it unable to identify obstacles such as stones and hard soil clods. To avoid mechanical failure, manual shutdown and clearing are often required, further reducing the continuity of operation. Ultimately, the overall spatial coverage is low, requiring secondary harvesting and increasing operating costs.

[0004] Traditional soybean harvesters rely on fixed program logic for control, making it impossible to adjust parameters based on soil characteristics and climate variations in the planting area. This leads to difficulties in cross-regional adaptation: in sandy soil areas, a fixed safety margin can cause excessive digging, increasing energy consumption; in clay soil areas, a fixed travel speed can result in wasted runs, reducing operational efficiency; in areas with significant annual rainfall variations, the tuber integrity rate varies greatly under the same digging parameters due to differences in soil compaction, requiring manual recalibration, which is complex and time-consuming; long-term performance degradation: as the planting season progresses, changes in soil fertility and abnormal climate can cause the tuber growth model to fail, but traditional machinery cannot self-correct using historical operational data, requiring replacement of core components to restore performance, increasing maintenance costs. Summary of the Invention

[0005] This invention provides an intelligent control system for a tiger nut harvester based on machine learning. By constructing a multi-source heterogeneous sensor data fusion architecture and combining a historical growth model of the plot with a real-time underground tuber distribution prediction algorithm, it achieves accurate determination of the optimal harvesting time for tiger nuts and dynamic planning of the harvesting path. Before operation, the system collects data on soil moisture, plant canopy spectral reflectance, root zone thermodynamic distribution, and historical yield. Through joint modeling using convolutional neural networks and long short-term memory networks, it outputs a spatial heat map of tuber maturity and a harvesting priority sequence. During operation, the system adjusts the digging depth and travel speed in real time based on underground radar echo signals and vibration sensor feedback, ensuring a tuber integrity harvesting rate of over 95% while avoiding resource waste and mechanical damage caused by premature harvesting or path deviation.

[0006] According to one aspect of the present invention, a machine learning-based intelligent control system for a tiger nut harvester is provided, comprising: The multi-dimensional sensing module for the site environment is used to collect data on soil moisture content, surface temperature gradient, near-infrared reflectance spectrum of plant canopy, historical fertilization records, and accumulated temperature data of the target site before harvesting. The tuber maturity spatiotemporal prediction module is used to input the data collected by the multi-dimensional perception module of the plot environment into a pre-trained three-dimensional convolutional long short-term memory hybrid neural network model. The model takes the plot spatial coordinates as the input dimension and the time series growth parameters as the temporal dimension, and outputs the maturity probability distribution map of tiger pea tubers and the optimal harvest time window in each sampling unit. The underground tuber distribution positioning module is used to transmit low-frequency electromagnetic pulses through a ground-penetrating radar array installed at the front end of the digging shovel during the harvester's movement, receive echo signals reflected by the underground tuber group, and combine echo amplitude attenuation rate, phase offset and multi-channel time difference positioning algorithm to construct a three-dimensional spatial coordinate point cloud map of the underground tubers in real time. An adaptive excavation control module is used to dynamically adjust the soil penetration depth, travel speed, and lateral swing amplitude of the excavation shovel based on the maturity probability distribution map output by the tuber maturity spatiotemporal prediction module and the three-dimensional spatial coordinate point cloud map output by the underground tuber distribution positioning module. The soil penetration depth is determined based on the median of the tuber burial depth distribution plus a preset safety margin. The travel speed is adjusted inversely proportional to the tuber density per unit area and the excavation efficiency function. The lateral swing amplitude is compensated and corrected based on the overlap of adjacent rows of tuber distribution. The damage avoidance feedback module is used to monitor the sudden change signal of digging resistance during the digging operation by using a vibration acceleration sensor array installed on the side wall of the digging shovel. When the instantaneous acceleration exceeds the preset threshold, the digging depth fine adjustment command is immediately triggered, so that the digging shovel is raised by 5cm within 0.5s and maintained for 3s before returning to the original trajectory, in order to avoid mechanical impact caused by hard obstacles or dense tuber groups. The path planning and missed mining module is used to compare the spatial coverage of the pre-harvest maturity heat map with the actual harvesting trajectory after the first round of mining operations. It generates a secondary mining path for areas with a coverage of less than 90%, and schedules the harvester to perform fixed-point re-harvesting along the path in half-speed mode until the spatial coverage reaches the target. The online model update module is used to collect data on actual harvest volume, tuber damage rate, unharvested residue rate, and operational energy consumption after each harvest season. This data is used as a supervisory signal to incrementally train the three-dimensional convolutional long short-term memory hybrid neural network model, enabling the model parameters to adaptively optimize as the soil characteristics and climate patterns of the planting area evolve year by year.

[0007] As a preferred embodiment of the present invention, the plant canopy near-infrared reflectance spectrum acquisition device in the multi-dimensional sensing module of the plot environment adopts a multispectral imager with a spectral channel covering the 700nm to 1000nm band, a spatial resolution of 5cm per pixel, an acquisition height of 1.2m above the top of the canopy, and a scanning speed synchronized with the harvester's travel speed to ensure that each plant unit is scanned at least three times to eliminate momentary shadow interference.

[0008] In a preferred embodiment of the present invention, the three-dimensional convolutional long short-term memory hybrid neural network model used in the tuber maturity spatiotemporal prediction module receives five-dimensional tensor data in its input layer. The first dimension is the row coordinates of the plot, the second dimension is the column coordinates of the plot, the third dimension is the vertical profile depth layer, the fourth dimension is the time step, and the fifth dimension is the number of environmental parameter channels. Its hidden layer is composed of three layers of three-dimensional convolutional kernels stacked together, each convolutional kernel having a size of 3×3×3 and a stride of 1, followed by batch normalization and modified linear unit activation functions. Its temporal modeling layer is composed of bidirectional long short-term memory units, with a hidden state dimension of 256. The final output layer is a fully connected layer with a soft maximum function, outputting the maturity probability value of each spatial unit.

[0009] In a preferred embodiment of the present invention, the ground-penetrating radar array in the underground tuber distribution positioning module consists of 12 transceiver antennas, arranged equidistantly along the digging shovel, with a center frequency of 900MHz, a pulse repetition frequency of 20KHz, a sampling interval of 0.1ns, an effective detection depth of 0.3m to 0.8m, a lateral resolution of 8cm, and a longitudinal resolution of 3cm. Its echo signal processing flow includes: removing DC offset from the original signal, median filtering of background noise, extracting the envelope using Hilbert transform, temporal threshold segmentation, peak clustering positioning, and spatial interpolation to generate a continuous point cloud.

[0010] In a preferred embodiment of the present invention, the digging depth adjustment mechanism in the adaptive digging control module has a safety margin set to twice the standard deviation of the tuber burial depth distribution. When the standard deviation is less than 5cm, the safety margin is fixed at 10cm. The travel speed adjustment function is: the speed is equal to the base speed multiplied by the tuber density threshold divided by the current area tuber density, where the base speed is 3km / h and the tuber density threshold is 150 tubers per square meter. The lateral swing compensation is equal to the width of the overlapping area of ​​adjacent rows of tuber distribution multiplied by 0.8, and the maximum swing is less than 20% of the width of the digging shovel.

[0011] In a preferred embodiment of the present invention, the vibration acceleration sensor array in the damage avoidance feedback module consists of eight triaxial accelerometers, which are respectively installed on the left and right side walls of the digging shovel at distances of 10cm, 20cm, 30cm and 40cm from the cutting edge. The sampling frequency is 5KHz, the trigger threshold is set to 15 times the value of gravitational acceleration, and the fine-tuning command is generated by the embedded programmable logic controller within 5ms after the threshold is detected to be exceeded. The hydraulic cylinder is controlled by the proportional-integral-derivative servo driver to perform the lifting action.

[0012] As a preferred embodiment of the present invention, the spatial coverage calculation method in the path planning and missed mining module is as follows: the area with a maturity probability greater than 0.7 in the pre-harvest maturity heat map is defined as the mining area, the actual harvesting trajectory is projected onto the same coordinate system, and the proportion of the intersection area of ​​the mining area and the trajectory coverage area to the total area of ​​the mining area is calculated; the secondary mining path generation algorithm adopts an improved ant colony optimization algorithm, and its pheromone update rule introduces a terrain slope penalty factor and a historical omission frequency weighting factor to ensure that the mining path prioritizes coverage of high omission risk areas.

[0013] In a preferred embodiment of the present invention, the incremental training strategy in the online model update module adopts an elastic weight solidification algorithm, which optimizes newly collected data by small-batch gradient descent while retaining historical task knowledge. The initial learning rate is 0.001, decays by 5% per training round, and the number of training rounds is fixed at 50. The loss function is a weighted sum of cross-entropy loss and mean squared error loss, with weight coefficients of 0.7 and 0.3, respectively.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention completely abandons the traditional model of relying on human experience to judge the timing of harvest. By constructing an intelligent decision-making architecture that integrates multi-source sensor data and spatiotemporal prediction models, it achieves centimeter-level spatial positioning and hour-level time window prediction of the maturity status of tiger nut tubers, thereby reducing the error rate of harvest decision-making.

[0015] During operation, the system uses real-time underground radar positioning and vibration feedback control to improve the integrity rate of tuber harvesting, which is better than traditional mechanical operations, while controlling the tuber damage rate.

[0016] Path planning and a missed sampling mechanism ensure comprehensive coverage across large areas, improving spatial harvesting coverage and avoiding secondary operation costs due to omissions. The model's online update capability enables continuous system evolution, adapting to different soil types, climate conditions, and planting patterns without requiring manual parameter recalibration, thus reducing operational complexity.

[0017] The entire system upgrades the mechanized harvesting of tiger nuts from extensive to precise and intelligent operations, improving the economic benefits per unit area and providing core technical support for large-scale industrialized tiger nut cultivation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent control system for a tiger nut harvester based on machine learning proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the tuber maturity spatiotemporal prediction module in this invention; Figure 3 This is a flowchart illustrating the logical framework of the linkage between underground tuber distribution location and adaptive excavation control in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the damage avoidance feedback and path planning supplementary mining collaborative mechanism in this invention; Figure 5 This is a logical flowchart of the online model update module and the closed-loop evolution architecture of the entire system in this invention; Detailed Implementation

[0019] Please refer to Figures 1 to 5As one embodiment of the present invention, the machine learning-based intelligent control system for tiger pea harvesters activates a multi-dimensional field environment perception module before operation. This module is deployed on a multi-sensor integrated platform mounted on the front end of the harvester. Its function is to collect soil moisture content, surface temperature gradient, near-infrared reflectance spectrum of plant canopy, historical fertilization records, and meteorological accumulated temperature data of the target operation area.

[0020] Soil moisture content was obtained using a frequency domain reflectance method sensor array. The sensor probe was inserted vertically into the ground to a depth of 10 cm. The sampling interval was one measurement point per square meter, the data acquisition frequency was once per second, and the acquisition duration was the entire process of the harvester passing through the measurement point area.

[0021] The surface temperature gradient was measured by an infrared thermal imager array with the imager lens facing the ground, the field of view covering a 3m×3m area in front, the spatial resolution being 2cm, the temperature measurement accuracy being ±0.5℃, and the data frame rate being 10fps. The effects of instantaneous thermal disturbances were eliminated by the time averaging method.

[0022] The near-infrared reflectance spectrum of the plant canopy was acquired by a multispectral imager mounted on a stable gimbal 1.2m above the top of the canopy. The imager covered the 700nm to 1000nm band and had eight narrowband filters with center wavelengths of 720nm, 760nm, 800nm, 840nm, 880nm, 920nm, 960nm, and 990nm. Each channel was imaged independently with a spatial resolution of 5cm per pixel. The scanning speed was synchronized with the harvester's travel speed to ensure that each plant unit was scanned three times in the direction of travel. The three scan data were fused using the maximum reflectance retention method to eliminate momentary leaf shading or shadow interference.

[0023] Historical fertilization records are retrieved from the vehicle-mounted database. The database pre-enters the types, amounts, times, and methods of fertilization applied to the plot over the past three planting seasons. The data is in a structured table format, with fields including fertilization date, nitrogen, phosphorus, and potassium content, organic matter content, application depth, and type of application machinery.

[0024] The accumulated temperature data is obtained in real time from the meteorological station interface. The meteorological station is located at the edge of the plot and uploads temperature, humidity, wind speed and sunshine duration data every 10 minutes. The system internally calculates the effective accumulated temperature since the sowing date. The calculation formula is the daily average temperature minus the biological zero degree, i.e. 10℃, and the positive values ​​are added together, while the negative values ​​are recorded as 0.

[0025] As one embodiment of the present invention, the five types of data collected by the multi-dimensional perception module of the plot environment are synchronously transmitted to the tuber maturity spatiotemporal prediction module. This module is deployed in the vehicle-mounted industrial control computer, and its core is a pre-trained three-dimensional convolutional long short-term memory hybrid neural network model.

[0026] The model's input layer receives five-dimensional tensor data. The first dimension is the row coordinates of the plot, with the harvester's direction of travel as the reference, a sampling interval of 10cm, covering a working width of four meters, for a total of forty sampling points. The second dimension is the column coordinates of the plot, perpendicular to the direction of travel, with a sampling interval of 10cm, and the working length is dynamically adjusted according to the plot boundaries. The third dimension is the vertical profile depth stratification, divided into five levels: 0 to 10cm, 10 to 20cm, 20 to 30cm, 30 to 40cm, and 40 to 50cm, corresponding to the main distribution depth range of tubers. The fourth dimension is the time step, in weeks, covering all weeks from the sowing date to the current date in the current planting season. The fifth dimension is the number of environmental parameter channels, a total of five channels, corresponding to soil moisture content, mean surface temperature gradient, first principal component of near-infrared reflectance spectrum, cumulative value of historical fertilizer nitrogen content, and cumulative accumulated temperature value.

[0027] The input tensor is processed by three stacked 3D convolutional kernels, each with a size of 3×3×3, a stride of 1, and no padding. After convolution, a batch normalization layer and a modified linear unit activation function are applied. The first layer has 64 output channels, the second has 128, and the third has 256. The convolutional feature map is then input into a bidirectional long short-term memory (LSTM) unit, whose hidden state dimension is 256. The forward and backward states are concatenated along the time dimension to output a 512-dimensional feature vector. The final output layer is a fully connected layer with a soft-maximization function. The fully connected layer has an input dimension of 512 and an output dimension of 2, corresponding to the immature and mature probabilities, respectively. The soft-maximization function normalizes the output into a probability distribution.

[0028] The model outputs the maturity probability value for each spatial sampling unit. Units with a probability value greater than 0.7 are marked as areas to be harvested. The model also outputs the optimal harvesting time window for that unit. The time window is calculated from the current date, extending forward 72 hours to the next 48 hours. The time point corresponding to the peak probability within the time window is the recommended harvesting time.

[0029] As one embodiment of the present invention, the maturity probability distribution map output by the tuber maturity spatiotemporal prediction module is transmitted to the path planning and missed harvesting module. This module first converts the probability distribution map into a grid map. The grid resolution is consistent with the input tensor, which is 10cm×10cm. Each grid stores the maturity probability value and the recommended harvesting timestamp.

[0030] The path planning algorithm employs an improved ant colony optimization algorithm. The pheromone matrix is ​​initialized as an all-zero matrix, and the number of ants is set to 100. Each ant starts from the top-left corner of the plot and selects the next moving grid according to the transition probability formula. The transition probability is jointly determined by the pheromone concentration, heuristic factor, terrain slope penalty factor, and historical omission frequency weighting factor. The heuristic factor is defined as the square of the maturity probability value. The terrain slope penalty factor is calculated by the digital elevation model: 0.5 for areas with a slope greater than 15%, 1.2 for areas with a slope less than 5%, and 1 for all other areas. The historical omission frequency weighting factor is generated from the previous season's operation records, with an initial value of 1, increasing by 0.3 for each omission, up to a maximum of 2.5.

[0031] After completing one traversal, the ants update the pheromone based on the product of the total maturity probability of the path and the terrain slope penalty factor. The pheromone volatility coefficient is 0.9, and the number of iterations is 500. Finally, the path with the highest total fitness is selected as the first round of harvesting trajectory. This trajectory is output as a sequence of coordinate points with an adjacent point spacing of 20cm, and includes instructions on the direction of travel and turning angle, which is then sent to the harvester's navigation and control system for execution.

[0032] In one embodiment of the present invention, as the harvester travels along the planned path, the underground tuber distribution positioning module is activated in real time. This module consists of a ground-penetrating radar array installed at the front end of the digging shovel. The array includes 12 sets of transceiver antennas, equidistantly arranged along the transverse side of the digging shovel, with an adjacent antenna spacing of 25 cm, a center frequency of 900 MHz, a pulse repetition frequency of 20 kHz, a sampling interval of 0.1 ns, and an effective detection depth of 0.3 m to 0.8 m. The radar emits low-frequency electromagnetic pulses, which are reflected at the tuber interface during underground propagation. The reflected echo is captured by the receiving antenna. The original signal is first processed to remove DC offset, subtracting the signal mean to eliminate the system's DC component.

[0033] Then, median background noise filtering is performed with a filtering window length of 50 sampling points to eliminate random high-frequency noise.

[0034] The filtered signal is subjected to Hilbert transform to extract the signal envelope, and the peak value of the envelope corresponds to the location of the reflecting interface.

[0035] The envelope signal is then segmented by a time-domain threshold, with the threshold value set at 30% of the maximum value of the envelope. Continuous segments above the threshold are marked as effective reflection segments.

[0036] Peak clustering was used to locate the effective reflection segment. The clustering algorithm adopted density clustering, with a neighborhood radius of 3cm and a minimum number of points of 3. The cluster center is the coordinate of the tuber center.

[0037] Clustered coordinates are spatially interpolated to generate a continuous point cloud. The interpolation method is the inverse distance weighted method with a weight exponent of 2 and an interpolation grid resolution of 5cm×5cm×5cm. The final output is a three-dimensional spatial coordinate point cloud map with a point cloud density of more than 200 points per cubic meter and a coordinate accuracy better than 3cm.

[0038] As one embodiment of the present invention, the three-dimensional spatial coordinate point cloud map output by the underground tuber distribution positioning module and the maturity probability distribution map output by the tuber maturity spatiotemporal prediction module are synchronously input into the adaptive excavation control module. This module is deployed in the hydraulic servo controller and its function is to dynamically adjust the soil penetration depth, travel speed and lateral swing amplitude of the excavation shovel.

[0039] The burial depth is determined by adding a preset safety margin to the median of the tuber burial depth distribution. The burial depth distribution is obtained by statistically analyzing the vertical coordinates of all points in the point cloud map. The median is calculated using a fast selection algorithm. The safety margin is set to twice the standard deviation of the burial depth distribution. When the standard deviation is less than 5cm, the safety margin is fixed at 10cm.

[0040] The travel speed is inversely adjusted based on the tuber density per unit area and the excavation efficiency function. Tuber density is determined by projecting the point cloud map onto the ground plane and counting the number of points per square meter. The excavation efficiency function is defined as speed equal to a base speed multiplied by a tuber density threshold divided by the current area's tuber density. The base speed is 3 km / h, and the tuber density threshold is 150 tubers per square meter. Speed ​​increases when density is below the threshold and decreases when density is above the threshold, with the speed adjustment range limited to 1 km / h to 5 km / h. Lateral swing amplitude is compensated based on the overlap of tuber distribution in adjacent rows. The overlap is calculated by projecting the tuber point clouds of the current row and the adjacent previous row onto the same lateral profile and calculating the ratio of the intersection area to the union area of ​​the two projected regions. The compensation amount is equal to the width of the overlapping area multiplied by 0.8. The maximum swing amplitude is less than 20% of the excavator shovel width. The swing amplitude command is implemented by a servo motor driving the excavator shovel's lateral slide rail, with a response delay of less than 0.1 seconds. In one embodiment of the present invention, during the execution of the digging command by the adaptive digging control module, the damage avoidance feedback module synchronously monitors the changes in digging resistance. This module consists of eight triaxial accelerometers, which are installed on the left and right side walls of the digging shovel at distances of 10cm, 20cm, 30cm and 40cm from the cutting edge, respectively. The sampling frequency is 5KHz, and the triaxial data is combined into a total acceleration value after low-pass filtering. The filter cutoff frequency is 500Hz.

[0041] The total acceleration value is compared with a preset threshold, which is set to 15 times the gravitational acceleration value. When any sensor detects that the instantaneous acceleration exceeds the threshold, a fine-tuning command for digging depth is immediately triggered. The fine-tuning command is generated by the embedded programmable logic controller within 5ms, and the command content is to raise the digging shovel by 5cm and maintain it for 3s before returning to the original trajectory.

[0042] The lifting action is controlled by a proportional-integral-derivative servo drive that controls the hydraulic cylinder. A hydraulic cylinder stroke sensor provides real-time position feedback, achieving a closed-loop control accuracy of ±1 millimeter. The acceleration during lifting is limited to five times the acceleration due to gravity to prevent secondary impacts. During fine-tuning, the travel speed is reduced to 50% of its original speed, and lateral swing is paused. Normal parameters are restored after the original trajectory is resumed. This mechanism ensures that the digging shovel can instantly avoid hard rocks or dense tuber clusters, preventing blade breakage or tuber crushing damage.

[0043] As one embodiment of the present invention, after the harvester completes the first round of harvesting, the path planning and missed harvesting module starts a second analysis. The module first calculates the spatial coverage rate by defining the area with a maturity probability greater than 0.7 in the pre-harvest maturity heat map as the area to be harvested, projecting the actual harvesting trajectory onto the same coordinate system, and calculating the proportion of the intersection area of ​​the area to be harvested and the trajectory coverage area to the total area of ​​the area to be harvested.

[0044] If the coverage rate is below 90%, a secondary sampling path is generated. The sampling path generation algorithm reuses the improved ant colony optimization algorithm, but the initial pheromone matrix inherits the results of the first round and is superimposed with a historical omission frequency weighting factor to ensure priority coverage of high omission risk areas. The sampling path is output as a sequence of coordinate points with a spacing of 10cm between adjacent points. The travel speed is set to 50% of the baseline speed of the first round, while the digging depth and swing amplitude parameters remain unchanged.

[0045] The harvester performs targeted re-harvesting along the re-harvesting path. During the re-harvesting process, the underground tuber distribution positioning module continuously works, updating the point cloud map in real time. When the tuber density in a certain area is detected to be less than ten tubers per square meter, it is determined that the harvest is complete, and the area is marked as covered. The re-harvesting is repeated until the spatial coverage rate reaches the standard or the plot boundary is closed.

[0046] As one embodiment of the present invention, the online model update module is activated after the harvest operation of each season. This module collects data on actual harvest volume, tuber damage rate, unharvested residue rate, and operation energy consumption.

[0047] The actual harvested amount is recorded in real time at the unloading hopper by weighing sensors, and the data is stored in plots. The tuber breakage rate is collected by a vision sorting machine on the conveyor belt. Breakage is defined as tubers with a broken skin area greater than 5% of the total area. The sorting machine captures images at 10fps, and the broken tubers are identified and counted by a convolutional neural network classifier. The unharvested residual rate is obtained by aerial photography within three days after harvesting using a drone equipped with a multispectral camera. The aerial photography altitude is 50m and the resolution is 5cm. By comparing with the pre-harvest heat map, the unharvested areas are identified and the area percentage is calculated. Energy consumption during operation is recorded by fuel flow meter and electricity meter, with data accumulated over operating hours. These four types of data serve as supervisory signals for incremental training of the 3D convolutional long short-term memory hybrid neural network model. The training strategy employs an elastic weighting algorithm, first calculating the importance weights of the model parameters. These weights are determined by the gradient variance of the parameters in historical tasks; the larger the variance, the higher the weight. New data is input in mini-batches of 32, with an initial learning rate of 0.001, decaying by 5% per training round, and a fixed training duration of 50 rounds.

[0048] The loss function is a weighted sum of cross-entropy loss and mean squared error loss. The weighting coefficient of cross-entropy loss is 0.7, which is used to optimize maturity probability prediction, and the weighting coefficient of mean squared error loss is 0.3, which is used to optimize harvest time window regression.

[0049] During training, the parameter update amount is multiplied by the reciprocal of the importance weight to ensure that the change in important parameters is limited and to prevent catastrophic forgetting.

[0050] After training is complete, the new model version overwrites the old version, and the new model is loaded after the system restarts, thus achieving closed-loop evolution.

[0051] As one embodiment of the present invention, the machine learning-based intelligent control system for tiger nut harvesters was deployed and applied in a 2,000-mu tiger nut planting base in the North China Plain during field trials over three consecutive planting seasons.

[0052] The soil type of the experimental plot is sandy loam, the previous crop was corn, the average annual precipitation is 600 mm, and the frost-free period is 180 days.

[0053] Before the first harvest season, the system initialized its model using historical data, including soil testing reports, meteorological records, fertilization logs, and manual harvesting assessment forms from the same plot over the past five years. During the first harvest season, the system's average error in determining the optimal harvesting time window was 4.3 hours, lower than the 36-hour error of manual judgment. During harvesting, the measured intact tuber harvest rate, damage rate, and spatial coverage rate were all superior to those of traditional mechanical harvesting methods. After the first harvest season, the model's online update module collected supervisory signals, and incremental training improved the model's accuracy in predicting maturity probability in the second season, while reducing the harvesting time window error. The third season further optimized accuracy and error. The cumulative economic benefit per unit area improved across the three seasons, and the mechanical failure rate decreased, demonstrating the system's ability to continuously evolve and its stable operational performance.

[0054] In one embodiment of the present invention, data transmission between the modules of the system is achieved through a controller local area network bus with a bus baud rate of 500kbps and a data frame format conforming to the International Organization for Standardization 1898 standard.

[0055] The multi-dimensional sensing module for the site environment generates 8,000 bytes of data packets per second, containing raw data from five types of sensors and timestamps; The tuber maturity spatiotemporal prediction module generates a 200-byte result package per square meter of area processed, containing the maturity probability and timestamp of 40 grids; the underground tuber distribution positioning module generates a 12,000-byte point cloud data package per second, containing approximately 300 three-dimensional coordinate points. The adaptive excavation control module generates a 64-byte control instruction packet every 0.1 seconds, containing three parameters: depth, velocity, and swing amplitude; the damage avoidance feedback module generates a 16-byte status packet every 5 ms, containing acceleration values ​​from eight sensors. The path planning and missing data collection module generates a 4000-byte path package containing a sequence of 200 coordinate points for every hectare of work completed; the online model update module generates a 50-byte training package each quarter, containing supervision signals and model parameters.

[0056] All data packets carry a cyclic redundancy check code. Data packets that fail the check are discarded and retransmission is requested to ensure the reliability of system communication.

[0057] The power system is powered by a 24V DC battery. The power consumption of each module has been optimized, and the total power consumption is less than 800W, which meets the needs of long-term field operations.

[0058] As one embodiment of the present invention, the system has a fault-tolerant mechanism under extreme working conditions. When the ground-penetrating radar array experiences signal attenuation due to mud coverage, the system automatically switches to pure prediction mode, adjusting the excavation parameters solely based on the output of the tuber maturity spatiotemporal prediction module, while simultaneously activating an audible and visual alarm to prompt the operator to clean the radar antenna.

[0059] When part of the vibration acceleration sensor array fails, the system uses interpolation of data from adjacent sensors as a replacement. The interpolation method is linear interpolation. If three sensors fail consecutively, the excavation is suspended and manual intervention is requested.

[0060] When the online model update module is unable to upload data due to network interruption, the monitoring signals are stored locally and uploaded in batches after the network is restored, triggering delayed training. When the path planning module times out, the system activates a backup path, which is a fixed pattern of spiraling inward along the plot boundary to ensure uninterrupted operation. All fault tolerance mechanisms are managed through a state machine, with clearly defined state transition conditions to prevent the system from getting stuck in deadlock or waiting indefinitely.

[0061] In one embodiment of the present invention, the human-machine interface of the system is deployed on the touch screen in the cab, and the interface is divided into four main pages: operation monitoring page, parameter setting page, history page, and system diagnostic page.

[0062] The operation monitoring page displays a real-time maturity heatmap, underground point cloud map, excavation parameter curves, and spatial coverage progress bar. The parameter settings page allows operators to manually adjust parameters such as safety margin, density threshold, and swing amplitude limit, with the adjustment range protected by software limits. The history page stores operational reports from the past ten quarters, including harvest yield curves, damage rate statistics, and model version change logs. The system diagnostic page displays the communication status of each module, sensor health, power supply voltage, and storage space utilization. Abnormal conditions are highlighted in red.

[0063] The interface implements hierarchical access control. Regular operators can only view monitoring and historical records, administrators can modify parameters and trigger diagnostics, and engineers can access underlying logs and firmware upgrades. All operation records are encrypted and stored in an unalterable manner, meeting agricultural machinery data compliance requirements.

[0064] As one embodiment of the present invention, the hardware platform of the system adopts industrial-grade components, the central processing unit is an Intel Core i7 processor with a main frequency of 2.4GHz, the memory is 16GB, the solid-state drive capacity is 1TB, the operating system is a real-time Linux kernel, and the task scheduling cycle is 1ms.

[0065] The sensor interface uses a hybrid architecture of Universal Serial Bus (USB) and Controller Area Network (CAN). The USB is used for high-bandwidth devices such as multispectral imagers, while the CAN is used for low-speed sensors such as temperature probes.

[0066] The actuator driver is a proportional-integral-derivative servo type with a response bandwidth of 100Hz and a position resolution of 0.1 millimeters.

[0067] The communication module supports dual-mode 4G Long Term Evolution and WLAN to ensure field data transmission and remote monitoring.

[0068] The enclosure has an international protection rating of Class 67, is dustproof and waterproof, and has an operating temperature range of -20°C to 50°C, meeting the requirements for all-weather operation.

[0069] The system has a mean time between failures (MTBF) of more than 5,000 hours, and the maintenance cycle is to replace the hydraulic oil and clean the sensors every 500 hours, resulting in lower maintenance costs than traditional machinery.

[0070] In one embodiment of the present invention, the adaptability of the system to different soil types and climatic conditions is achieved through an online model update module.

[0071] In sandy soil areas, where tubers are buried at a shallow depth, the model automatically reduces the safety margin baseline value after one season of training. In clay regions, the tubers are dispersed, and the model adjusts the density threshold to 120 tubers per square meter; in arid climates, the surface temperature gradient characteristics are enhanced, and the model increases the weight of temperature channels. In rainy climates, soil moisture content fluctuates greatly, so the model is enhanced with temporal difference features of moisture content.

[0072] Three seasons of trial data show that the system's complete recovery rate fluctuates little in sandy soil, loam, and clay soil, and its performance degradation is minimal within the annual precipitation range of 400mm to 800mm, proving its ability to be promoted across regions. The system does not require manual parameter recalibration and can adapt to new environments simply by online updates, significantly reducing the technical threshold and maintenance burden for users.

[0073] In one embodiment of the invention, the system interfaces with an external agricultural management platform via an application programming interface (API). The interface protocol is Hypertext Transfer Protocol (HTTP), and the data format is JavaScript object notation. After completing one hectare of work, the system automatically uploads a work report, including plot number, work time, harvest yield, damage rate, coverage rate, energy consumption, and model version number. The external platform can issue work instructions, such as forcing harvesting of a specific area, adjusting harvesting priority, or pausing specific modules. The interface features authentication and data encryption. Authentication uses a hash message authentication code, and encryption uses a 256-bit Advanced Encryption Standard (AES) key to ensure data security. The system also supports offline mode; when there is no network connection, all data is stored locally and automatically synchronized upon network restoration, without affecting the continuity of field operations.

[0074] As one embodiment of the present invention, the economic benefit analysis of the system is based on three years of data from a 2,000-mu experimental field. The traditional manual judgment plus mechanical harvesting model costs 320 yuan per mu, including 50 yuan for manual judgment, 180 yuan for mechanical operation, 60 yuan for secondary harvesting, and 30 yuan for loss depreciation. The cost per mu for this system model is 260 yuan, including 80 yuan for equipment depreciation, 60 yuan for energy, 40 yuan for maintenance, no secondary harvesting cost, and 20 yuan for loss depreciation. The increased income per mu comes from improved integrity and reduced omissions. The traditional model yields 400 kg per mu, while this system yields 510 kg. At a market price of 4 yuan per kg, this translates to an increased income of 440 yuan per mu. The cumulative net income increase per mu over three years is 500 yuan, with a payback period of one and a half years, demonstrating a high internal rate of return and economic advantages. Simultaneously, the system reduces diesel consumption, lowering carbon emissions and aligning with the trend of green agriculture.

[0075] As one embodiment of the invention, this system automates harvesting decisions, reducing reliance on manual experience and enabling young practitioners to quickly master large-scale planting. The system outputs standardized operation reports, providing reliable data for agricultural insurance loss assessment, government subsidy calculation, and scientific research data collection, thus promoting the transformation of the tiger nut industry from extensive to precision management. In demonstration and promotion, surrounding farmers have increased their per-acre income by leasing the system's services, promoting balanced regional agricultural economic development.

[0076] As one embodiment of the present invention, future expansion directions of the system include integrating a pest and disease identification module, which adds an ultraviolet band to the multispectral imager to identify plant lesions and pest traces, providing early warnings and adjusting harvesting strategies; integrating a soil nutrient rapid testing module, which collects soil samples during excavation and analyzes nitrogen, phosphorus, and potassium content on-site using a near-infrared spectrometer to generate fertilization recommendation maps; and integrating an unmanned driving module, deeply coupled with path planning, to achieve fully automated operation. All expansions are based on reserved interfaces on the existing hardware platform, and the software supports hot-swapping through modular design, ensuring sustainable system upgrades and maintaining technological leadership.

[0077] In one embodiment of the present invention, the system, in actual deployment, is equipped with an operation training manual, a fault code quick reference table, and a remote technical support hotline. The training manual includes system startup procedures, daily operation guidelines, and emergency handling steps, presented in a clear and concise manner with illustrations. The fault code quick reference table lists all possible alarm codes and their meanings; for example, code E001 indicates radar signal loss, code E002 indicates acceleration exceeding limits, and code E003 indicates model training failure, facilitating rapid on-site diagnosis. The remote technical support hotline provides 24 / 7 service, allowing engineers to access the system remotely to view logs, adjust parameters, and push firmware updates, ensuring timely resolution of user issues and improving customer satisfaction and brand loyalty.

[0078] In one embodiment of the present invention, the system's extreme weather response strategy includes automatically reducing its travel speed to 70% of the baseline value when encountering heavy rain causing a sudden increase in soil moisture content, increasing the safety margin to 15cm to prevent the digging shovel from slipping or the tubers from being trapped in mud and difficult to separate. When encountering high temperatures causing the surface temperature to exceed 40°C, the system increases the weight of the infrared reflectance spectral channel and reduces the influence of the temperature gradient channel to prevent the model from misinterpreting plant stress as a maturity signal. When encountering strong winds causing the harvester to shake, the vibration sensor array initiates dynamic threshold adjustment, temporarily raising the threshold to 20 times the value of gravitational acceleration to avoid falsely triggering lifting commands. All strategies are triggered in real time through environmental perception modules, requiring no manual intervention and ensuring stable operation of the system under harsh conditions.

[0079] As one embodiment of the present invention, the system's data privacy protection mechanism complies with agricultural data security standards. All collected plot data, operational data, and model parameters are stored on local solid-state drives and may not be uploaded to the cloud without user authorization. Uploaded data is anonymized, removing precise plot coordinates and farmer identity information, retaining only the area code and crop type. Model training is completed locally, with only gradient update values ​​uploaded, not the original data, to prevent data leakage. The system has a built-in data erasure function, allowing users to clear all stored data with a single click, satisfying data sovereignty requirements. When deployed on cooperative farms, a data usage agreement is signed, clarifying data ownership and scope of use, building a foundation of trust, and promoting technology adoption.

[0080] In one embodiment of the present invention, the system, after minor adjustments based on a small amount of data, was successfully applied to peanut and potato harvesting scenarios in a cross-crop adaptability test. Peanut tuber distribution is similar to that of tiger nuts, requiring only adjustment of the ground-penetrating radar center frequency to 120MHz to accommodate shallower burial depths; potato tubers are larger, necessitating modification of the clustering algorithm's neighborhood radius to 5cm. The online model update module, trained on two seasons of data in a ten-hectare experimental field, showed improved accuracy in predicting maturity probability, demonstrating the system architecture's versatility and its ability to be quickly extended to other underground tuber crops, thus expanding market application and enhancing the value of the technology platform.

[0081] As one embodiment of the invention, the system includes a built-in subsidy calculator that automatically estimates the amount that can be claimed based on the operating area, crop type, and regional coefficient, thereby increasing farmers' enthusiasm for using the system. Pilot programs for intelligent agricultural machinery subsidies are being conducted in cooperation with local governments. This dual drive of policy dividends and economic benefits creates a virtuous cycle, promoting the upgrading of agricultural mechanization towards intelligence.

[0082] As one embodiment of the present invention, an open application programming interface (API) is provided for use by agricultural colleges and research institutions, enabling researchers to acquire raw sensor data and intermediate model features for crop growth model research, soil-crop interaction analysis, and intelligent algorithm improvement. The system includes a built-in data annotation tool, allowing researchers to manually correct historical data to generate high-quality labeled datasets that can then be used to improve model training.

[0083] As one embodiment of the present invention, precise harvesting reduces tuber damage and omissions, thus lowering post-harvest losses, equivalent to reducing grain waste by 50 tons per 10,000 mu per year. Optimizing travel speed and digging depth reduces fuel consumption, resulting in a 200-ton reduction in carbon dioxide emissions per 10,000 mu per year. Online model updates reduce the need for manual calibration.

[0084] In one embodiment of the present invention, during disaster emergency response, when floods inundate parts of the land, the system can quickly assess the maturity of tubers in unflooded areas and prioritize harvesting from high-risk areas to reduce disaster losses. When drought causes crops to mature prematurely, the system shortens the predicted harvest time window and starts operations earlier to prevent tubers from over-ripening and falling off. The system is linked to a meteorological early warning platform to receive extreme weather forecasts and automatically adjust operational plans, such as suspending operations before heavy rains and reducing intensity during periods of high temperatures. These emergency functions enhance the resilience of agriculture against risks.

[0085] As one embodiment of the invention, a smart harvesting insurance program is launched in cooperation with agricultural insurance companies. Premiums are dynamically priced based on historical system data, with discounts offered to farmers with high harvest integrity and low damage rates. Agricultural machinery financing leases are also provided, allowing farmers to use future earnings as collateral to pay for equipment in installments, thus lowering the initial investment threshold.

[0086] As one embodiment of the present invention, the system records the experience data of veteran growers, such as traditional harvest times and methods for dealing with special weather conditions in specific plots, and encodes this data into model features or rule bases to achieve digital inheritance of experience. The system generates plot growth stories, including fertilization, weather, and harvest data over the years, forming digital agricultural history archives and enhancing farmers' emotional connection to the land.

[0087] As one embodiment of the present invention, data use follows the principle of informed consent, and farmers can view, export, and delete their personal data at any time. The algorithm is highly transparent, providing a mature probability calculation basis and path planning logic explanation to avoid black-box operations.

[0088] In one embodiment of the invention, the machine features a streamlined body to reduce wind resistance and mud adhesion. The color scheme combines agricultural green and technological gray. The interface icons are simple and intuitive, conforming to ergonomics and reducing operator fatigue. The lighting system automatically adjusts brightness during nighttime operation to avoid light pollution.

[0089] As one embodiment of the present invention, data visualization allows farmers to see the breath of the land and the growth of crops.

[0090] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based intelligent control method for a tiger nut harvester, characterized in that, include: The soil moisture content, surface temperature gradient, near-infrared reflectance spectrum of plant canopy, historical fertilization records and meteorological accumulated temperature data of the target plot are collected through the multi-dimensional sensing module of the plot environment. The data collected by the multi-dimensional perception module of the plot environment is input into a pre-trained three-dimensional convolutional long short-term memory hybrid neural network model, which outputs the maturity probability distribution map of tiger nut tubers and the optimal harvest time window in each sampling unit. During the harvester's movement, a low-frequency electromagnetic pulse is emitted through the underground tuber distribution positioning module, and the echo signal reflected by the underground tuber group is received. Combining the echo amplitude attenuation rate, phase offset, and multi-channel time difference positioning algorithm, a three-dimensional spatial coordinate point cloud map of the underground tuber is constructed in real time. Based on the maturity probability distribution map and the three-dimensional spatial coordinate point cloud map, the soil penetration depth, travel speed and lateral swing amplitude of the digging shovel are dynamically adjusted. The soil penetration depth is determined by the median of the tuber burial depth distribution plus a preset safety margin. The travel speed is adjusted inversely proportional to the tuber density per unit area and the digging efficiency function. The lateral swing amplitude is compensated and corrected based on the overlap of tuber distribution in adjacent rows. During the excavation operation, the vibration acceleration sensor array monitors the sudden change signal of the excavation resistance. When the instantaneous acceleration exceeds the preset threshold, the excavation depth fine-tuning command is triggered, causing the excavator to be raised by 5cm within 0.5s and maintained for 3s before returning to the original trajectory. After completing the first round of mining operations, the spatial coverage of the pre-harvest maturity heat map and the actual harvesting trajectory is compared. For areas with a spatial coverage of less than 90%, a secondary harvesting path is generated, and the harvester is dispatched to perform fixed-point re-harvesting along the path in half-speed mode. After each harvest season, data on actual harvest volume, tuber damage rate, unharvested residue rate, and operational energy consumption are collected and used as supervisory signals to incrementally train the three-dimensional convolutional long short-term memory hybrid neural network model, so that the model parameters can be adaptively optimized year by year according to the soil characteristics and climate patterns of the planting area.

2. The intelligent control method for a tiger nut harvester based on machine learning according to claim 1, characterized in that, The plant canopy near-infrared reflectance spectral acquisition device uses a multispectral imager with spectral channels covering the 700nm to 1000nm band, a spatial resolution of 5cm per pixel, an acquisition height of 1.2m above the top of the canopy, and a scanning speed synchronized with the harvester's travel speed to ensure that each plant unit is scanned at least three times to eliminate momentary shadow interference.

3. The intelligent control method for a tiger nut harvester based on machine learning according to claim 2, characterized in that, The input layer of the three-dimensional convolutional long short-term memory hybrid neural network model receives five-dimensional tensor data: the first dimension is the row coordinates of the land parcel, the second dimension is the column coordinates of the land parcel, the third dimension is the vertical profile depth layer, the fourth dimension is the time step, and the fifth dimension is the number of environmental parameter channels. Its hidden layer consists of three stacked 3D convolutional kernels, each with a kernel size of 3×3×3 and a stride of 1, followed by batch normalization and modified linear unit activation functions; its temporal modeling layer consists of bidirectional long short-term memory units, with a hidden state dimension of 256; the final output layer is a fully connected layer with a soft maximum function, outputting the maturity probability value of each spatial unit.

4. The intelligent control method for a tiger nut harvester based on machine learning according to claim 3, characterized in that, The ground-penetrating radar array in the underground tuber distribution positioning module consists of 12 sets of transceiver antennas, which are equidistantly arranged along the digging shovel. The center frequency is 900MHz, the pulse repetition frequency is 20KHz, the sampling interval is 0.1ns, the effective detection depth is 0.3m to 0.8m, the lateral resolution is 8cm, and the longitudinal resolution is 3cm. Its echo signal processing flow includes: DC offset removal of the original signal, median filtering of background noise, envelope extraction by Hilbert transform, time-domain threshold segmentation, peak clustering localization, and spatial interpolation to generate continuous point clouds.

5. The intelligent control method for a tiger nut harvester based on machine learning according to claim 4, characterized in that, In the aforementioned excavation depth adjustment mechanism, the safety margin is set to twice the standard deviation of the tuber burial depth distribution. When the standard deviation is less than 5cm, the safety margin is fixed at 10cm. The travel speed adjustment function is: the speed is equal to the base speed multiplied by the tuber density threshold divided by the current area tuber density, where the base speed is 3km / h and the tuber density threshold is 150 tubers per square meter. The lateral swing compensation is equal to the width of the overlapping area of ​​adjacent rows of tuber distribution multiplied by 0.8, and the maximum swing is less than 20% of the width of the excavation shovel.

6. The intelligent control method for a tiger nut harvester based on machine learning according to claim 5, characterized in that, The vibration acceleration sensor array consists of eight triaxial accelerometers, which are installed on the left and right side walls of the digging shovel at distances of 10cm, 20cm, 30cm and 40cm from the cutting edge, respectively. The sampling frequency is 5KHz, and the trigger threshold is set to 15 times the value of gravitational acceleration. The fine-tuning command is generated by the embedded programmable logic controller within 5ms after the threshold is detected to be exceeded, and the hydraulic cylinder is controlled by the proportional-integral-derivative servo driver to perform the lifting action.

7. The intelligent control method for a tiger nut harvester based on machine learning according to claim 6, characterized in that, The spatial coverage calculation method is as follows: the area with a maturity probability greater than 0.7 in the pre-harvest maturity heat map is defined as the area to be harvested. The actual harvesting trajectory is projected onto the same coordinate system, and the proportion of the intersection area of ​​the area to be harvested and the trajectory coverage area to the total area of ​​the area to be harvested is calculated. The secondary replenishment path generation algorithm adopts an improved ant colony optimization algorithm. Its pheromone update rule introduces the terrain slope penalty factor and the historical omission frequency weighting factor to ensure that the replenishment path prioritizes coverage of high omission risk areas.

8. The intelligent control method for a tiger nut harvester based on machine learning according to claim 7, characterized in that, The incremental training strategy employs an elastic weight fixation algorithm, which optimizes newly collected data using mini-batch gradient descent while retaining historical task knowledge. The initial learning rate is 0.001, decays by 5% per training round, and the number of training rounds is fixed at 50. The loss function is a weighted sum of cross-entropy loss and mean squared error loss, with weight coefficients of 0.7 and 0.3, respectively.

9. A machine learning-based intelligent control system for a tiger nut harvester, characterized in that, include: The multi-dimensional sensing module for the site environment is used to collect data on soil moisture content, surface temperature gradient, near-infrared reflectance spectrum of plant canopy, historical fertilization records, and accumulated temperature data of the target site before harvesting. The tuber maturity spatiotemporal prediction module is used to input the data collected by the multi-dimensional perception module of the plot environment into a pre-trained three-dimensional convolutional long short-term memory hybrid neural network model, and output the maturity probability distribution map of tiger pea tubers and the optimal harvest time window in each sampling unit. The underground tuber distribution positioning module is used to transmit low-frequency electromagnetic pulses through a ground-penetrating radar array installed at the front end of the digging shovel during the harvester's movement, receive echo signals reflected by the underground tuber group, and combine echo amplitude attenuation rate, phase offset and multi-channel time difference positioning algorithm to construct a three-dimensional spatial coordinate point cloud map of the underground tubers in real time. An adaptive excavation control module is used to dynamically adjust the soil penetration depth, travel speed, and lateral swing amplitude of the excavation shovel based on the maturity probability distribution map output by the tuber maturity spatiotemporal prediction module and the three-dimensional spatial coordinate point cloud map output by the underground tuber distribution positioning module. The soil penetration depth is determined based on the median of the tuber burial depth distribution plus a preset safety margin. The travel speed is adjusted inversely proportional to the tuber density per unit area and the excavation efficiency function. The lateral swing amplitude is compensated and corrected based on the overlap of adjacent rows of tuber distribution. The damage avoidance feedback module is used to monitor the sudden change signal of digging resistance during the digging operation by using a vibration acceleration sensor array installed on the side wall of the digging shovel. When the instantaneous acceleration exceeds the preset threshold, the digging depth fine adjustment command is immediately triggered, so that the digging shovel is raised by 5cm within 0.5s and maintained for 3s before returning to the original trajectory. The path planning and missed mining module is used to compare the spatial coverage of the pre-harvest maturity heat map with the actual harvesting trajectory after the first round of mining operations. It generates a secondary mining path for areas with a coverage of less than 90% and schedules the harvester to perform fixed-point re-harvesting along the path in half-speed mode. The online model update module is used to collect data on actual harvest volume, tuber damage rate, unharvested residue rate, and operational energy consumption after each harvest season. This data is used as a supervisory signal to incrementally train the three-dimensional convolutional long short-term memory hybrid neural network model, enabling the model parameters to adaptively optimize as the soil characteristics and climate patterns of the planting area evolve year by year.

10. The intelligent control system for a tiger nut harvester based on machine learning according to claim 9, characterized in that, The near-infrared reflectance spectral acquisition device for the plant canopy in the multi-dimensional perception module of the plot environment adopts a multispectral imager with a spectral channel covering the 700nm to 1000nm band, a spatial resolution of 5cm per pixel, an acquisition height of 1.2m above the top of the canopy, and a scanning speed synchronized with the speed of the harvester to ensure that each plant unit is scanned at least three times to eliminate momentary shadow interference.

Images