Machine learning based subsoiling operation depth prediction method and system

Abstract

The application relates to the field of intelligent agricultural technology and provides a deep ploughing operation ploughing depth prediction method and system based on machine learning. The deep ploughing operation ploughing depth prediction method comprises the following steps: collecting multi-dimensional working condition data in a deep ploughing operation implement operation process through a CAN bus, performing a pretreatment operation on the multi-dimensional working condition data, and obtaining a multi-dimensional feature vector; based on the multi-dimensional feature vector, a current working condition state vector of the deep ploughing operation implement is constructed, the weight of each dimension feature vector is determined based on the working condition state vector, each dimension feature vector is adjusted through the weight of each dimension feature vector, and an adaptive input feature vector is obtained; the adaptive input feature vector is input into a ploughing depth prediction model for processing, and a deep ploughing operation ploughing depth prediction result is obtained. The deep ploughing operation ploughing depth prediction method and system based on machine learning can improve the accuracy and stability of ploughing depth measurement in the deep ploughing operation process and reduce the installation and debugging complexity.

Description

Technical Field

[0001] This application relates to the field of smart agriculture technology, and in particular to a method and system for predicting tillage depth in deep tillage operations based on machine learning. Background Technology

[0002] Deep tillage is a crucial step in modern conservation tillage and soil improvement practices. Its main functions are to break up the plow pan, improve soil permeability, enhance soil water retention capacity, and provide a more suitable soil environment for crop root growth. Tillage depth, as one of the core parameters for evaluating the quality of deep tillage, directly relates to whether the operation meets predetermined agronomic requirements and has a significant impact on soil structure remodeling, water and fertilizer transport efficiency, and subsequent crop growth. Therefore, accurately obtaining tillage depth parameters during deep tillage is of great significance for monitoring agricultural machinery operation quality, intelligent control, and the development of precision agriculture.

[0003] Currently, methods for measuring tillage depth mainly rely on external sensors or additional measuring mechanisms. These methods suffer from poor accuracy and stability during actual deep tillage operations, and are also complex to install and debug. Summary of the Invention

[0004] This application provides a machine learning-based method and system for predicting tillage depth in deep tillage operations, in order to solve the problems of poor accuracy and stability of related technologies in actual deep tillage operations, as well as the complexity of installation and debugging.

[0005] This application provides a machine learning-based method for predicting tillage depth in deep tillage operations, including: Multi-dimensional working condition data of agricultural implements during deep tillage operations are collected via the CAN bus of the controller area network, and preprocessing operations are performed on the multi-dimensional working condition data to obtain multi-dimensional feature vectors. Based on multi-dimensional feature vectors, a current working state vector of deep tillage tools is constructed, and the weights of each dimension feature vector are determined based on the working state vector. The feature vectors of each dimension are then adjusted by the weights of the feature vectors of each dimension to obtain an adaptive input feature vector. The adaptive input feature vector is fed into the tillage depth prediction model for processing to obtain the tillage depth prediction results for deep tillage operations.

[0006] In one embodiment, the multi-dimensional operating condition data includes wheel speed, horizontal traction force, rear suspension position, engine speed, engine torque-load ratio, and instantaneous fuel consumption.

[0007] In one embodiment, the operating condition state vector includes at least one of the following: The working environment state component is used to characterize soil resistance. The power load state component is used to characterize the engine load level. The slip state component is used to characterize the adhesion state of the drive wheel. The geometric height component is used to characterize the spatial geometric reference of the rear suspension of deep tillage implements; The operating speed component is used to characterize the transient speed regulation state of the power system. The energy consumption variation component is used to characterize the dynamic load.

[0008] In one embodiment, the tillage depth prediction model is obtained by training a random forest model. The training process of the random forest model includes tuning the hyperparameters of the random forest model using a Bayesian optimization method. The hyperparameters include the number of decision trees, the maximum tree depth, the minimum number of split samples per node, and the minimum number of leaf node samples.

[0009] In one embodiment, the preprocessing operation includes at least one of the following: Normalize the working condition data for each dimension; Outliers in multi-dimensional working condition data are replaced by the middle value of the sliding window. Outliers are feature values ​​whose deviation from the middle value of the sliding window is greater than a preset threshold. Linear interpolation is performed on missing items in multi-dimensional working condition data; The operating condition data for each dimension is filtered and smoothed.

[0010] This application also provides a machine learning-based deep tillage depth prediction system, including: The data acquisition controller is connected to the CAN bus network of the deep tillage implement via the CAN diagnostic connector of the implement's controller area network. It is used to collect multi-dimensional working condition data of the deep tillage implement during operation. The host computer is used to perform preprocessing operations on the multi-dimensional working condition data collected by the data acquisition controller to obtain multi-dimensional feature vectors. Based on the multi-dimensional feature vectors, it constructs the current working condition state vector of the deep tillage operation implements, determines the weight of each dimension feature vector based on the working condition state vector, and adjusts each dimension feature vector through the weight of each dimension feature vector to obtain an adaptive input feature vector. The adaptive input feature vector is then input into the tillage depth prediction model for processing to obtain the tillage depth prediction result of the deep tillage operation. The host computer is also used to display the predicted tillage depth results of deep tillage operations.

[0011] This application also provides a machine learning-based deep tillage depth prediction device, including: The acquisition module is used to collect multi-dimensional working condition data of agricultural implements during deep tillage operations via the CAN bus of the controller area network; The processing module is used to perform preprocessing operations on multi-dimensional working condition data to obtain multi-dimensional feature vectors. The processing module is also used to construct the current working state vector of the deep tillage operation tools based on the multi-dimensional feature vector, determine the weight of each dimension feature vector based on the working state vector, and adjust each dimension feature vector through the weight of each dimension feature vector to obtain an adaptive input feature vector. The processing module is also used to input the adaptive input feature vector into the tillage depth prediction model for processing, so as to obtain the tillage depth prediction result of deep tillage operation.

[0012] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the machine learning-based deep tillage depth prediction method as described above.

[0013] This application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the machine learning-based deep tillage depth prediction method as described above.

[0014] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the machine learning-based deep tillage depth prediction method as described above.

[0015] The machine learning-based method and system for predicting tillage depth in deep tillage operations provided in this application can directly extract multi-dimensional working condition data from the CAN bus of deep tillage implements and perform data preprocessing, avoiding reliance on external physical sensors and additional measuring mechanisms, thus reducing the hardware complexity, installation difficulty, and subsequent maintenance costs of agricultural machinery. Furthermore, by constructing a working condition state vector and dynamically adjusting the weights of each feature dimension online, adaptive dynamic modulation of feature importance can be achieved, effectively overcoming the technical defects of traditional geometric calibration relationships that are prone to drift due to sudden changes in soil properties, physical tire sinking, or severe slippage of drive wheels. Simultaneously, combined with a tillage depth prediction model, it can improve the accuracy, operational stability, and anti-interference robustness of deep tillage operations in complex and variable farmland environments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a machine learning-based method for predicting tillage depth in deep tillage operations, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a model training process provided in an embodiment of this application.

[0019] Figure 3 This is a schematic diagram of a parameter tuning process provided in an embodiment of this application.

[0020] Figure 4 This is a schematic diagram of the structure of a machine learning-based deep tillage depth prediction system provided in an embodiment of this application.

[0021] Figure 5 This is a schematic diagram of the architecture of a machine learning-based deep tillage depth prediction system provided in an embodiment of this application.

[0022] Figure 6 This is a schematic diagram of a machine learning-based deep tillage depth prediction device provided in an embodiment of this application.

[0023] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] To facilitate understanding of the machine learning-based deep tillage depth prediction method and system provided in the embodiments of this application, the following is a brief introduction.

[0026] Currently, most methods for measuring tillage depth rely on external sensors or additional measuring mechanisms. While these methods can achieve tillage depth measurement to a certain extent or under certain conditions, they have poor characterization capabilities for working conditions, poor accuracy and stability during actual deep tillage operations, and are complex to install and debug.

[0027] Specifically, these methods are highly dependent on external sensors and additional measuring mechanisms. For example, they typically rely on external devices such as tilt sensors, displacement sensors, ultrasonic sensors, and contouring mechanisms, which are additionally deployed on tractors or implements. This not only increases the complexity of the system but also places high demands on the sensor installation location, installation accuracy, and on-site deployment conditions.

[0028] Furthermore, these methods are sensitive to initial calibration and structural parameters. For example, measurement methods that calculate tillage depth based on geometric relationships or attitude typically rely on initial installation dimensions, structural parameters, and calibration relationships to establish a tillage depth estimation model. However, the relevant calibrations are mostly based on static or ideal conditions. When the working conditions change, the original calibration relationships shift, which can easily lead to increased measurement errors.

[0029] Furthermore, these methods are insufficiently adaptable to actual working conditions. During tractor tillage, the working environment will change dynamically, such as soil differences, operating speed, and operating depth. These factors will cause changes in the unit's load and vibration. Relying solely on sensor measurements or geometric calculations to estimate tillage depth is insufficient to fully characterize the comprehensive impact of actual working conditions on tillage depth, resulting in insufficient adaptability to actual working environments and making it difficult to meet the needs of continuous operation in real field scenarios.

[0030] Considering the continuous improvement of the intelligence level of agricultural equipment, tractors already possess relatively complete electronic control systems and on-board bus communication capabilities, enabling them to output a large amount of operational information related to their working status in real time. Therefore, it is necessary to fully utilize the operational data of the tractor's Electronic Control Unit (ECU) and establish a mapping relationship between tillage depth and operational information using artificial intelligence methods, thereby achieving tillage depth measurement without the need for additional equipment.

[0031] Based on this, this application provides a machine learning-based method and system for predicting tillage depth in deep tillage operations. It directly extracts multi-dimensional operational data from the Controller Area Network (CAN) bus of the deep tillage implements and performs data preprocessing, avoiding reliance on external physical sensors and additional measuring mechanisms, thus reducing the hardware complexity, installation difficulty, and subsequent maintenance costs of agricultural machinery. Furthermore, by constructing an operational state vector and dynamically adjusting the weights of each feature dimension online, it achieves adaptive dynamic modulation of feature importance, effectively overcoming the technical shortcomings of traditional geometric calibration relationships that are prone to drift due to sudden changes in soil properties, tire sinking, or severe slippage of the drive wheels. Simultaneously, combined with a tillage depth prediction model, it can improve the accuracy, operational stability, and robustness against interference in deep tillage operations under complex and variable farmland environments.

[0032] The following is combined Figures 1-5 This application describes the method and system for predicting tillage depth in deep tillage operations.

[0033] Figure 1 This is a flowchart illustrating the machine learning-based method for predicting tillage depth in deep tillage operations provided in an embodiment of this application. Figure 1 As shown, the method includes the following: Step 110: Collect multi-dimensional working condition data of agricultural implements during deep tillage operations via CAN bus, and perform preprocessing operations on the multi-dimensional working condition data to obtain multi-dimensional feature vectors.

[0034] The CAN bus refers to the CAN bus network on subsoiling implements. Subsoiling implements are agricultural tools used for subsoiling operations in farmland, such as tractors.

[0035] In one embodiment, the multi-dimensional operating condition data includes wheel speed, horizontal traction force, rear suspension position, engine speed, engine torque-load ratio, and instantaneous fuel consumption.

[0036] The rear suspension position directly determines the relative height of the implement in space, and is a major geometric parameter affecting tillage depth, directly reflecting the overall trend of tillage depth. Horizontal traction reflects the force between the implement and the soil; as soil resistance increases, the horizontal traction increases accordingly, showing a clear positive correlation. Wheel speed reflects the machine's driving state; its changes affect implement stability and soil disturbance, indirectly affecting tillage depth stability. Engine speed and engine torque-load ratio characterize the power system's output state; when tillage depth or soil resistance increases, the engine load relatively increases, and related parameters change accordingly. Instantaneous fuel consumption reflects the engine's output level; its trend is related to load conditions, and to some extent indirectly characterizes tillage depth changes.

[0037] As an example, considering that the data in the CAN bus communication process of deep tillage tools is transmitted in the form of messages, a working condition data parsing program can be designed to identify and parse the data according to relevant industry protocols and communication standards to obtain multi-dimensional feature vectors.

[0038] For example, the multi-dimensional feature vectors obtained from the receiving parsing program can be shown in Table 1 below.

[0039] Table 1 For example, the receiving and parsing procedure can be as follows: che_su = a1 / 256 + a2; fdj_zhuan_su=(b4*256+b3) / 8; fu_zai_lv=c2; qian_yin_li=(d0+d1*256)*10-320000; xg_wei_zhi=e0*0.4; ran_you_lv=(r1*256+r0)*0.05.

[0040] Specifically, "che_su=a1 / 256+a2" is used to proportionally combine bytes a1 and a2 to obtain wheel speed. "fdj_zhuan_su=(b4*256+b3) / 8" is used to proportionally convert bytes b3 and b4 to engine speed. "fu_zai_lv=c2" is used to directly read the signal from byte c2 to obtain the engine torque-load ratio. "qian_yin_li=(d0+d1*256)*10-320000" is used to obtain horizontal traction force by offsetting and scaling the combined bytes d0 and d1. "xg_wei_zhi=e0*0.4" is used to multiply the signal from the single byte e0 by a proportional coefficient to obtain the rear suspension position. "ran_you_lv=(r1*256+r0)*0.05" is used to multiply the combined bytes r0 and r1 by a proportional coefficient to obtain instantaneous fuel consumption.

[0041] In one embodiment, in order to reduce abnormal pulses, missing data and high-frequency noise in the operating condition data, preprocessing operations can be performed on the multi-dimensional operating condition data to identify and remove outliers, complete missing values ​​and suppress noise, thereby improving the overall quality of the original data and ensuring the high quality and continuity of the input feature vector.

[0042] As an example, the CAN diagnostic interface on the deep tillage implement can be connected to the vehicle's CAN bus network to extract multi-dimensional working condition data according to a preset communication protocol, forming a raw input data set: .

[0043] in, For the first The speed of the wheels at all times. For the first Horizontal traction at all times. For the first The position after the moment of suspension. For the first Constant engine speed. For the first Engine torque-load ratio at any given time. For the first Real-time fuel consumption.

[0044] Furthermore, to ensure the time consistency of operating condition data, a unified time benchmark can be introduced to synchronously reconstruct the operating condition data and construct a standardized time-series input sequence: .

[0045] in, This is the sampling length. This is the original input data set for time 1 (i.e., multi-dimensional operating condition data). This is the original input data set at time 2. Let N be the set of original input data at time N.

[0046] Furthermore, to eliminate the influence of different units on the model, preprocessing operations can include normalization, which involves normalizing each input feature to obtain standardized input features. .

[0047] in, For the first The first sample (or the first) The first of the original input data sets at time (time) The eigenvalues ​​of each feature. For the normalized first The first sample The eigenvalues ​​of each feature. For the first The minimum value of a feature in all samples. For the first The maximum value of a feature across all samples.

[0048] Based on this, the final input vector is: .

[0049] in, Normalized The set of raw input data at each moment. For the normalized first The speed of the wheels at all times. For the normalized first Horizontal traction at all times. For the normalized first The position after the moment of suspension. For the normalized first Constant engine speed. For the normalized first Engine torque-load ratio at any given time. For the normalized first Real-time fuel consumption.

[0050] In this way, by directly calling the internal operating condition data of the tractor ECU, a unified time series and unified scale input feature vector can be constructed, realizing the foundation for expressing tillage depth information without the need for external measurement sensors, and providing a stable input for subsequent tillage depth prediction models.

[0051] Step 120: Based on the multi-dimensional feature vector, construct the current working condition state vector of the deep tillage operation tool, determine the weight of each dimension feature vector based on the working condition state vector, and adjust each dimension feature vector through the weight of each dimension feature vector to obtain the adaptive input feature vector.

[0052] In one embodiment, in order to achieve higher prediction accuracy of the tillage depth prediction model under different deep tillage operation conditions, a feature dynamic weighting method based on working condition perception can be used to adjust the feature vectors of each dimension during the input stage of the tillage depth prediction model, thereby dynamically adjusting the influence strength of the feature vectors of each dimension under different operation conditions.

[0053] Specifically, a multi-dimensional feature vector can be constructed to extract key parameters that reflect changes in the operational status as descriptive quantities for the current working environment, load level, and operating status. For example, the multi-dimensional feature vector and the operational status vector can be: ; .

[0054] in, It is a multi-dimensional feature vector. This is the operating condition state vector.

[0055] In one embodiment, some dimensional feature vectors can be selected from the multi-dimensional feature vectors as components of the working condition state vector, and / or, multiple different dimensional feature vectors in the multi-dimensional feature vectors can be nonlinearly composited to obtain the components of the working condition state vector.

[0056] The operating condition state vector may include at least one of the following: slip state component, operating environment state component, geometric height component, operating speed component, power load state component, and energy consumption change component.

[0057] The slip component characterizes the adhesion state of the drive wheels. In actual operation, the slip component is obtained by nonlinearly combining the wheel speed and engine speed features from a multi-dimensional feature vector. When a tractor experiences severe slippage while operating in wet, sticky, or high-resistance soil, the wheel speed rapidly decreases while the engine speed remains high. At this point, the value of the slip component instantaneously decreases. By observing the change in this component, the adhesion failure state of the drive wheels can be identified in real time.

[0058] The working environment state component is used to characterize soil resistance and can be obtained by nonlinearly combining the horizontal traction force characteristics and the rear suspension position characteristics. Since the traction force is not only related to soil hardness but also positively correlated with the working depth (i.e., the rear suspension position), when the rear suspension position does not change depth but the resistance suddenly increases (such as when driving into hard soil or cutting into the plow pan), the working environment state component will increase significantly, thus accurately sensing changes in the external soil environment.

[0059] The geometric height component is used to characterize the spatial geometric reference of the rear suspension of subsoiling implements and can be directly determined by the positional characteristics of the rear suspension. The geometric height component directly reflects the relative vertical lifting height of the rear suspension mechanism of the subsoiling implement and is the dominant spatial geometric factor determining the macroscopic variation trend of tillage depth.

[0060] The operating speed component is used to characterize the transient speed regulation state of the powertrain and can be determined by directly selecting engine speed characteristics. When encountering sudden resistance, the engine will experience a speed drop due to overload. The operating speed component can sensitively capture the speed feedback of the powertrain in response to sudden external resistance.

[0061] The power load state component is used to characterize the engine load level and can be determined by directly selecting the engine torque-load ratio characteristic. It can directly reflect the percentage of the engine's current output torque to its limit torque. When the tillage depth increases or the soil hardens, causing the tractor's pulling load to increase, the power load state component will increase proportionally and significantly, intuitively reflecting the overall machine power load situation.

[0062] The energy consumption variation component is used to characterize the dynamic load and can be determined by directly selecting the instantaneous fuel consumption characteristic. In the engine injection control loop, instantaneous fuel consumption directly reflects the opening of the injector, and its variation trend has a highly nonlinear correlation with torque load and resistance load. It can help indicate the overall engine resistance fluctuation state from the perspective of energy consumption flow.

[0063] Furthermore, the operating condition vector can be mapped to a feature weight vector to obtain the weights of the feature vectors in each dimension, enabling dynamic adjustment of feature importance under different operating conditions. For example, weight generation can use a normalized mapping method, the expression of which is: .

[0064] in, For the first The dynamic weights of each feature satisfy... and , That is, the sum of the weights of the feature vectors of each dimension. The sum of the weighted activation values ​​of all p feature channels is an exponential sum of the natural constant e, where p is the total number of dimensions of the multi-dimensional feature vector (e.g., 6), and l is a loop index variable used to iterate and accumulate the sum of each feature channel. For the first The weighted activation values ​​of each feature are obtained by linear mapping from the working condition state vector: .

[0065] in, For the first The mapping parameter vector corresponding to each feature. For the first Each feature corresponds to a bias term. After obtaining the feature weights, the multi-dimensional feature vector can be element-wise weighted to obtain an adaptive input feature vector: .

[0066] in, For the weighted number of The eigenvalues ​​of each feature are the adaptive input feature vectors. The final e-weighted feature vector (i.e., the adaptive input feature vector) is: .

[0067] Based on this, the model input can be transformed from fixed features into a feature expression that is dynamically adjusted according to changes in working conditions. Under conditions such as fluctuations in soil, working speed, and load, the stability and accuracy of tillage depth prediction can be effectively improved.

[0068] Step 130: Input the adaptive input feature vector into the tillage depth prediction model for processing to obtain the tillage depth prediction result for deep tillage operation.

[0069] In one embodiment, the tillage depth prediction model can be obtained by training a random forest model. The training process of the random forest model includes tuning the hyperparameters of the random forest model using a Bayesian optimization method. The hyperparameters include the number of decision trees, maximum tree depth, minimum number of split samples per node, and minimum number of leaf node samples.

[0070] Figure 2 This is a schematic diagram illustrating a model training process provided in an embodiment of this application. Figure 2 As shown, the random forest model achieves a stable output of the final tillage depth prediction value by integrating and averaging the prediction results of multiple decision regression trees.

[0071] Specifically, the modeling and computation process of a random forest regression model may include data source input, bootstrap resampling, out-of-bag data validation, independent decision tree modeling, regression prediction output, and result ensemble averaging.

[0072] During the data source input phase, the model's input source is the standard time-series feature data generated after the aforementioned preprocessing, cleaning, and dynamic weighting modulation, which is used to construct the training set. The training set can be represented as: .

[0073] in, This indicates that the i-th sample in the model is an adaptive input feature vector containing 6 working condition features. Indicates and The corresponding actual measured value of tillage depth. N is the total number of training samples. If the random forest model consists of T regression trees, the tillage depth result of the t-th tree is... The final tillage depth value y is obtained by averaging.

[0074] During the Bootstrap resampling phase, the model randomly samples T subsets of the same size from the training set using random sampling with replacement, denoted as Training Subset 1, Training Subset 2, ..., Training Subset n (where n corresponds to the total number of regression decision trees T). In constructing each training subset, samples not selected from the training set are automatically assigned to their corresponding out-of-bag subsets 1, 2, ..., n.

[0075] During the out-of-bag data validation phase, internal cross-validation can be performed using out-of-bag samples corresponding to each decision tree to evaluate the generalization error of each decision tree in real time and ensure the output accuracy of the overall model.

[0076] In the independent decision tree modeling stage, decision trees 1, 2, ..., n can be constructed independently and in parallel based on each training subset. During the growth and splitting process of each decision tree, not only are the samples randomly selected, but the feature subsets of the splitting nodes are also randomly selected, thereby minimizing the correlation between the trees and enabling each decision tree to gradually learn the complex nonlinear correspondence between working condition features and the actual tillage depth.

[0077] In the regression prediction output stage, after the adaptive input feature vector is input into the trained model, each decision tree computes independently in parallel and outputs its corresponding prediction result, namely, the result of decision tree 1, the result of decision tree 2, ..., the result of decision tree n. For the t-th decision tree, its tillage depth prediction result is denoted as... .

[0078] In the results integration and averaging stage, the prediction results of all decision trees can be aggregated. By performing an averaging calculation, the final output of the random forest regression model (i.e., the final tillage depth prediction result) is merged and output. : .

[0079] Figure 3 This diagram illustrates a parameter tuning process provided in an embodiment of this application. To further improve the prediction accuracy and robustness of the model under complex, alternating farmland conditions, key model configuration parameters can be tuned online using Bayesian optimization methods. The hyperparameters to be tuned specifically include the number of decision trees, maximum tree depth, minimum number of split samples per node, and minimum number of leaf node samples. The specific execution steps are as follows: Step 1: Initialize the parameter space and samples.

[0080] Specifically, after starting the optimization process, the search space X of the hyperparameters to be optimized can be initialized. Let the objective function to be optimized be f(x), where x represents the vector of hyperparameter combinations to be optimized (including the number of decision trees, maximum tree depth, minimum number of split samples per node, and minimum number of leaf node samples), and d is the dimension of the hyperparameters (e.g., d=4). Then, the problem of finding the optimal parameter combination can be transformed into the following mathematical problem of minimizing the objective function: .

[0081] Where X is the parameter search space. This represents the optimal combination of hyperparameters.

[0082] Step 2: Construct a Gaussian process proxy model.

[0083] Specifically, considering that the true evaluation objective function f(x) is extremely complex and cannot be directly expressed analytically, a Gaussian process can be used to approximate it. Assume the objective function follows a Gaussian process distribution: .

[0084] in, It is a mean function. Let be the covariance function. GP represents the Gaussian process distribution.

[0085] Let the obtained n observation samples be: .

[0086] in, Let be the hyperparameter combination for the i-th sampling. For the corresponding objective function value, Bayesian inference can be used to obtain the mean and variance of the posterior distribution of the objective function at any new parameter coordinate x: in, This is the covariance vector between the new sampling point and each observation point. , is a kernel-symmetric matrix constructed based on historical samples. , which is the column vector of observations. This represents the mean output of the current Gaussian process prediction, used to reflect the model's preference for prediction accuracy with the current parameter combination. The output variance is used to reflect the uncertainty of the model in the current parameter region.

[0087] Step 3: Calculate the acquisition function EI.

[0088] Specifically, instead of directly utilizing the current optimal point, the posterior distribution output by the Gaussian process can be used. Instead, a data acquisition function is introduced to balance development and exploration. In this embodiment, the data acquisition function can be Expected Improvement (EI), which is defined as: .

[0089] By integrating it, its analytical form can be obtained: .

[0090] in, This represents the currently observed optimal objective function value (i.e., the lowest verification error). This is the cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution. , which are standardized variables. This data collection function comprehensively considers both the output mean and uncertainty. When Smaller or When the value is large, the EI value is high, achieving a balanced relationship.

[0091] Step 4: Maximize EI to select the next sampling point.

[0092] Specifically, a global search algorithm can be used to find the function that improves the current desired performance. The coordinates of the hyperparameter that reaches its maximum value are used as the position for the next iteration sampling: .

[0093] Step 5: Calculate performance metrics for model training.

[0094] Specifically, at the selected new hyperparameter point In this case, the random forest model is actually configured and retrained, and then run on the test and validation set to calculate its tillage depth estimation error (such as mean square error), which serves as the performance index of the true objective function for evaluating the merits of this parameter combination.

[0095] Step 6: Update the sample dataset.

[0096] Specifically, the newly sampled hyperparameters can be combined The corresponding performance evaluation results are added to the historical observation set, thereby updating the sample set. .For example, .

[0097] Step 7: Determine whether the termination condition is met.

[0098] Specifically, the termination condition can be reaching a preset maximum number of iterations, or the change between two adjacent optimal points being less than a preset convergence threshold.

[0099] If the determination result is negative, then return to the step of constructing the Gaussian process surrogate model along the update feedback branch, based on the updated sample set. Reconstruct the Gaussian process and begin a new round of optimization iterations.

[0100] If the determination result is yes, then stop the iteration, proceed to the step of outputting the optimal parameters, extract the hyperparameter combination corresponding to the historical best target value as the final model parameter configuration, and end.

[0101] Based on this, the optimal hyperparameter configuration for the tillage depth prediction model can be determined by performing the Bayesian optimization process described above. For example, the number of decision trees is 396, the maximum tree depth is 17 layers, the minimum number of leaf node samples is 2, and the minimum number of split samples is 4. This allows the trained random forest model to exhibit good fitting performance and anti-overfitting ability in the adaptive feature space, achieving optimal output results.

[0102] In one embodiment, the preprocessing operation in step 110 above includes at least one of the following: normalizing the operating condition data for each dimension, replacing outliers in the multi-dimensional operating condition data with the median value in the sliding window, performing linear interpolation on missing items in the multi-dimensional operating condition data, and filtering and smoothing the operating condition data for each dimension. Outliers are feature values ​​whose deviation from the median value in the sliding window is greater than a preset threshold.

[0103] As an example, outlier identification and adaptive replacement can be performed on each channel signal at each sampling time. Specifically, a local sliding window with a time span of W can be constructed, where the current sampling time is the end time of the local sliding window, and the window width W can be set according to the sampling frequency (e.g., 5 or 7). Then, the median of all sampled data within the local sliding window can be calculated, and the median of the sliding window can be used as a reference value for the baseline physical level within the sliding window. Simultaneously, the Median Absolute Deviation (MAD) of the data within this sliding window can be calculated to characterize the self-tuning dispersion of the current signal segment.

[0104] Therefore, when the operating data meets the following judgment formula, it can be judged as an outlier: .

[0105] in, For operating condition data at time The sampled values. This represents the middle value in the sliding window corresponding to the operating condition data. The degree of dispersion of the operating condition data can be calculated based on the absolute deviation of the middle value in the sliding window corresponding to the operating condition data. This is the threshold coefficient.

[0106] If any dimension of the operating data is determined to be an outlier, the median of that dimension in the sliding window can be automatically used to replace the operating data of that dimension, thereby completely eliminating high-frequency pulse and distortion spike interference while preserving the macroscopic timing characteristics of the physical signal.

[0107] As an example, linear interpolation can be performed on missing items in multi-dimensional operating condition data collected at any given time to ensure the continuous evolution of multi-dimensional operating condition data over time.

[0108] Specifically, the nearest previous valid sampling point to the missing time t can be extracted. With the next valid sampling point Linear interpolation is performed using time-proportional correlation, and the calculation formula is as follows: .

[0109] in, and These are the valid data at both ends of the missing time point. This is a time to be completed.

[0110] Based on this, by performing linear interpolation calculations on all missing moments, the originally discrete and discontinuous operating condition sequence can be transformed into a complete operating condition data stream with complete time sequence and high continuity.

[0111] As an example, during actual deep tillage operations, the engines of deep tillage implements may vibrate violently, and the resistance of the tires interacting with heterogeneous soil may cause vibrations. These factors introduce continuous high-frequency fluctuation noise into the bus data. Therefore, dynamic Kalman recursive filtering smoothing can be performed independently on the operating data for each dimension.

[0112] Specifically, at each sampling time, the estimated value from the previous time step can be optimally fused with the actual observation value at the current time step, and the recursive update state estimation equation can be expressed as: .

[0113] in, This is the filtered output at the current moment. For the current observation value, For filtering efficiency.

[0114] Let k be the estimated value after smoothing by Kalman filtering. It is the estimated value of time k-1 (i.e., the time before time k) after Kalman filtering smoothing. Let k be the actual observed value at time k. This is the Kalman filter gain at the current iteration step. Its magnitude is dynamically calculated and updated during the recursive process by the preset process state noise covariance, measurement noise covariance, and estimation error covariance of the previous time step.

[0115] Based on this, by performing preprocessing operations on multi-dimensional operating condition data, the original operating condition data can be transformed from abnormal, discontinuous, and fluctuating states into a continuous, smooth, and stable standard data sequence, providing high-quality input for the model.

[0116] This application also provides a machine learning-based deep tillage depth prediction system, including a data acquisition controller and a host computer.

[0117] The data acquisition controller is connected to the CAN bus network of the subsoil tillage implement via the CAN diagnostic connector, and is used to collect multi-dimensional working condition data of the subsoil tillage implement during operation.

[0118] The host computer is used to perform preprocessing operations on the multi-dimensional working condition data collected by the data acquisition controller to obtain multi-dimensional feature vectors. Based on the multi-dimensional feature vectors, a current working condition state vector of the deep tillage operation tools is constructed. The weights of each dimension feature vector are determined based on the working condition state vector, and the feature vectors of each dimension are adjusted by the weights of each dimension feature vector to obtain an adaptive input feature vector. The adaptive input feature vector is then input into the tillage depth prediction model for processing to obtain the tillage depth prediction result of the deep tillage operation.

[0119] The host computer is also used to display the predicted tillage depth results of deep tillage operations.

[0120] As an example, Figure 4 This is a schematic diagram of the structure of a machine learning-based deep tillage depth prediction system provided in an embodiment of this application.

[0121] The data acquisition controller, acting as both an acquisition controller and a signal receiving device, may include an acquisition control host and a CAN transceiver module. The acquisition control host can be a controller from the NI-C DAQ-913 series (such as the NI-C DAQ-9135). The CAN transceiver module can be an NI-9862 module.

[0122] The CAN transceiver module connects to the CAN bus network of the subsoil tillage implement via its CAN diagnostic connector. It receives and parses data extracted from the implement's CAN bus network, enabling online acquisition of ECU data. Furthermore, the acquisition and control host can centrally manage multi-dimensional operating condition data. This multi-dimensional data can include wheel speed, horizontal traction force, rear suspension position, engine speed, engine torque-load ratio, and instantaneous fuel consumption.

[0123] During data acquisition, the data acquisition controller can establish a global time reference based on a unified clock, and collect and schedule different data according to a set sampling frequency to form a data sequence with consistent timestamps. Compared with the traditional multi-device independent acquisition method, the embodiment of this application can achieve advantages such as clear data acquisition path, high system integration and strong scalability through centralized access via CAN bus and modular acquisition architecture. Furthermore, the unified sampling frequency and time synchronization mechanism can effectively avoid the problem of time misalignment of multiple data, providing high-quality model input data.

[0124] The data acquisition controller supports external device access, such as a host computer and a 12V vehicle power supply. The host computer supports data processing, model operation, and human-machine interface. The 12V vehicle power supply is used for power supply.

[0125] As an example, to meet the real-time monitoring needs of operational status information and tillage depth, a host computer human-machine interface can be developed using graphical programming software (such as the LabVIEW platform) to integrate functions such as parameter configuration, real-time data display, and storage. Furthermore, the interactive platform can also support saving the collected data in a timestamped table format.

[0126] The host computer human-machine interface serves as the core for data acquisition and scheduling. After completing the CAN data parsing and processing, it can encapsulate the current working condition data and pass it to the Python side by calling the Python script interface. After the Python program receives the input data, it calls the trained tillage depth prediction model for real-time calculation and returns the output tillage depth prediction result to the host computer human-machine interface. Finally, it is displayed in real time on the human-machine interface, and the overall latency can be controlled within the sampling period, meeting the real-time monitoring requirements of deep tillage operations.

[0127] As an example, Figure 5This is a schematic diagram of the architecture of a machine learning-based deep tillage depth prediction system provided in an embodiment of this application. The system is divided into an input end, an execution end, and an output end in terms of space and data link. The input end is built upon the tillage unit. During deep tillage operations, the system collects multi-dimensional operating condition data reflecting the machine's movement and load status in real time via the vehicle's CAN network. Specifically, this includes horizontal traction force, rear suspension position, instantaneous fuel consumption, torque-load ratio, wheel speed, and engine speed. These six types of operating condition data converge at the bus physical layer to the CAN diagnostic connector and are sent to the execution end in real time, achieving lossless acquisition of basic operating condition data without the need for external physical sensors.

[0128] The execution end is deployed on the host computer, and its core operating platform is the tillage depth measurement model (i.e., tillage depth prediction model). This model deeply integrates four sub-algorithm modules: a dynamic feature weighting mechanism, a random forest regression model, a sample distribution optimization strategy, and Bayesian parameter optimization. The dynamic feature weighting mechanism dynamically assigns weights online based on input features to reshape the feature space. The random forest regression model acts as the prediction agent, performing ensemble computation of parallel trees. The sample distribution optimization strategy is responsible for balancing heterogeneous datasets during the training phase. Bayesian parameter optimization uses a surrogate model to globally optimize and determine the optimal hyperparameters, which are then embedded into the regression model, collectively ensuring the algorithm's superior adaptive computational capability in complex and variable soil environments.

[0129] The real-time tillage depth prediction results obtained from the execution end can be transmitted to the output end. The output end includes two branches in parallel: a human-machine interface and a tillage depth output branch. The human-machine interface can display the current tillage depth value, historical trend curves, and various collected working condition characteristic parameters at a high frequency of milliseconds, and supports automatic storage of historical data with timestamps. The tillage depth output branch directly outputs the high-precision calculation results as physical feedback signals to the electro-hydraulic regulating mechanism or the remote cloud, realizing online real-time monitoring and depth closed-loop adaptive control at the work site.

[0130] Based on this, efficient acquisition of tillage depth information can be achieved by directly accessing operating condition data from the vehicle's CAN bus, significantly reducing reliance on external sensors and complex measurement structures, and lowering system installation complexity. Furthermore, unified access and synchronous processing of various data types through the acquisition controller effectively solves the problem of timing inconsistencies between multi-source data. Building upon this, a data-driven approach is used to model and dynamically correct tillage depth, constructing a tillage depth measurement model based on an improved random forest. This model adapts to actual dynamic operating conditions, improving the stability and accuracy of tillage depth measurement. Simultaneously, the model calculation results and operating condition data are displayed in real-time on the human-machine interface, achieving integrated operation of data monitoring, processing, and result display. Through these methods, the tillage depth measurement process can be made simpler, more efficient, and more stable, enhancing the system's practicality and engineering value.

[0131] The following describes the machine learning-based deep tillage depth prediction device provided in the embodiments of this application. The machine learning-based deep tillage depth prediction device described below can be referred to in correspondence with the machine learning-based deep tillage depth prediction method described above.

[0132] Figure 6 This is a schematic diagram of a machine learning-based deep tillage depth prediction device provided in an embodiment of this application. Figure 6 As shown, the machine learning-based deep tillage depth prediction device may include: The acquisition module 610 is used to collect multi-dimensional working condition data of agricultural implements during deep tillage operations via the CAN bus of the controller area network; The processing module 620 is used to perform preprocessing operations on multi-dimensional working condition data to obtain multi-dimensional feature vectors. The processing module 620 is also used to construct the current working state vector of the deep tillage operation tools based on the multi-dimensional feature vector, determine the weight of each dimension feature vector based on the working state vector, and adjust each dimension feature vector through the weight of each dimension feature vector to obtain an adaptive input feature vector. The processing module 620 is also used to input the adaptive input feature vector into the tillage depth prediction model for processing, so as to obtain the tillage depth prediction result of deep tillage operation.

[0133] In one embodiment, the multi-dimensional operating condition data includes wheel speed, horizontal traction force, rear suspension position, engine speed, engine torque-load ratio, and instantaneous fuel consumption.

[0134] In one embodiment, the operating condition state vector includes at least one of the following: The working environment state component is used to characterize soil resistance. The power load state component is used to characterize the engine load level. The slip state component is used to characterize the adhesion state of the drive wheel. The geometric height component is used to characterize the spatial geometric reference of the rear suspension of deep tillage implements; The operating speed component is used to characterize the transient speed regulation state of the power system. The energy consumption variation component is used to characterize the dynamic load.

[0135] In one embodiment, the tillage depth prediction model is obtained by training a random forest model. The training process of the random forest model includes tuning the hyperparameters of the random forest model using a Bayesian optimization method. The hyperparameters include the number of decision trees, the maximum tree depth, the minimum number of split samples per node, and the minimum number of leaf node samples.

[0136] In one embodiment, the preprocessing operation includes at least one of the following: Normalize the operating condition data for each dimension; Outliers in multi-dimensional working condition data are replaced by the middle value of the sliding window. Outliers are feature values ​​whose deviation from the middle value of the sliding window is greater than a preset threshold. Linear interpolation is performed on missing items in multi-dimensional working condition data; The operating condition data for each dimension is filtered and smoothed.

[0137] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communications interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communications interface 720, and the memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a machine learning-based deep tillage depth prediction method, which includes: Multi-dimensional working condition data of agricultural implements during deep tillage operations are collected via the CAN bus of the controller area network, and preprocessing operations are performed on the multi-dimensional working condition data to obtain multi-dimensional feature vectors. Based on multi-dimensional feature vectors, a current working state vector of deep tillage tools is constructed, and the weights of each dimension feature vector are determined based on the working state vector. The feature vectors of each dimension are then adjusted by the weights of the feature vectors of each dimension to obtain an adaptive input feature vector. The adaptive input feature vector is fed into the tillage depth prediction model for processing to obtain the tillage depth prediction results for deep tillage operations.

[0138] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] On the other hand, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the machine learning-based deep tillage depth prediction method provided by the above methods, which includes: Multi-dimensional working condition data of agricultural implements during deep tillage operations are collected via the CAN bus of the controller area network, and preprocessing operations are performed on the multi-dimensional working condition data to obtain multi-dimensional feature vectors. Based on multi-dimensional feature vectors, a current working state vector of deep tillage tools is constructed, and the weights of each dimension feature vector are determined based on the working state vector. The feature vectors of each dimension are then adjusted by the weights of the feature vectors of each dimension to obtain an adaptive input feature vector. The adaptive input feature vector is fed into the tillage depth prediction model for processing to obtain the tillage depth prediction results for deep tillage operations.

[0140] Furthermore, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the machine learning-based deep tillage depth prediction method provided by the methods described above, the method comprising: Multi-dimensional working condition data of agricultural implements during deep tillage operations are collected via the CAN bus of the controller area network, and preprocessing operations are performed on the multi-dimensional working condition data to obtain multi-dimensional feature vectors. Based on multi-dimensional feature vectors, a current working state vector of deep tillage tools is constructed, and the weights of each dimension feature vector are determined based on the working state vector. The feature vectors of each dimension are then adjusted by the weights of the feature vectors of each dimension to obtain an adaptive input feature vector. The adaptive input feature vector is fed into the tillage depth prediction model for processing to obtain the tillage depth prediction results for deep tillage operations.

[0141] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0142] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting tillage depth in deep tillage operations based on machine learning, characterized in that, include: Multi-dimensional working condition data of agricultural implements during deep tillage operations are collected via the CAN bus of the controller local area network, and preprocessing operations are performed on the multi-dimensional working condition data to obtain multi-dimensional feature vectors. Based on the multi-dimensional feature vectors, the current working condition state vector of the deep tillage operation tool is constructed, and the weights of each dimension feature vector are determined based on the working condition state vector. The feature vectors of each dimension are then adjusted by the weights of the feature vectors of each dimension to obtain an adaptive input feature vector. The adaptive input feature vector is input into the tillage depth prediction model for processing to obtain the tillage depth prediction result for deep tillage operations.

2. The method for predicting tillage depth in deep tillage operations according to claim 1, characterized in that, The multi-dimensional operating condition data includes wheel speed, horizontal traction force, rear suspension position, engine speed, engine torque-load ratio, and instantaneous fuel consumption.

3. The method for predicting tillage depth in deep tillage operations according to claim 2, characterized in that, The operating condition state vector includes at least one of the following: The working environment state component is used to characterize soil resistance; Power load state component, which is used to characterize the engine load level; Slip state component, which is used to characterize the adhesion state of the drive wheel; Geometric height component, which is used to characterize the spatial geometric reference of the rear suspension of deep tillage implements; Operating speed component, which is used to characterize the transient speed regulation state of the power system; Energy consumption variation component, which is used to characterize dynamic load.

4. The method for predicting tillage depth in deep tillage operations according to claim 1, characterized in that, The tillage depth prediction model is obtained by training a random forest model. The training process of the random forest model includes tuning the hyperparameters of the random forest model using a Bayesian optimization method. The hyperparameters include the number of decision trees, the maximum tree depth, the minimum number of split samples per node, and the minimum number of leaf node samples.

5. The method for predicting tillage depth in deep tillage operations according to claim 1, characterized in that, The preprocessing operation includes at least one of the following: Normalize the working condition data for each dimension; The outlier values ​​in the multi-dimensional working condition data are replaced with the middle value of the sliding window, wherein the outlier value is a feature value whose deviation from the middle value of the sliding window is greater than a preset threshold. Linear interpolation is performed on the missing items in the multi-dimensional working condition data; The operating condition data for each dimension is filtered and smoothed.

6. A machine learning-based deep tillage depth prediction system, characterized in that, include: The data acquisition controller is connected to the CAN bus network of the deep tillage implement via the CAN diagnostic connector of the controller area network of the deep tillage implement, and is used to collect multi-dimensional working condition data of the deep tillage implement during the operation process. The host computer is used to perform preprocessing operations on the multi-dimensional working condition data collected by the data acquisition controller to obtain multi-dimensional feature vectors. Based on the multi-dimensional feature vectors, it constructs the current working condition state vector of the deep tillage operation implements, determines the weight of each dimension feature vector based on the working condition state vector, and adjusts each dimension feature vector through the weight of each dimension feature vector to obtain an adaptive input feature vector. The adaptive input feature vector is then input into the tillage depth prediction model for processing to obtain the tillage depth prediction result of the deep tillage operation. The host computer is also used to display the predicted tillage depth of the deep tillage operation.

7. A machine learning-based deep tillage depth prediction device, characterized in that, include: The acquisition module is used to collect multi-dimensional working condition data of agricultural implements during deep tillage operations via the CAN bus of the controller area network; The processing module is used to perform preprocessing operations on the multi-dimensional working condition data to obtain multi-dimensional feature vectors; The processing module is also used to construct the current working state vector of the deep tillage operation tool based on the multi-dimensional feature vector, determine the weight of each dimension feature vector based on the working state vector, and adjust each dimension feature vector through the weight of each dimension feature vector to obtain an adaptive input feature vector. The processing module is also used to input the adaptive input feature vector into the tillage depth prediction model for processing, so as to obtain the tillage depth prediction result of deep tillage operation.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the deep tillage depth prediction method as described in any one of claims 1 to 5.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep tillage depth prediction method as described in any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep tillage depth prediction method as described in any one of claims 1 to 5.

Images