A real-time identification method and system for the ground type of tractor field work

Abstract

The application discloses a tractor field operation ground type real-time identification method and system, and belongs to the technical field of ground and road identification. The application collects wheel vibration, soil characteristics and vehicle dynamics multidimensional data by using a sensor, comprehensively reflects ground characteristics, estimates a slip rate by combining an extended Kalman filtering algorithm, and uses the ground adhesion state discrimination ability of a ground adhesion feature enhancement system. A variation mode decomposition is introduced to extract features of a wheel vibration signal, and a multidimensional feature matrix is formed by combining the soil characteristics and the slip rate, and then unified manifold approximation and projection algorithm is used for dimension reduction. A genetic algorithm optimized extreme learning machine is used as an intelligent identification model of the ground type, data acquisition and ground type identification are completed in a single rotation period of the tractor wheel, good real-time performance and accuracy are achieved, the stability and automation level of agricultural operation are effectively improved, and technical support is provided for intelligent agricultural machinery environment perception and operation parameter self-adaptive control.

Description

Technical Field

[0001] This invention relates to the field of ground and road surface recognition, and in particular to a method and system for real-time identification of ground types for tractor field operations. Background Technology

[0002] In agricultural operations and engineering construction, the driving performance and operating strategies of vehicles or machinery vary significantly on different terrain types (such as muddy ground, dry hard ground, wet ground, and saline-alkali ground). Accurately identifying terrain types is of great significance for optimizing vehicle power distribution, braking control, and path planning.

[0003] Currently, the technologies used for ground and road surface recognition are mainly applicable to conventional road vehicles or some engineering machinery. The technical approaches fall into two main categories: First, recognition methods based on vehicle kinematics and dynamics models rely on precise mathematical modeling, typically requiring detailed construction of the vehicle's dynamic behavior and analysis of the intrinsic relationships between its motion parameters. These methods suffer from complex modeling, high computational resource consumption, and poor response time, and are particularly weak in terms of portability and universality in practical applications. Second, recognition methods based on extracting vehicle operating characteristic parameters extract feature data (such as acceleration, vibration, and speed) during vehicle operation to construct a mapping relationship between feature parameters and ground types. However, these methods have poor adaptability to complex, unstructured ground surfaces. Especially in agricultural operation scenarios, ground conditions vary significantly, including soil looseness, moisture levels, surface unevenness, and other variables, making it difficult for traditional feature recognition methods used for road vehicles to accurately adapt.

[0004] Tractors, as the core power machinery of modern agriculture, are typically used to pull various implements in farmland with complex terrain and poor road conditions. Their operating status is highly dependent on surface conditions. The type and condition of the field surface directly affect the tire adhesion, which plays a crucial role in the tractor's traction efficiency and driving performance, making it a key factor that cannot be ignored in the working environment. Uneven, soft, or slippery ground conditions can weaken the tractor's chassis system's adhesion, reduce the vehicle's stability during operation, and even lead to loss of control, resulting in safety hazards such as skidding and rollover.

[0005] Different types of soil exhibit significant differences in physical structure and mechanical properties, which in turn affect vehicle behavior parameters such as wheel speed, vehicle speed, and rolling resistance. Therefore, accurately identifying the current ground type is crucial for rationally adjusting operational parameters (such as driving speed and implement posture). However, due to the frequent changes and complex distribution of field ground conditions, the availability and accuracy of ground type information during operations are often lacking in real-time precision.

[0006] Based on the above problems, a method is proposed that can quickly and accurately identify ground types, providing a basis for intelligent control of tractor operating systems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for real-time identification of ground type during tractor field operations, in order to solve the problems in the background art.

[0008] To achieve the above objectives, the present invention provides a method for real-time identification of ground type during tractor field operations, comprising the following steps:

[0009] S1. Acquire wheel vibration signals, soil properties, wheel speed and ground speed data through sensors;

[0010] S2. Construct state variables based on wheel speed and ground speed, and recursively estimate the slip ratio using an extended Kalman filter as a ground adhesion characteristic;

[0011] S3. Perform variational mode decomposition on the wheel vibration signal to obtain the intrinsic mode function, extract the statistical features of each mode, and combine them with soil properties and slip rate to form a multidimensional feature matrix.

[0012] S4. Use UMAP to perform nonlinear dimensionality reduction on the multidimensional feature matrix to obtain a low-dimensional feature matrix;

[0013] S5. Input the low-dimensional feature matrix into the extreme learning machine model optimized by the genetic algorithm for classification reasoning, and output the ground type recognition result.

[0014] Preferably, in S1, the soil properties include soil moisture and electrical conductivity.

[0015] Preferably, in S2, the state variables are represented as follows:

[0016] x = [v, s] T ;

[0017] Where v is the ground speed and s is the slip ratio, expressed as:

[0018]

[0019] In the formula, r is the wheel diameter and w is the wheel angular velocity.

[0020] Preferably, in S2, the state transition model of the extended Kalman filter is:

[0021] X k =f(X) k-1 u k-1 )+W k ;

[0022] In the formula, X kLet W be the current state, f(·) be the evolution function from the previous time step, and W be the evolution function from the previous time step. k This is process noise;

[0023] The observation model is:

[0024] Z k =h(X) k )+V k ;

[0025] In the formula, Z k For the observed value, h(X) k V is the state-to-observation mapping function. k To observe noise;

[0026] The slip rate is estimated through the state transition update and observation update process using an extended Kalman filter.

[0027] Preferably, the specific steps of S3 are as follows:

[0028] 1) The collected wheel vibration signals were subjected to variational mode decomposition using VMD to obtain the intrinsic mode functions {u k The objective of variational mode decomposition is to minimize the sum of bandwidths for each mode, expressed as:

[0029]

[0030] In the formula, {u k ω(t)} is the k-th eigenmode function, ω k For k modal center frequencies, Let be the derivative with respect to time t, and j be the imaginary unit;

[0031] 2) Iteratively solve using ADMM to extract the statistical features of each mode, forming a complete feature vector and constructing a ground response feature description space;

[0032] The feature vector is represented as:

[0033] F = [f1, f2, ..., f n ] T ;

[0034] In the formula, f i The statistical eigenvalue of the i-th modality;

[0035] 3) Combine the obtained feature vectors with soil moisture, electrical conductivity, and slip ratio to form a multidimensional feature matrix, represented as:

[0036] X∈R N×D ;

[0037] In the formula, X is the multidimensional feature matrix, N is the number of samples, and D is the number of feature bits per sample.

[0038] Preferably, in step S4, the optimization objective of UMAP is to minimize the KL divergence of the high- and low-dimensional neighborhood distributions, thereby achieving the mapping from high-dimensional features to low-dimensional features, expressed as:

[0039]

[0040] In the formula, P ij Q ij This represents the proximity probability between samples in high-dimensional and low-dimensional spaces. The loss function is defined as follows (the optimization objective is to maximize the consistency of neighborhood relationships in both high-dimensional and low-dimensional spaces).

[0041] Preferably, in S5, the extreme learning machine is a single hidden layer feedforward network. The network structure includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer corresponds to the low-dimensional feature dimension, and the number of neurons in the output layer corresponds to the number of ground type categories.

[0042] Preferably, in step S5, the Extreme Learning Machine constructs the hidden layer output matrix by randomly initializing the input weights and biases, and the output weights are expressed as:

[0043] β=H + T;

[0044] Where β is the output weight, H is the output matrix, and H + is the Moore-Penrose pseudoinverse, and T is the target output matrix.

[0045] Preferably, in step S5, the input weights and biases are globally optimized using a genetic algorithm, with the root mean square error (RESM) as the fitness function, expressed as:

[0046]

[0047] In the formula, y i For the true value of the i-th sample, Let N be the predicted value for the i-th sample, and N be the number of samples.

[0048] Genetic algorithms generate individuals with optimal parameters through encoding, selection, crossover, and mutation mechanisms, replacing the random initialization process, enhancing the predictive ability of extreme learning machines, and avoiding overfitting.

[0049] The present invention also provides a real-time identification system for tractor field operation ground type, which is used to realize the real-time identification method for tractor field operation ground type, including a multi-source sensor acquisition unit, a wireless communication transmission module and an intelligent decision recognition module. The multi-source sensor acquisition unit includes a wheel multi-axis acceleration sensor, a soil moisture sensor, a soil conductivity sensor, a wheel speed sensor and a GNSS receiver.

[0050] The wireless communication transmission module uses Bluetooth wireless communication to transmit multi-source data from the multi-source sensor acquisition unit, and works with the STM32 receiver module and CAN bus converter to achieve data conversion and synchronization.

[0051] The intelligent decision recognition module is set as a laptop computing platform. The laptop computing platform integrates the UMAP-GA-ELM model to perform feature dimensionality reduction and classification reasoning on the multi-source data of the multi-source sensor acquisition unit and output the ground type recognition result.

[0052] The aforementioned hardware structure employs magnetic metal encapsulation to achieve high-strength impact-resistant fixation of the sensor, providing excellent rapid deployment and adaptability to highly dynamic operating environments.

[0053] Therefore, the real-time identification method and system for tractor field operation ground type of the present invention has the following beneficial effects:

[0054] (1) By integrating multiple types of sensors such as acceleration, wheel speed, GNSS, soil moisture and conductivity, and using the extended Kalman filter (EKF) algorithm, the slip ratio is dynamically estimated in real time, thereby enhancing the system's comprehensive perception of ground adhesion status and disturbance conditions.

[0055] (2) The variational mode decomposition (VMD) method is used to adaptively decompose the wheel vibration signal into multiple frequency bands, extract statistical features including kurtosis factor, impulse factor and waveform factor, and construct a high-dimensional working condition feature vector with physical interpretability and statistical discriminative power, which significantly improves the system's perception and analysis accuracy of complex ground working conditions.

[0056] (3) To address the issues of high feature space dimensionality and redundancy, a unified manifold approximation and projection algorithm (UMAP) is introduced to achieve low-dimensional embedding of the feature vector space. This preserves key discriminative information while reducing the complexity of model training and inference, thereby improving the overall system's response speed and computational stability.

[0057] (4) A lightweight ground recognition model based on Extreme Learning Machine (ELM) is constructed. Genetic Algorithm (GA) is integrated to optimize the global parameters of network input weights and hidden layer biases, effectively solving the performance uncertainty caused by random initialization. The model has high-speed training, strong generalization ability and real-time deployment capability. It can complete ground type recognition within a single wheel cycle, meeting the requirements of field operations for high real-time performance and robustness.

[0058] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0059] Figure 1This is a flowchart of the identification method according to an embodiment of the present invention;

[0060] Figure 2 This is a diagram illustrating the multi-sensor parameter fusion process based on Kalman aluminum foil according to an embodiment of the present invention.

[0061] Figure 3 This is a VMD exploded view of an embodiment of the present invention;

[0062] Figure 4 This is a diagram of the neural network structure of the Extreme Learning Machine according to an embodiment of the present invention;

[0063] Figure 5 The following is a flowchart of the UMAP dimensionality reduction process according to an embodiment of the present invention:

[0064] Figure 6 This is a flowchart of the genetic algorithm optimization ELM process according to an embodiment of the present invention;

[0065] Figure 7 The following is a graph showing the variation of error with the number of generations in an embodiment of the present invention:

[0066] Figure 8 This is a diagram showing the test set identification results of an embodiment of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0069] Example

[0070] This invention also provides a real-time identification system for tractor field operation ground types, used to realize a method for real-time identification of tractor field operation ground types. It includes a multi-source sensor acquisition unit, a wireless communication transmission module, and an intelligent decision recognition module. The multi-source sensor acquisition unit includes a wheel multi-axis acceleration sensor, a soil moisture sensor, a soil conductivity sensor, a wheel speed sensor, and a GNSS receiver; used to acquire ground disturbance response, moisture status, conductivity characteristics, and vehicle dynamic parameters.

[0071] The wireless communication transmission module uses Bluetooth wireless communication to transmit multi-source data from the multi-source sensor acquisition unit, avoiding the risk of wheel entanglement. It works with the STM32 receiver module to complete TTL signal conversion and relay, and the CAN bus converter synchronously reads wheel speed information for dynamic adjustment of the acceleration signal sampling period.

[0072] The intelligent decision recognition module is set up on a laptop computing platform. The laptop computing platform integrates the UMAP-GA-ELM model to perform feature dimensionality reduction and classification reasoning on the multi-source data of the multi-source sensor acquisition unit and output the ground type recognition result.

[0073] In this embodiment, a multi-axis accelerometer sensor (MPU6050) is fixed to the wheel hub, soil moisture and conductivity sensors are buried 5cm above the ground at the front of the vehicle, wheel speed sensors are mounted on the axle, and the GNSS receiver antenna is placed on the roof in an unobstructed position. The wheel-end sensors transmit data via a Bluetooth module (e.g., HC-05), an STM32 receiver module (STM32F103) converts TTL to USB signals, and a CAN bus converter (e.g., USBCAN-II) connects the wheel speed sensors to a laptop computer. The sensor sampling frequency is set to 100Hz, the Bluetooth communication baud rate is 115200bps, the STM32 processor clock speed is 72MHz, and the laptop computing platform is configured with an i7 processor and 16GB of memory.

[0074] In this system, all data is centrally processed by a laptop computing platform. Based on the UMAP-GA-ELM model, feature dimensionality reduction and classification reasoning are performed to quickly identify ground types (such as muddy ground, dry hard ground, wetland, saline-alkali land, etc.) within a single wheel rotation cycle. The system adopts a magnetic metal encapsulation structure, which is suitable for agricultural operating environments with strong vibrations and high-frequency impacts, and has good environmental adaptability and engineering deployment capabilities.

[0075] like Figure 1 As shown, the present invention provides a method for real-time identification of ground type during tractor field operations, comprising the following steps:

[0076] S1. Acquire wheel vibration signals, soil properties, wheel speed, and ground speed data through sensors; soil properties include soil moisture and electrical conductivity.

[0077] S2. Construct state variables based on wheel speed and ground speed, expressed as follows:

[0078] x = [v, s] T ;

[0079] Where v is the ground speed and s is the slip ratio, expressed as:

[0080]

[0081] In the formula, r is the wheel diameter and w is the wheel angular velocity.

[0082] Then, the slip ratio is estimated recursively using an extended Kalman filter. First, the state transition model of the extended Kalman filter is defined as follows:

[0083] X k=f(X) k-1 u k-1 )+W k ;

[0084] In the formula, X k Let W be the current state, f(·) be the evolution function from the previous time step, and W be the evolution function from the previous time step. k This is process noise.

[0085] The observation model is:

[0086] Z k =h(X) k )+V k ;

[0087] In the formula, Z k For the observed value, h(X) k V is the state-to-observation mapping function. k To observe noise.

[0088] During the prediction phase, the state estimate from the previous time step is used. Predict the current state using the state transition function:

[0089]

[0090] Simultaneously, the predicted covariance matrix is ​​updated accordingly, as follows:

[0091]

[0092] Among them, F k Let Q be the first-order Jacobian matrix of the state transition function f(·). k To observe the noise covariance matrix, this predicted covariance... It characterizes the uncertainty of the current estimated state and will be used as a weighting factor in the observation update.

[0093] Correcting the predicted state using observed values ​​is expressed as follows:

[0094]

[0095] In the formula, K k Here is the Kalman gain matrix.

[0096] And update the state covariance matrix, expressed as:

[0097]

[0098] In the formula, I is the identity matrix, and H is the identity matrix. k For the observation function h(X) k For the state variable, X k Jacobian matrix;

[0099] The prediction error is adjusted through the covariance update process, thereby closing the state recursion loop and improving the stability and accuracy of the estimation.

[0100] Adopting such Figure 2 The process shown involves multi-sensor parameter fusion, estimating the slip ratio through the state transition update and observation update process of the extended Kalman filter, effectively suppressing multi-source measurement errors, and obtaining key parameters of ground contact characteristics. Specifically:

[0101] The system first initializes by setting initial estimates of the vehicle's state and its covariance matrix. Then, it acquires measurements from wheel speed sensors and observations from a navigation satellite system (GNSS) in parallel. Based on the previous state estimate and the current control input, the system predicts the vehicle's state and covariance. Next, it calculates the Kalman gain to balance the error between the predicted and observed values, thereby updating the state estimate and covariance matrix. Finally, based on the updated state variables, it calculates the vehicle's slip rate, achieving dynamic estimation and fusion of the ground adhesion state.

[0102] S3. Perform variational mode decomposition on the wheel vibration signal to obtain the intrinsic mode functions, extract the statistical features of each mode, and combine them with soil properties and slip rate to form a multidimensional feature matrix; such as Figure 3 As shown, the specific steps are as follows:

[0103] 1) The collected wheel vibration signal is subjected to variational mode decomposition (VMD). The wheel vibration signal is used as the original signal for input to obtain the intrinsic mode functions {u k The objective of variational mode decomposition is to minimize the sum of bandwidths for each mode, expressed as:

[0104]

[0105] In the formula, {u k ω(t)} is the k-th eigenmode function, ω k For k modal center frequencies, Let be the derivative with respect to time t, and j be the imaginary unit.

[0106] In this process, a penalty factor and Lagrange multipliers are introduced to construct the augmented Lagrange objective function, which is expressed as:

[0107]

[0108] in, To augment the Lagrange objective function; u k (t) represents the k-th modal component; ω k λ is the center frequency of the k-th mode; λ(t) is the Lagrange multiplier; K is the number of modes; and α is the penalty factor.

[0109] Initialize the eigenfunction u of each mode k (0), center frequency w k (t), Lagrange multiplier λ(t).

[0110] 2) Iterative solution is performed using the Alternating Direction Multiplier Method (ADMM), which iteratively solves the objective function while continuously updating the modal function u. k and its corresponding center frequency w k After each update, it is determined whether the convergence condition is met. If convergence is not achieved, the iteration continues. When the convergence condition is met, K modal components with different frequency characteristics are obtained. Finally, the corresponding statistical feature parameters are extracted from each modal component to form a complete feature vector, and the ground response feature description space is constructed as the input of the subsequent ground type recognition model.

[0111] The feature vector is represented as:

[0112] F = [f1, f2, ..., f n ] T ;

[0113] In the formula, f i Let be the statistical characteristic value of the i-th modality.

[0114] 3) Combine the obtained feature vectors with soil moisture, electrical conductivity, and slip ratio to form a multidimensional feature matrix, represented as:

[0115] X = [f1, f2, ..., f n Humidity, conductivity, v, s]∈R N×D ;

[0116] In the formula, X is the multidimensional feature matrix, N is the number of samples, and D is the number of feature bits per sample.

[0117] S4. Considering the issues of high feature dimensionality and strong redundancy, UMAP is further used to perform nonlinear dimensionality reduction on the multidimensional feature matrix, employing methods such as... Figure 4 The process shown will transform the high-dimensional feature matrix X∈R N×D Mapped to a low-dimensional feature matrix X∈R N×d First, the high-dimensional feature matrix (modal statistical features, soil properties, and dynamic features) X is used as the initial data. A high-dimensional neighborhood graph is constructed by calculating the distance relationship between samples. Then, based on this neighborhood structure, UMAP embeds the high-dimensional features into the low-dimensional space while maintaining the local data structure, thereby obtaining a dense and clearly distributed low-dimensional feature representation.

[0118] The optimization objective of UMAP is to minimize the KL divergence of the high- and low-dimensional neighborhood distributions, thereby achieving a mapping from high-dimensional features to low-dimensional features, expressed as:

[0119]

[0120] In the formula, P ij Q ij This represents the proximity probability between samples in high-dimensional and low-dimensional spaces. This is the loss function.

[0121] S5. Input the low-dimensional feature matrix into the extreme learning machine model optimized by the genetic algorithm for classification reasoning, and output the ground type recognition result.

[0122] Extreme Learning Machine (ELM) is a single-hidden-layer feedforward network, such as... Figure 5 As shown, the network structure includes an input layer, hidden layers, and an output layer. The number of neurons in the input layer corresponds to the low-dimensional feature dimension, and the number of neurons in the output layer corresponds to the number of ground type categories. The Extreme Learning Machine constructs the hidden layer output matrix by randomly initializing the input weights and biases. The output weights are represented as follows:

[0123] β=H + T;

[0124] Where β is the output weight, H is the output matrix, and H + is the Moore-Penrose pseudoinverse, and T is the target output matrix.

[0125] To overcome the instability caused by random initialization, a genetic algorithm is introduced to globally optimize the input weights and biases. The specific process is as follows: Figure 6 As shown, the fitness function, expressed as the root mean square error (RMSE), is:

[0126]

[0127] In the formula, y i For the true value of the i-th sample, Let N be the predicted value for the i-th sample, and N be the number of samples.

[0128] Genetic algorithms generate individuals with optimal parameters through encoding, selection, crossover, and mutation mechanisms, replacing the random initialization process, enhancing the predictive ability of extreme learning machines, and avoiding overfitting.

[0129] The performance of the above-mentioned recognition system and method was verified, and the specific steps are as follows:

[0130] For field data collection, a serial communication script was first written in Python to read sensor data in real time and store it as a .csv file, containing fields such as timestamp, acceleration, humidity, conductivity, wheel speed, GNSS coordinates, and velocity. Then, after processing according to the steps described in this invention, a dataset with a multidimensional feature matrix was constructed. After dimensionality reduction using a unified manifold approximation and projection algorithm, 4320 sets of valid data containing 19 principal components were obtained. The sample data was divided into a training set and a test set, with 3320 sets used for model training and 1000 sets used for model validation. The training and test data were completely independent. The test set samples from different ground types were roughly evenly distributed, labeled as follows: cement ground as 1, sandy ground as 2, and clay ground as 3.

[0131] In the modeling process, the Extreme Learning Machine (GA-ELM) optimized by the genetic algorithm was used as the ground classification model. The number of input layer nodes was set to 19, corresponding to the principal component features after dimensionality reduction; the output layer had 3 nodes, corresponding to the three ground types. The genetic algorithm parameters were set as follows: population size of 20, maximum number of generations of evolution of 100, crossover probability of 0.7, mutation probability of 0.01, and fitness inheritance ratio of 0.95. The activation function of the Extreme Learning Machine was chosen as the Sigmoid function, and the number of hidden layer neurons was set to 40. During model training, the input weights and bias parameters of the neural network were iteratively optimized through the genetic algorithm. Figure 7 As shown, the root mean square error reaches its minimum in the 57th generation, indicating good model convergence.

[0132] The accuracy of the trained GA-ELM model was then validated using test set data, and the recognition results are as follows: Figure 8 As shown in the figure, the overall performance indicates that the model is most accurate in identifying cement ground samples, and the identification results for other types of ground also have good discrimination and a low misclassification rate, indicating that the working condition parameters vary significantly under different ground conditions and that the model has good classification characteristics.

[0133] Overall, for the 1000 samples in the test set, the model correctly identified 956 of them, with an accuracy rate of 95.6%, which is quite good.

[0134] The recognition results for the three types of terrain in the test set are shown in Table 1, with Precision, Recall, and F1 scores respectively.

[0135] Table 1. Precision, Recall, and F1 of the identification model.

[0136] Ground type Precison Recall F1 cement 99.7％ 94.6％ 97.1％ sand 95.1％ 93.1％ 94.1％ clay 92.8％ 93.4％ 93.1％

[0137] Analysis of Table 1 shows that the identification model performs best on cement surfaces in the test set, while its performance on clay surfaces is relatively the worst. Overall, the average precision is 95.9%, the average recall is 93.7%, and the average F1 score is 94.8%, indicating that the identification method of this invention has high accuracy and stability, and is particularly suitable for rapid identification of ground types in agricultural scenarios. Furthermore, the system only requires 1 second of tractor operating parameter data to identify the working surface, demonstrating high efficiency and adaptability, meeting the needs of rapid environmental perception and real-time control in agricultural operations.

[0138] Therefore, the present invention provides a real-time identification method and system for tractor field operation ground types. It relies on extended Kalman filter (EKF) to estimate slip ratio as the basis of operating condition parameters, combines variational mode decomposition (VMD) to extract multi-frequency feature information, uses unified manifold approximation and projection (UMAP) to achieve dimensionality reduction of high-dimensional features, and then uses extreme learning machine (ELM) neural network to construct a classification model and optimize its parameters through genetic algorithm (GA) to achieve efficient and intelligent identification under complex agricultural ground conditions.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for real-time identification of the ground type for tractor field work, characterized in that, The method comprises the following steps: S1, acquiring wheel vibration signals, soil characteristics, wheel speed and ground speed data through sensors; S2, constructing a state variable according to the wheel speed and the ground speed, recursively estimating the slip rate through an extended Kalman filter, and taking the slip rate as a ground adhesion characteristic; the state variable is expressed as: ； wherein is the ground speed, is the slip ratio, expressed as: ； wherein is the wheel radius, is the wheel angular velocity; S3, performing variational modal decomposition on the wheel vibration signals to obtain intrinsic mode functions, extracting statistical characteristics of each mode, and combining the soil characteristics and the slip rate to form a multi-dimensional feature matrix; the specific steps are as follows: 1) performing variational modal decomposition on the collected wheel vibration signals by using VMD to obtain intrinsic mode functions; the target of the variational modal decomposition is to minimize the total bandwidth of each mode, which is expressed as: ； In the formula, is the kth eigenmode function, is the kth modal center frequency, is the derivative with respect to time t, is the imaginary unit; 2) iteratively solving by using ADMM to extract the statistical characteristics of each mode, form a complete feature vector, and construct a ground response characteristic description space; The feature vector is expressed as: ； In the formula, is the i-th modal statistical feature value; 3) combining the soil humidity, the electrical conductivity and the slip rate to form a multi-dimensional feature matrix, which is expressed as: ； wherein is a multi-dimensional feature matrix, is the number of samples, is the number of feature bits for each sample; S4, performing nonlinear dimension reduction on the multi-dimensional feature matrix by using UMAP to obtain a low-dimensional feature matrix; S5, inputting the low-dimensional feature matrix into a limit learning machine model optimized by a genetic algorithm for classification reasoning to output a ground type recognition result.

2. A method of real-time ground type recognition for tractor field work according to claim 1, characterized in that: In the S1, the soil characteristics include soil humidity and electrical conductivity.

3. The method of claim 1, wherein: In the S2, the state transition model of the extended Kalman filter is: ； wherein is the current time state, is the evolution function of the previous time, is the process noise; The observation model is: ； wherein is the observation, is the state-to-observation mapping function, is the observation noise; The slip rate is estimated through the state transition update and observation update processes of the extended Kalman filter.

4. The method of claim 1, wherein: In the S4, the optimization target of the UMAP is to minimize the KL divergence of the high-dimensional and low-dimensional neighborhood distribution, realize the mapping of the high-dimensional features to the low-dimensional features, and is expressed as: ； In the formula, , represents the proximity probability between samples in high-dimensional and low-dimensional spaces, is a loss function.

5. The method of claim 1, wherein: In the S5, the limit learning machine is a single-hidden-layer feedforward network, the network structure includes an input layer, a hidden layer and an output layer, the number of input layer neurons corresponds to the low-dimensional feature dimension, and the number of output layer neurons corresponds to the number of ground type categories.

6. The method of claim 1, wherein: In the S5, the limit learning machine initializes the input weight and the bias randomly, constructs a hidden layer output matrix, and outputs the weight, which is expressed as: ； wherein, is an output weight, is an output matrix, is a Moore-Penrose pseudo-inverse, is a target output matrix.

7. The method of claim 1, wherein: In the S5, the input weight and the bias are globally optimized by using the genetic algorithm, and the root mean square error RESM is taken as the fitness function, which is expressed as: ； In the formula, is the true value of the i-th sample, is the predicted value of the i-th sample, is the number of samples; The genetic algorithm generates optimal parameter individuals through the coding, selection, crossover and mutation mechanisms, replaces the random initialization process, enhances the prediction ability of the limit learning machine, and avoids overfitting.

8. A tractor field work ground type real-time identification system for implementing the tractor field work ground type real-time identification method according to any one of claims 1-7, characterized in that: The system comprises a multi-source sensor acquisition unit, a wireless communication transmission module and an intelligent decision recognition module, the multi-source sensor acquisition unit comprises a wheel multi-axis acceleration sensor, a soil humidity sensor, a soil electrical conductivity sensor, a wheel speed sensor and a GNSS receiver; The wireless communication transmission module transmits multi-source data of the multi-source sensor acquisition unit in a Bluetooth wireless communication mode, and realizes data conversion and synchronization in cooperation with an STM32 receiving module and a CAN bus converter; The intelligent decision recognition module is a notebook computing platform, the notebook computing platform integrates a UMAP-GA-ELM model, performs feature dimension reduction and classification reasoning on multi-source data of the multi-source sensor acquisition unit, and outputs a ground type recognition result.

Images