Orchard intelligent weeding robot advancing track control method

Abstract

The invention discloses an orchard intelligent weeding robot advancing track control method, and relates to the technical field of agricultural robots, and the system comprises the following components: S1, multi-sensor environment information collection, S2, environment information fuzzification processing and fuzzy rule generation, S3, neural network model construction and parameter determination, and S4, neural network model construction and parameter determination. Neural network training and fuzzy rule optimization are carried out; and S5, proceeding track control execution is carried out. The collected environment information is fuzzified to generate fuzzy rules aiming at factors such as the distance from an obstacle, the ground gradient and the wind direction and wind speed, the rules provide a basis for subsequent advancing decision making, meanwhile, the fuzzy rules are continuously optimized by the constructed neural network model through training, and the accuracy of the decision making process is improved. The robot can make more accurate and reasonable advancing decisions according to different orchard environment conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of agricultural robots, and particularly to a method for controlling the traveling trajectory of an intelligent orchard weeding robot. Background Art

[0002] In modern agricultural production, orchard management is a crucial link. With the progress of technology, the intelligent management of orchards has gradually become a trend. Among them, intelligent weeding robots are an important part of orchard automation operations. The orchard environment is complex and changeable, and factors such as fruit trees, weeds, terrain undulations, and climatic conditions pose challenges to the control of the traveling trajectory of weeding robots.

[0003] There are deficiencies in the traditional technical environment. Traditional fuzzy processing methods often only consider single or limited environmental factors, such as the distance from obstacles, and ignore the influence of complex factors such as ground slope, wind direction, and wind speed, resulting in incomplete and inaccurate fuzzy rules. In addition, traditional methods may not be able to fully learn and optimize fuzzy rules due to insufficient training data or simple neural network structures, thereby affecting the accuracy and rationality of traveling decisions.

[0004] In summary, there are obvious deficiencies in the traveling trajectory control of traditional orchard intelligent weeding robots, especially in the fuzzy processing of environmental information, the generation of fuzzy rules, the training of neural networks, and the optimization of fuzzy rules. To overcome these shortcomings, it is particularly important to propose a method for controlling the traveling trajectory of an intelligent orchard weeding robot. Summary of the Invention

[0005] The purpose of the present invention is to make up for the deficiencies of the existing technology and provide a method for controlling the traveling trajectory of an intelligent orchard weeding robot. This method comprehensively uses a variety of sensors to collect orchard environmental information, conducts comprehensive fuzzy processing and fuzzy rule generation, and at the same time uses an advanced neural network model for training and optimization, enabling the robot to make more accurate and reasonable traveling decisions according to different orchard environmental conditions.

[0006] The present invention provides the following technical solutions to solve the above technical problems: A method for controlling the traveling trajectory of an intelligent orchard weeding robot, and the specific steps of this method are as follows;

[0007] S1, Multi-sensor environmental information collection:

[0008] Use a variety of sensors installed on the robot to collect orchard environmental information. The sensors include but are not limited to lidar, cameras, and ultrasonic sensors. The lidar is used to accurately measure the distance information between the robot and surrounding objects, and its measurement accuracy can reach ±d preccentimeters. The camera is used to obtain image information of fruit trees, weeds, and other objects in the orchard, and the image resolution is not lower than res img pixels. The ultrasonic sensor is used to assist in detecting obstacles at close range;

[0009] S2, Ambient information fuzzification and fuzzy rule generation:

[0010] Fuzzify the collected ambient information to generate fuzzy rules. The fuzzification is carried out for environmental factors such as the distance to the obstacle and the ground slope. Specifically, let the actual measured value of the distance to the obstacle be d, and define the fuzzy set of distance as: near N(0 ≤ d < d1), medium M(d1 ≤ d < d2), far F(d2 ≤ d), where the values of d1 and d2 are determined according to the analysis of the distance sensitivity in the actual orchard operation scenario. For the ground slope s, define the fuzzy set as: gentle P(0 ≤ s < s1), medium Q(s1 ≤ s < s2), steep R(s2 ≤ s), and the values of s1 and s2 are obtained through cluster analysis of a large amount of measurement data of the ground slope in different areas of the orchard. Assume that after analysis, s1 = 10 degrees and s2 = 20 degrees are determined to divide the influence degree of different slope levels on the travel trajectory. According to the above division of fuzzy sets, generate fuzzy rules;

[0011] S3, Neural network model construction and parameter determination

[0012] Construct a neural network model, which includes an input layer, a hidden layer, and an output layer. The input layer receives the fuzzy rule-related information processed by fuzzy logic. The number of neurons in the input layer, n input is determined according to the types and complexity of the fuzzy rules. Assume that after comprehensive analysis of the above distance and slope fuzzy rules and other possible environmental factor fuzzy rules, n input = 10. The hidden layer performs arithmetic processing through the connection and activation function between neurons. The number of neurons in the hidden layer, n hidden is determined based on the number of neurons in the input layer, the complexity of the expected output of the output layer, and the pre-evaluation of the neural network training effect. After multiple simulation trainings and performance evaluations, when n input = 10 and considering that the output layer needs to output accurate travel direction and speed values, n hidden = 20. The activation function of the hidden layer neurons adopts a custom function f(x), and its expression is:

[0013]

[0014] Among them, α and β are custom parameters, and their values are determined through the feature analysis of orchard environment data and the optimization experiment of the neural network training effect. Specifically, principal component analysis is performed on a large amount of collected orchard environment data to extract the feature vectors that have a greater impact on the robot's travel trajectory decision. According to the distribution of these feature vectors in different environmental scenarios, through multiple iterative training experiments, it is determined that when α = 0.5 and β = 1, the neural network can achieve a better non-linear mapping effect when processing fuzzy rule inputs, so as to more accurately output a reasonable travel decision. The output of the output layer is the optimized travel decision, including the accurate travel direction θ (unit: degree) and speed v (unit: cm / s);

[0015] S4. Neural network training and fuzzy rule optimization

[0016] The neural network is trained with training data, and the fuzzy rules are continuously learned and optimized, enabling the robot to make more accurate and reasonable travel decisions according to different orchard environmental conditions. The training data is sourced from the robot's travel situations and corresponding environmental information simulated or actually collected in different orchard environments. The collected data includes, but is not limited to, the actual travel direction and speed of the robot when encountering obstacles at different distances, the operation performance of the robot under different ground slopes, and the distribution of fruit trees and weeds in different areas of the orchard. During the training process, let the input vector of the neural network be X, the output vector be Y, and the weight matrix be W;

[0017] The forward propagation formula of the neural network is:

[0018] Z hidden =f(W inputhidden X + b hidden )

[0019] Y output =g(W hidden:output Z hidden + b output )

[0020] Among them, Z hidden is the output vector of the hidden layer, f(x) is the activation function of the hidden layer, g(x) is the output code, W input：hidden is the weight matrix connecting the input layer and the hidden layer, b hidden is the bias vector of the hidden layer. W hidden：output is the weight matrix connecting the hidden layer and the output layer, b output is the bias vector of the output layer. The initial values of the weight matrix W and the bias vector b are given by the random initialization method, and the value ranges are determined according to the scale of the neural network and the characteristics of the training data;

[0021] During the training process, a custom loss function L is used to measure the difference between the neural network output and the expected output. The expression of the loss function is as follows:

[0022]

[0023] where N is the number of samples in the training data, i is an index variable, is the neural network output of the i-th training sample, is the true expected output of the i-th training sample, λ is the regularization parameter, j is an index variable. Through the backpropagation algorithm, the weight matrix W and the bias vector b are updated according to the loss function, continuously adjusting the parameters of the neural network, enabling it to continuously learn and optimize the fuzzy rules during the training process, so that the robot can make more accurate and reasonable travel decisions according to different orchard environmental conditions;

[0024] S5, Execute the travel trajectory control

[0025] The robot controls its own travel trajectory in real time according to the optimized travel decision output by the neural network, achieving efficient, accurate and safe weeding operations.

[0026] Furthermore, the installation position and angle of the S1 lidar are carefully designed to ensure that it can collect the distance information around the robot omnidirectionally and without dead angles. The lidar is installed at the center position on the top of the robot, and the vertical downward angle deviation is controlled within the range of ±Δθ radar degrees, and it can perform 360° rotational scanning in the horizontal direction with a rotational speed of v rot，radar rotations per second. Through such installation position and angle settings, the space around the robot can be covered to the greatest extent, providing more accurate and comprehensive distance data for subsequent environmental information collection and travel trajectory control.

[0027] Even further, the S1 camera uses a high-definition wide-angle lens with a viewing angle range of not less than θ view，img degrees. The high-definition wide-angle lens can obtain a wider orchard scene image in one shot, facilitating more accurate identification of the positions and shapes of fruit trees, weeds and other objects. At the same time, the camera is equipped with an autofocus and anti-shake function. The autofocus function can automatically adjust the focal length according to the distance of the shooting object to ensure clear images, and the anti-shake function can reduce image blurring caused by vibration factors during the robot's travel, further improving the image quality, thus providing more reliable image information for the fuzzification process and subsequent travel trajectory control.

[0028] Even further, the operating frequency range of the S1 ultrasonic sensor is from f min，ultra to f max，ultraHz. In this frequency range, the ultrasonic sensor has good reflection characteristics for the common obstacle materials in the orchard, and can detect obstacles at close range more accurately. In addition, the detection distance accuracy of the ultrasonic sensor can reach ±d pred，ultr a centimeter. By combining with the information collected by lidar and camera, a more perfect environmental perception system is formed, providing richer environmental information for generating accurate fuzzy rules and optimizing neural networks.

[0029] Furthermore, in the S2 fuzzy processing, in addition to the two environmental factors of the distance to the obstacle and the ground slope, the two factors of wind direction and wind speed in the orchard are also considered. Let the wind direction be w and the wind speed be v w , for the wind direction, the fuzzy sets are defined as: downwind S (w ∈ [w1, w2]), upwind N w (w ∈ [w3, w4]), crosswind L w , where the values of w1, w2, w3, and w4 are determined according to the statistical data of the perennial wind direction in the area where the orchard is located. Suppose it is determined after analysis that w1 = 0°, w2 = 45°, w3 = 135°, w4 = 180°. For the wind speed, the fuzzy sets are defined as: gentle breeze B (0 ≤ v w < v1), moderate wind M w (v1 ≤ v w < v2), strong wind G (v2 ≤ v w ), where the values of v1 and v2 are obtained through statistical analysis of a large amount of measured wind speed data in different seasons in the orchard. Suppose it is determined after analysis that v1 = 2 m / s and v2 = 5 m / s. When considering the factors of wind direction and wind speed, the generated fuzzy rules will be more complex and comprehensive, so that the influence of orchard environmental factors on the robot's travel trajectory can be considered more comprehensively, making the travel decision more accurate.

[0030] Furthermore, the hidden layer of the S3 neural network model adopts a multi-layer structure, that is, it contains multiple hidden layers. A three-layer hidden layer structure is adopted, denoted as H1, H2, and H3 respectively. The number of neurons in each hidden layer is determined according to the number of neurons in the input layer, the complexity of the expected output of the output layer, and the pre-evaluation of the training effect of the neural network. Suppose after analysis and experiments, it is determined that the number of neurons in the H1 layer is n hidden1, The number of neurons in the H2 layer is n hidden2 , and the number of neurons in the H3 layer is n hidden3 , let when the number of neurons in the input layer is n input , after analysis and experiments, it is determined that the number of neurons in the H1 layer is nhidden1 = 15, the number of neurons in the H2 layer is n hidden2 = 25, and the number of neurons in the H3 layer is n hidden3= 20. Adopting a multi-layer hidden layer structure can increase the expressive power of the neural network, enabling it to better handle complex fuzzy rule inputs and non-linear mappings, thus more accurately outputting optimized travel decisions. At the same time, the connection weight matrix and bias vector between different hidden layers also need to be optimized through the training process. Their initial values and update methods are similar to those described in Claim 1 for the single hidden layer case, but the characteristics of the multi-layer structure need to be considered during the training process, such as using layer-by-layer training or joint training methods, to ensure that the neural network can effectively learn and optimize fuzzy rules and provide more accurate travel decisions for the robot.

[0031] Furthermore, during the training process of the S4 neural network, in addition to using the conventional random initialization method to assign initial values to the weight matrix and bias vector, a pre-training strategy is also adopted. According to the basic characteristics of the orchard environment information, a simplified neural network model is constructed. This model is of a smaller scale, with the number of neurons in the input layer being ninput s imple, and the number of neurons in the hidden layer being nhidden s imple, assuming that when constructing the simplified neural network model, the number of neurons in the input layer is n input，imple = 5, and the number of neurons in the hidden layer is The output layer outputs some preliminary travel decision-related information, and this simplified neural network model is pre-trained. The training data used is a representative part extracted from the complete training dataset. The purpose of pre-training is to let this simplified neural network model first learn some basic laws about the orchard environment and travel decisions. Then, the weight matrix and bias vector obtained from pre-training are used as the initial values and assigned to the neural network model of the full scale described in Claim 1, and then the formal training process is carried out. Through this pre-training strategy, the training speed of the neural network can be accelerated, the training effect can be improved, enabling the robot to learn the optimized travel decisions faster, and thus more effectively control the travel trajectory. Especially when dealing with large-scale training data, the advantages of this pre-training strategy are more obvious.

[0032] Furthermore, in addition to controlling the movement of the robot according to the traveling direction and speed output by the neural network, the traveling control system of the robot also has an emergency handling mechanism. When the environmental information collected by the sensor indicates that the robot is about to collide with an obstacle and the traveling decision output by the neural network cannot avoid the collision in time, the traveling control system will immediately activate the emergency handling mechanism. The emergency handling mechanism includes but is not limited to emergency braking, that is, instantly reducing the speed of the robot to zero, or emergency steering to change the traveling direction at the maximum angle to avoid colliding with the obstacle. At the same time, after activating the emergency handling mechanism, the robot will immediately re-collect the environmental information and re-perform the fuzzification process and neural network training steps to quickly restore the normal traveling trajectory control and ensure the safe operation of the robot in the orchard.

[0033] Compared with the prior art, a method for controlling the traveling trajectory of an intelligent orchard weeding robot has the following beneficial effects:

[0034] First, the present invention generates fuzzy rules for factors such as the distance from obstacles, ground slope, and wind direction and speed by fuzzifying the collected environmental information. These rules provide a basis for subsequent traveling decisions. At the same time, the constructed neural network model continuously optimizes these fuzzy rules through training, enabling the robot to make more accurate and reasonable traveling decisions according to different orchard environmental conditions.

[0035] Second, the present invention realizes a comprehensive and accurate perception of the orchard environment through multi-sensor environmental information collection and fusion. The combined use of lidar, cameras, and ultrasonic sensors not only improves the distance measurement accuracy of the robot for surrounding objects but also enhances the image recognition ability of fruit trees and weeds in the orchard and the auxiliary detection ability for close-range obstacles. This all-round environmental perception system provides rich and accurate data support for subsequent fuzzification, neural network model construction, and training, thus significantly improving the traveling trajectory control accuracy and safety of the robot in a complex orchard environment. Brief Description of the Drawings

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 An operation flowchart of a method for controlling the traveling trajectory of an intelligent orchard weeding robot. Detailed Embodiments

[0038] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] Embodiment 1

[0040] This embodiment describes that in spring, the fruit trees in the orchard are in bloom and the weeds begin to grow. The wind is moderate but the wind direction is changeable. At this time, the intelligent weeding robot in the orchard needs to avoid colliding with fruit trees and weeds while taking into account the impact of wind direction and wind speed on its movement, and perform efficient and safe weeding operations. The terrain of the orchard is complex, including flat land, slopes and gullies, which poses higher requirements for the control of the robot's movement trajectory.

[0041] Precisely measure the distances between the robot and fruit trees, weeds, other obstacles, and the terrain edge to provide high-precision spatial positioning information for the robot. Capture the image information of fruit trees and weeds in the orchard. The autofocus function ensures clear images. At the same time, the camera can also be used to identify the types and growth states of fruit trees so that the robot can intelligently avoid fruit trees or adjust the weeding strategy. Assist in detecting short-range obstacles, especially in areas with low weeds and large terrain undulations, to provide additional safety guarantees. Real-time monitor the wind direction and wind speed in the orchard to provide dynamic environmental information for the robot to adjust its movement direction and speed. Measure the height and tilt angle of the ground to provide terrain information for the robot to better adapt to the complex terrain.

[0042] Fuzzify the collected distance, slope, wind direction, wind speed, and terrain information to convert continuous data into discrete fuzzy sets. According to the characteristics of the spring orchard, set the fuzzy sets and fuzzy rules for distance, slope, wind direction, wind speed, and terrain. Let the actual measured value of the distance to the obstacle be d, and define the fuzzy set of distance as: near N(0≤d<d1), medium M(d1≤d<d2), far F(d2≤d), where the values of d1 and d2 are determined based on the analysis of the distance sensitivity in the actual orchard operation scenario. After multiple experiments and statistical analyses, it is found that when d1 = 200 cm and d2 = 500 cm, it can ensure the safe operation of the robot while effectively distinguishing the influence of different distances on the travel decision. For the ground slope s, define the fuzzy set as: gentle P(0≤s<s1), medium Q(s1≤s<s2), steep R(s2≤s). The values of s1 and s2 are obtained through cluster analysis of a large amount of measured data of the ground slope in different areas of the orchard. Assume that after analysis, s1 = 10 degrees and s2 = 20 degrees are determined to divide the influence degree of different slope levels on the travel trajectory. According to the above division of fuzzy sets, generate fuzzy rules, and through fuzzy logic operations, fuse multiple fuzzy information into a comprehensive fuzzy decision to provide a basis for the input of the neural network.

[0043] Construct a neural network model with multiple hidden layers, which includes an input layer, hidden layers, and an output layer. The input layer receives the fuzzy rule-related information processed by fuzzy logic. The number of neurons in the input layer is n input Determined according to the types and complexities of the fuzzy rules. Assume that after comprehensive analysis of the above distance and slope fuzzy rules and other possible environmental factor fuzzy rules, n is determined input = 10. The hidden layers perform arithmetic processing through the connections and activation functions between neurons. The number of neurons in the hidden layers is n hidden Determined based on the number of neurons in the input layer, the complexity of the expected output of the output layer, and the pre-evaluation of the training effect of the neural network. After multiple simulation trainings and performance evaluations, when m input = 10 and considering that the output layer needs to output accurate travel direction and speed values, n is determined hidden = 20. The activation function of the hidden layer neurons adopts a custom function f(x), and its expression is:

[0044]

[0045] Among them, α and β are custom parameters, and their values are determined through the feature analysis of orchard environment data and the optimization experiment of the neural network training effect. Specifically, principal component analysis is performed on a large amount of collected orchard environment data to extract the feature vectors that have a greater impact on the robot's travel trajectory decision-making. According to the distribution of these feature vectors in different environmental scenarios, through multiple iterative training experiments, it is determined that when α = 0.5 and β = 1, the neural network can achieve a better non-linear mapping effect when processing fuzzy rule inputs, so as to more accurately output a reasonable travel decision. The output of the output layer is the optimized travel decision, including the precise travel direction θ (unit: degree) and speed v (unit: cm / s). The deep learning algorithm is used for training. The input layer receives the fuzzified environmental information, including distance, slope, wind direction, wind speed, and terrain. The hidden layer performs non-linear mapping through a custom activation function to extract the deep features of the environmental information. The output layer outputs the precise travel direction and speed, as well as possible obstacle avoidance strategies. According to the training data and experimental results, the parameters and structure of the neural network are adjusted to improve the accuracy and generalization ability of the model.

[0046] The neural network is trained using training data simulated or actually collected in the spring orchard environment.

[0047]

[0048] Among them, N is the number of samples of the training data, i is an index variable, is the neural network output of the i-th training sample, is the true expected output of the i-th training sample, λ is the regularization parameter, j is an index variable. The training data should include environmental information under different weather conditions, different terrains, and different fruit tree distributions. During the training process, the fuzzy rules are continuously optimized so that the robot can make accurate travel decisions according to different environmental conditions. For example, when the wind speed is large, the robot should choose the travel direction more carefully to avoid being blown off by the strong wind. Through cross-validation and performance evaluation, the optimal neural network model and fuzzy rule combination are selected.

[0049] The robot controls the travel trajectory in real time according to the travel decision output by the neural network. During the travel process, the robot should continuously collect environmental information and update the travel decision according to the new information. In case of emergencies, such as suddenly appearing obstacles or terrain changes, the robot should be able to immediately activate the emergency handling mechanism, such as emergency braking, emergency steering, or adjusting the weeding strategy. The robot should also have the functions of autonomous navigation and obstacle avoidance, and be able to independently find the optimal path in a complex environment and avoid collisions with fruit trees and weed obstacles.

[0050] A remote monitoring center is established to monitor the working status and travel trajectory of the robot in real time. When the robot encounters complex situations that it cannot handle, the monitoring center can intervene remotely to provide guidance or adjust strategies for the robot, and collect the working data of the robot and user feedback for subsequent optimization and improvement.

[0051] Embodiment 2

[0052] This embodiment describes that in a complex and changeable orchard environment, especially in areas with steep slopes, the operation of intelligent weeding robots faces severe challenges. These areas not only have large terrain undulations, but also the soil may become soft due to rain erosion, further increasing the instability of the robot's movement. In addition, the narrow passages between fruit trees and possible obstacles also require the robot to have a high level of environmental perception ability and flexible trajectory planning ability.

[0053] It is used for high-precision terrain scanning to generate three-dimensional point cloud data to help the robot identify slopes, obstacles and ground textures. Combining with deep learning algorithms, it can identify fruit trees, pedestrians and other potential obstacles, and at the same time monitor the lighting conditions to adjust the camera exposure to ensure image quality. It is used for close-range obstacle detection, especially to provide supplementary information in low light or visual blind areas. When used in combination, it provides the absolute position and attitude information of the robot. Especially when the GPS signal is weak or lost, INS can maintain the positioning accuracy in the short term. It evaluates the softness of the soil to provide a basis for the robot to adjust the driving force and tire pressure.

[0054] In addition to "gentle", "medium" and "steep", the level of "extremely steep" is added to more finely reflect the slope change. According to the size, quantity and type of obstacles (such as movable / fixed), different fuzzy sets are set. Let the actual measured value of the distance from the obstacle be d, and the fuzzy set of distance is defined as: near N(0≤d<d1), medium M(d1≤d<d2), far F(d2≤d), where the values of d1 and d2 are determined according to the analysis of the distance sensitivity in the actual orchard operation scenario. After multiple tests and statistical analyses, it is found that when d1 = 200 cm and d2 = 500 cm, while ensuring the safe operation of the robot, it can effectively distinguish the influence of different distances on the travel decision. For the ground slope s, the fuzzy set is defined as: gentle P(0≤s<s1), medium Q(s1≤s<s2), steep R(s2≤s), and the values of s1 and s2 are obtained through cluster analysis of a large amount of measured data of the ground slopes in different areas of the orchard. Suppose after analysis, it is determined that s1 = 10 degrees and s2 = 20 degrees, so as to divide the influence degree of different slope levels on the travel trajectory. According to the above division of fuzzy sets, fuzzy rules are generated to guide the obstacle avoidance strategy of the robot, and combined with the actual orchard operation experience, travel strategies are formulated for different slopes, obstacle densities and soil conditions, such as deceleration, detour or adjustment of the travel direction.

[0055] In addition to slope information, multi-dimensional features such as obstacle distance, type, soil humidity, and light intensity are also considered to enhance the richness of the model's input. A deep neural network structure is adopted, the number of hidden layers and neurons is increased, and ReLU and LeakyReLU non-linear activation functions are used to improve the model's learning ability, prevent overfitting of the model, and improve the generalization ability to ensure reasonable decisions can also be made under unseen slope and obstacle combinations.

[0056] By simulating scenarios with different slope, light, and obstacle combinations, more training data is generated to enhance the robustness of the model. During the robot's operation, new data is continuously collected for online fine-tuning of neural network parameters and fuzzy rules to achieve self-optimization. The operator is allowed to intervene when necessary, either through manual control or by providing feedback, to accelerate the model's learning process.

[0057] Based on real-time environmental information and the output of the neural network, the travel path is dynamically adjusted to ensure safety and efficiency. According to the slope change, the motor output torque and tire pressure are automatically adjusted to keep the robot stable. When the risk of an impending collision or instability is detected, emergency braking is immediately initiated or obstacle avoidance actions are performed to ensure safety.

[0058] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed by the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for controlling the trajectory of an orchard intelligent weeding robot, characterized in that: The specific steps of this method are: S1, Multi-sensor environmental information collection The orchard environment information is collected using a variety of sensors installed on the robot. The sensors include but are not limited to laser radar, cameras, and ultrasonic sensors. The laser radar is used to accurately measure the distance between the robot and surrounding objects, and its measurement accuracy can reach ±d prec The camera is used to obtain image information of fruit trees, weeds and other objects in the orchard. The image resolution is not less than res img ' pixels, and the ultrasonic sensor assists in detecting obstacles at close range; S2, Fuzzy processing of environmental information and generation of fuzzy rules The collected environmental information is fuzzified to generate fuzzy rules. The fuzzy processing is performed on the environmental factors of the distance to the obstacle and the ground slope, as follows: Assume that the actual measured value of the distance to the obstacle is d, and define the fuzzy set of distance as: near N (0≤d<d1), medium M (d1≤d<d2), far F (d2≤d), where the values ​​of d1 and d2 are determined according to the analysis of distance sensitivity in the actual operation scene of the orchard. For the ground slope s, the fuzzy set is defined as: gentle P (0≤s<s1), medium Q (s1≤s<s2), steep R (s2≤s), and the values ​​of s1 and s2 are obtained by clustering analysis of a large amount of measurement data of the ground slope in different areas of the orchard. It is assumed that after analysis, s1=10 degrees and s2=20 degrees are determined to divide the degree of influence of different slope levels on the travel trajectory. According to the division of the above fuzzy sets, fuzzy rules are generated; S3, neural network model construction and parameter determination Construct a neural network model, which includes an input layer, a hidden layer and an output layer. The input layer receives fuzzy rule related information after fuzzy logic processing. The number of neurons in the input layer is n. input According to the type and complexity of fuzzy rules, it is assumed that after comprehensive analysis of the above distance and slope fuzzy rules and other environmental factor fuzzy rules that may be involved, n is determined. input = 10, the hidden layer performs computations through connections between neurons and activation functions, and the number of neurons in the hidden layer is n hidden ′ is determined based on the number of neurons in the input layer, the complexity of the expected output of the output layer, and the pre-evaluation of the training effect of the neural network. After multiple simulation training and performance evaluation, when n input = 10 and considering that the output layer needs to output accurate direction and speed values, determine n hidden =20, the activation function of the hidden layer neuron uses a custom function f(x), which is expressed as: Among them, α and β are custom parameters, and their values ​​are determined by feature analysis of orchard environmental data and optimization experiments of neural network training effects. Specifically, principal component analysis is performed on a large amount of collected orchard environmental data to extract feature vectors that have a greater impact on the robot's trajectory decision. According to the distribution of these feature vectors in different environmental scenarios, through multiple iterative training experiments, it is determined that when α=0.5 and β=1, the neural network can achieve a better nonlinear mapping effect when processing fuzzy rule input, thereby more accurately outputting reasonable travel decisions. The output layer outputs optimized travel decisions, including accurate travel direction θ and speed v; S4, Neural Network Training and Fuzzy Rule Optimization The neural network is trained through training data, and fuzzy rules are continuously learned and optimized, so that the robot can make more accurate and reasonable travel decisions according to different orchard environmental conditions. The training data comes from the robot's travel conditions and corresponding environmental information simulated or actually collected in different orchard environments. The collected data includes but is not limited to the actual travel direction and speed of the robot at different distances from obstacles, the robot's operating performance under different ground slopes, and the distribution of fruit trees and weeds in different areas of the orchard. In the training process, the input vector of the neural network is X, the output vector is Y, and the weight matrix is ​​W; The forward propagation formula of the neural network is: Z hidden =f(W input:hidden X+b hidden ) Y output =g(W hidden:output Z hidden +b output ) Among them, Z hidden is the output vector of the hidden layer, f(x) is the activation function of the hidden layer, g(x) is the output code, W input:hidden is the weight matrix connecting the input layer and the hidden layer, b hidden is the bias vector W of the hidden layer hidden:output is the weight matrix connecting the hidden layer and the output layer, b output is the bias vector of the output layer. The initial values ​​of the weight matrix W and the bias vector b are given by random initialization method, and the value range is determined according to the scale of the neural network and the characteristics of the training data; During the training process, a custom loss function L is used to measure the difference between the neural network output and the expected output. The loss function expression is: Where N is the number of samples in the training data, i is an index variable, is the neural network output of the ith training sample, is the true expected output of the i-th training sample, λ is the regularization parameter, and j is an index variable. Through the back-propagation algorithm, the weight matrix W and the bias vector b are updated according to the loss function, and the parameters of the neural network are continuously adjusted so that the robot can continuously learn and optimize the fuzzy rules during the training process, so that the robot can make more accurate and reasonable travel decisions according to different orchard environmental conditions; S5, travel trajectory control execution The robot controls its own movement trajectory in real time based on the optimized movement decisions output by the neural network, achieving efficient, accurate and safe weeding operations.

2. According to claim 1, a method for controlling the trajectory of an orchard intelligent weeding robot is characterized in that: The installation position and angle of the S1 laser radar are carefully designed. The laser radar is installed at the top center of the robot, and its vertical downward angle deviation is controlled within ±Δθ radar Within the range of degrees, 360° rotation scanning can be achieved in the horizontal direction, and the rotation speed is v rot，radar Revolution / second.

3. According to claim 1, a method for controlling the trajectory of an orchard intelligent weeding robot is characterized in that: The S1 camera uses a high-definition wide-angle lens, and its viewing angle range is not less than θ view，img The high-definition wide-angle lens can capture a wider range of orchard scene images in one shot. The camera is also equipped with autofocus and anti-shake functions. The autofocus function can automatically adjust the focal length according to the distance of the subject, and the anti-shake function can reduce image blur caused by vibration during the robot's movement.

4. The method for controlling the trajectory of an orchard intelligent weeding robot according to claim 1, characterized in that: The operating frequency range of the S1 ultrasonic sensor is f min，ultra to f max，ultra In this frequency range, ultrasonic sensors have good reflection characteristics for common obstacle materials in orchards. In addition, the detection distance accuracy of ultrasonic sensors can reach ±d prec，ultra Centimeters, by combining with the information collected by lidar and cameras, a more complete environmental perception system is formed.

5. The method for controlling the trajectory of an orchard intelligent weeding robot according to claim 1, characterized in that: In the S2 fuzzy processing, in addition to the two environmental factors of the distance to the obstacle and the ground slope, the wind direction and wind speed in the orchard are also considered. Let the wind direction be w and the wind speed be v w For wind direction, the fuzzy set is defined as: downwind S (w∈[w1, w2]), headwind N w (w∈[w3,w4]), crosswind L w , where the values ​​of w1, w2, w3, and w4 are determined according to the annual wind direction statistics of the orchard area. Assuming that w1 = 0°, w2 = 45°, w3 = 135°, and w4 = 180° are determined after analysis, the fuzzy set for wind speed is defined as: breeze B (0≤v w <v1)、Stroke M w (v1≤v w <v2)、Strong wind G(v2≤v w ), where the values ​​of v1 and v2 are obtained by statistical analysis of a large amount of measurement data of wind speed in the orchard in different seasons. Assuming that v1 = 2 m / s and v2 = 5 m / s are determined after analysis, when the wind direction and wind speed factors are considered, the generated fuzzy rules will be more complex and comprehensive.

6. The method for controlling the trajectory of an orchard intelligent weeding robot according to claim 1, characterized in that: The hidden layer of the S3 neural network model adopts a multi-layer structure, that is, it contains multiple hidden layers, and adopts a three-layer hidden layer structure, which are denoted as H1, H2, and H3 respectively. The number of neurons in each hidden layer is determined according to the number of neurons in the input layer, the complexity of the expected output of the output layer, and the pre-evaluation of the neural network training effect. It is assumed that after analysis and experiments, the number of neurons in the H1 layer is determined to be n hidden1 , the number of neurons in layer H2 is n hidden2 , the number of neurons in layer H3 is n hidden3 , assuming that the number of neurons in the input layer is n input After analysis and experiments, it is determined that the number of neurons in the H1 layer is n hidden1 =15, the number of neurons in layer H2 is |n hidden2 =25, the number of neurons in layer H3 is n hidden3 =20, the use of a multi-layer hidden layer structure can increase the expressive power of the neural network, enabling it to better handle complex fuzzy rule inputs and nonlinear mappings, thereby more accurately outputting optimized travel decisions.

7. The method for controlling the trajectory of an orchard intelligent weeding robot according to claim 1, characterized in that: In addition to using the conventional random initialization method to give the initial values ​​of the weight matrix and the bias vector during the S4 neural network training process, a pre-training strategy was also used to construct a simplified neural network model based on the basic characteristics of the orchard environmental information. The model is small in scale and has a number of neurons in the input layer. The number of neurons in the hidden layer is Assume that when constructing a simplified neural network model, the number of neurons in the input layer is n input，imple =5, the number of neurons in the hidden layer is The output layer outputs some preliminary information related to the travel decision, and pre-trains this simplified neural network model. The training data used is a representative part of the data extracted from the complete training data set.

8. The method for controlling the trajectory of an orchard intelligent weeding robot according to claim 1, characterized in that: In addition to controlling the movement of the robot according to the travel direction and speed output by the neural network, the travel control system of the robot also has an emergency handling mechanism. When the environmental information collected by the sensor indicates that the robot is about to collide with an obstacle and the travel decision output by the neural network cannot avoid the collision in time, the travel control system will immediately activate the emergency handling mechanism. The emergency handling mechanism includes but is not limited to emergency braking, that is, instantly reducing the speed of the robot to zero, or emergency turning to change the travel direction at the maximum angle. At the same time, after starting the emergency handling mechanism, the robot will immediately re-collect environmental information and re-perform fuzzy processing and neural network training steps.

Images