Twisting machine inspection path dynamic programming method considering personnel comfort

Abstract

The present application relates to the technical field of equipment inspection control, and particularly relates to a method for dynamically planning an inspection path of a two-for-one twister considering comfort of personnel, which is based on a tense space modeling, takes the comfort of personnel as a part of input states, processes a social stress feature as a weight after a fusion processing of states of an inspection robot and pedestrians to calculate an attention score of the pedestrians to represent a relative importance of each pedestrian to the inspection robot, and takes the comfort of personnel as a part of a reward function to make an evaluation of a value network on a current state value more accurate, to make the value network select more social actions in a training process, to improve efficiency and success rate in navigation, and to make a decision of the inspection robot better meet a comfort demand of the pedestrians and be more social.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment inspection control, and particularly relates to a double twister inspection path dynamic programming method considering personnel comfort. BACKGROUND

[0002] The double twister in the spinning industry plays an important role in improving the strength, elasticity and uniformity of yarn. The stable operation of the double twister is undoubtedly a key link to ensure production efficiency and product quality, and therefore regular inspection by the inspection robot has become an indispensable management measure. Workers need to frequently shuttle between various workshops, and when the inspection robot meets the workers, the problems of obstacle avoidance and path planning need to be considered. The traditional reactive obstacle avoidance and navigation technology is prone to cause the path planning problem to fall into an extreme point, and the highly random dynamic crowd will increase the autonomous obstacle avoidance calculation amount of the inspection robot.

[0003] Although the traditional obstacle avoidance and path planning algorithm can better achieve the safe navigation task without collision, in the dynamic crowd environment, the workers move randomly before reaching the destination, and the trajectories of the workers and the inspection robot will influence each other. Simply treating the highly random pedestrians as general dynamic obstacles is not conducive to the inspection robot to achieve a good human-robot coexistence mode of obstacle avoidance and navigation task in a complex environment and to better meet the comfort needs of the pedestrians in the obstacle avoidance and navigation process.

[0004] Specifically, the existing obstacle avoidance and navigation method in the spinning industry has the following several methods:

[0005] 1. The relative distance of the worker when inspecting the double twister is detected by a ranging method. This method has the following disadvantages:

[0006] 1) Low efficiency: During the double twister inspection process, the computer needs to process a large amount of data from different sensors to accurately determine the relative position of the worker and the robot. However, the traditional ranging method is often inefficient in processing these data, making it difficult to make accurate control decisions in real time.

[0007] 2) Randomness: The high randomness of worker activities in the textile workshop brings great uncertainty to the obstacle avoidance task. Workers can change their walking routes or speeds at any time, making it difficult for the robot to accurately predict and avoid potential collisions.

[0008] 3) Incomplete data recording: Traditional mathematical analysis methods often rely on static or simplified environmental models, which may not fully reflect the dynamic changes in the textile workshop.

[0009] 2. The neural network deep learning method for double twister inspection has the following disadvantages:

[0010] 1) High training cost: Deep learning models require a large amount of labeled data for training, which is a time-consuming and labor-intensive task in the environment of doubling machine inspection. In addition, due to the special nature of the textile workshop, the working environment is very poor, and the collected data may have high diversity and complexity, which further increases the difficulty of model training.

[0011] 2) Difficulty in adapting to the special environment of the textile workshop: existing solutions are mostly general-purpose, lacking optimization for the special needs of doubling machines. In the environment of narrow passages and frequent personnel movement, the obstacle avoidance and navigation capabilities of the inspection robot are limited.

[0012] 3) Lack of consideration of personnel comfort, only training the model to achieve obstacle avoidance, which is relatively single.

[0013] The above two methods have the following problems:

[0014] 1. Lack of targeted optimization: existing obstacle avoidance and navigation methods in the spinning industry are mostly general-purpose and do not optimize for the special needs of doubling machines. In the environment of doubling machine inspection, the robot needs to be able to accurately identify and respond to various equipment failures and abnormal situations, while general-purpose solutions often cannot meet these needs.

[0015] 2. Insufficient environmental adaptability: the textile workshop has its unique environmental characteristics, such as narrow passages, frequent personnel movement, and equipment layout. Existing systems often lack adaptability to these special environments, resulting in poor performance in actual application.

[0016] 3. Insufficient real-time response capability: during the doubling machine inspection process, the robot needs to be able to sense environmental changes in real time and make corresponding adjustments. However, existing systems are mostly passive detection types, making it difficult to actively predict and respond to sudden situations. This may increase the risk of collision between the robot and the worker, and even affect the smooth progress of the production process. SUMMARY

[0017] The technical problem to be solved by the present application is: In order to overcome the above technical problems, the present application provides a doubling machine inspection path dynamic planning method considering personnel comfort, which models the environment in a tense space and considers personnel comfort, designs a SARL* model, introduces a reward function design of personnel comfort, and is deployed on a doubling inspection robot to ensure the safety and comfort of the surrounding crowd during the obstacle avoidance and navigation process of the inspection robot.

[0018] The technical solution adopted by the present application to solve its technical problems is: a doubling machine inspection path dynamic planning method considering personnel comfort, comprising the following steps:

[0019] S1: Model the inspection environment by laser radar to generate a map, initialize an obstacle avoidance space model centered on the twister inspection robot, and receive end point information;

[0020] S2: Generate an inspection path according to the current position of the twister inspection robot and the end point information;

[0021] S3: The twister inspection robot detects pedestrians near the production line and obtains their pose information, including moving direction, moving speed, and position;

[0022] S4: Model the pedestrians according to their pose information, combine the twister inspection robot obstacle avoidance space model and the physical model of the pedestrians to make judgments, and add personnel comfort as a weight to the calculation network;

[0023] S5: Generate a window path according to the obstacle avoidance space and the inspection path;

[0024] S6: Identify pedestrians from the industrial camera loaded on the twister inspection robot, combine laser radar data to measure the moving direction, moving speed, position of the pedestrians relative to the twister inspection robot, and the position of the window path relative to the twister inspection robot, and input them into the trained deep reinforcement learning model to output linear and angular velocities for controlling the twister inspection robot;

[0025] S7: The twister inspection robot navigates forward according to the linear and angular velocities;

[0026] S8: Determine whether the twister inspection robot has reached the end point. If it has, the navigation ends. If it has not, return to step S2.

[0027] In step S1, the obstacle avoidance space model uses a tight space-based environment modeling, including the following steps:

[0028] S11: The tight space of the twister inspection robot can be represented as a circle and a semi-ellipse. The circle represents the safety space of the twister inspection robot, and the semi-ellipse represents the comfort space of the twister inspection robot;

[0029] S12: The radius of the circle and the short axis radius b1 of the semi-ellipse are both equal to the radius r1 of the twister inspection robot, and the long axis radius a front1 of the semi-ellipse is related to the speed v t of the twister inspection robot. b1 = r1, and k1 is a parameter related to the front of the twister inspection robot.

[0030] In step S2, the inspection path is autonomously planned in real time by the dynamic programming SRAL* algorithm. In step S3, the pedestrian detection and pose information acquisition includes the following steps:

[0031] S31: Obtain the pedestrian moving speed through the pedestrian perception module, and generate the point cloud data of the pedestrian position, and locate the position of the worker in the map, wherein the pedestrian perception module is a time-of-flight ranging sensor;

[0032] S32: Use the industrial cameras carried by the front and rear of the double twister inspection robot to run the target detection program of the YOLO algorithm to detect pedestrians on the 2D images obtained by the industrial cameras, and judge whether the front is a static obstacle or a dynamic pedestrian.

[0033] In step S4, the physical model of the pedestrian is established, and the specific steps of constructing the value calculation network are as follows:

[0034] S41: Model the surrounding pedestrians, and the personal space of the pedestrian is composed of an inner circle and a 1 / 2 ellipse, wherein the inner circle represents the safety space of the pedestrian, and the 1 / 2 ellipse represents the comfort space of the pedestrian;

[0035] a2=[a front2 ,a back2 ] T b2=[b left2 ,b right2 ] T

[0036] a p =b p =[r2+c i ,r2+c i ] T

[0037] a2=[a front2 ,a back2 ] T =a p +V h [k2,k3] T

[0038] b2=[b left2 ,b right2 ] T =b p +V h [k4,k5] T

[0039] Wherein a2 represents the boundary of the personal space of the pedestrian in the front-back direction; b2 represents the boundary of the personal space of the pedestrian in the left-right direction; a p and b p represent the initial personal space boundary of the pedestrian in the front-back direction and the initial personal space boundary of the pedestrian in the left-right direction, respectively; r2 is the radius of the inner circle; c i is a constant representing the expansion amount of the personal space; V hIt is a weight vector, where k2 and k3 are parameters related to the pedestrian's forward and backward directions, respectively; k4 and k5 are parameters related to the pedestrian's left and right directions, respectively; a front2 ,a back2 b left2 ,b right2 These represent the tense spatial boundaries of pedestrians in front, behind, left, and right.

[0040] S42: Interaction between the twisting machine inspection robot and pedestrians: If the comfort spaces of the twisting machine inspection robot and pedestrians overlap but no collision occurs, then the comfort spaces of both the twisting machine inspection robot and the pedestrians are corrected. V1 and V2 are the movement velocities of the two pedestrians, and ∥O1T∥ and ∥O2T∥ can be calculated using the following formulas:

[0041]

[0042] ∥O2T∥=∥O1O2∥-∥O1T∥

[0043] Where V0 is the combined velocity of the robot and the pedestrian, ∥O1T∥ is the robot's velocity, ∥O2T∥ is the pedestrian's velocity, ∥O1O2∥ is the distance between them, T represents the target point position, and O1 and O2 represent the current positions of the robot and the pedestrian, respectively.

[0044] S43: Using an MLP (Multilayer Perceptron) with a ReLU activation function, denoted as φ e Learn the embedding representation of each human-machine state and generate n fixed-length tensors e. i Let i = 1 to n, where i represents the number of pedestrians and n is the total number of pedestrians; then, these embedding tensors are input into another tensor denoted as ρ. f In MLP, to obtain human-computer interaction features for f i , i = 1 to n; at the same time, the embedding tensor is also input to another denoted as In the MLP, to calculate each person's attention score. The attention score for i = 1 to n This reflects the relative importance of pedestrians to the twisting machine inspection robot; then, attention scores are used to... Using the weights, a linear weighted sum is performed on the human-computer interaction feature pairs to obtain the crowd feature C. t ;

[0045]

[0046] S44: Considering personnel comfort, social pressure is introduced. The social pressure exerted on pedestrians by the twisting machine inspection robot is denoted as I1, which is directly calculated from the degree of compression deformation in the pedestrian's tense space:

[0047]

[0048] where S T is the initial personal space of the pedestrian, S' T is the transformed personal space of the pedestrian; when the personal space of the pedestrian is not squeezed, the front and rear personal spaces do not deform, i.e., S' T =S T , then I1 is 0 at this time; I1 represents the degree of squeezing deformation, and the value range of I1 is [0, 1);

[0049] S45: the social pressure applied by the doubling and twisting inspection robot to a single pedestrian is recorded as I2, and the state of interaction between the pedestrian and the doubling and twisting inspection robot is recorded as Φ; when the social pressure on the pedestrian only comes from the doubling and twisting inspection robot, Φ = 1 is recorded, and I1 = I2 at this time; when the personal space of the doubling and twisting inspection robot does not affect the pedestrian, Φ = 0 is recorded, and I2 = 0 at this time; when the pedestrian is subjected to multiple social pressures from the doubling and twisting inspection robot and other pedestrians, Φ = 2 is recorded; the I2 formula of the three interactive state conditions between the pedestrian and the doubling and twisting inspection robot is as follows:

[0050]

[0051] where d r is the distance between the doubling and twisting inspection robot and the current target pedestrian, is the distance between other pedestrians 1, 2, …, n and the current target pedestrian, and the value range of I2 is [0, 1); the smaller the distance between the pedestrian and the doubling and twisting inspection robot, the greater the social pressure applied by the doubling and twisting inspection robot to the pedestrian, and the lowest comfort of the personnel at this time;

[0052] S46: the social pressure applied by the doubling and twisting inspection robot to the pedestrian group is recorded as I3:

[0053]

[0054] where d is the distance between the doubling and twisting inspection robot and each pedestrian around it, is the social pressure applied by the doubling and twisting inspection robot to each pedestrian;

[0055] I3 calculated reflects the comfort of the personnel, and the smaller the value, the better the comfort of the personnel;

[0056] S47: the personal space and the comfort demand of the personnel are introduced, and the personnel comfort I3 is combined with the SARL* algorithm as part of the environmental state input and the reward function.

[0057] In step S5, the window path generation includes the following steps:

[0058] S51: Extract discrete points from the inspection path planned by the twister inspection robot at a preset interval; discretize the continuous global path for subsequent processing; calculate the distance between each discrete point and the edge of the narrowest channel, and put these discrete points into a small top heap priority queue;

[0059] S52: Dequeue discrete points from the priority queue according to priority, and perform collision detection. Connect the discrete points with the current position of the twister inspection robot, and if the connection does not pass through the obstacles in the map, it is considered to pass the collision detection. Find a discrete point that passes the collision detection and use it as the window path.

[0060] The network model framework combined with the comfort of personnel is divided into four parts: feature extraction, social stress feature processing, pedestrian attention score acquisition, and value function estimation. Step S6 is mainly based on the policy network and value network of deep reinforcement learning, including the following steps:

[0061] S61: According to the laser radar data, the moving direction, moving speed, position of the pedestrian relative to the twister inspection robot, and the position of the window path relative to the twister inspection robot are measured and input into the policy network. The linear velocity and angular velocity output by the policy network are used as actions executed by the twister inspection robot;

[0062] S62: Introduce social stress feature processing to the value network;

[0063] The social stress exerted by the twister inspection robot on a certain i pedestrian is part of the input state; the social stress exerted by the twister inspection robot on each pedestrian is processed by the MLP with ReLU activation function described in step S43; and p i is obtained through the Sigmoid function i ; p i is used as the weight of each human-machine interaction pair, and is used for weighted processing of the features of human-machine interaction after the introduction of social stress;

[0064]

[0065] h i = p i

[0066] where f s is a multi-layer perceptron (MLP) that extracts features of social stress ; w s is the MLP fs learnable weights for adjusting the process of feature extraction; h i is the weighted fusion feature, representing the human-robot interaction feature after introducing social pressure; e i is the tensor mentioned in step S43, representing the original feature of the ith human-robot interaction;

[0067] S63: Pedestrian attention score acquisition:

[0068] After introducing the social pressure of the inspection robot of the double-twisting machine in the obstacle avoidance navigation process, the feature is extracted, and p i is obtained through the Sigmoid function. i p i is used as the weight of each human-robot interaction pair, and the feature e i of the human-robot interaction pair after introducing social pressure is weighted to obtain the fusion feature h i . The fusion feature h α is used as the input of the function to obtain the attention score of each pedestrian The attention score of a certain pedestrian represents the relative importance of the pedestrian to the inspection robot of the double-twisting machine.

[0069]

[0070] where w α is the weight matrix in the neural network, used for linear transformation of the input feature when calculating the attention score. Through the training process, the function can more accurately reflect the relative importance of pedestrians to the inspection robot of the double-twisting machine.

[0071] The attention scores of these pedestrians are used as weights, and each pair of human-robot interaction features is linearly weighted and summed to obtain the crowd-robot interaction feature c t .

[0072]

[0073] S64: After obtaining the crowd-robot interaction feature c t , the value of the crowd-robot interaction feature pair is calculated as an estimate of the value function V:

[0074] V = f v (s t ,c t ; w v )

[0075] where f v represents an MLP function with ReLU activation function, and w v is the weight parameter of the MLP network.

[0076] The reward function of deep reinforcement learning is:

[0077]

[0078] where d min represents the minimum Euclidean distance between the twister inspection robot and the pedestrian, p t represents the position of the twister inspection robot at time t, and p g represents the position of the navigation window target; o.w. means otherwise; For deep reinforcement learning, the state action a t The resulting immediate reward, represents the joint state of the robot and the pedestrian;

[0079] S65: When the surrounding pedestrians meet the twister inspection robot, or the comfort space is in contact with each other, the twister inspection robot executes the corresponding obstacle avoidance decision.

[0080] In step S8, the real-time position coordinates of the twister inspection robot in the world coordinate system are subscribed and compared with the end point coordinates in the world coordinate system. If the distance between the two coordinates is less than or equal to a threshold value, it means that it has arrived.

[0081] The beneficial effects of the present application are that the twister inspection path dynamic planning method considering personnel comfort of the present application, based on the tension space modeling, takes the personnel comfort as part of the input state and processes the social stress feature as a weight after the fusion processing of the state of the inspection robot and the pedestrian to calculate the attention score of the pedestrian to represent the relative importance of each pedestrian to the inspection robot; the personnel comfort is taken as part of the reward function, so that the value network is more accurate in evaluating the value of the current state, so that the value network will choose more social actions during the training process, improve the efficiency and success rate during navigation, and make the decision of the inspection robot better meet the comfort demand of the pedestrians and more social. BRIEF DESCRIPTION OF DRAWINGS

[0082] The present application will be further described below in conjunction with the drawings and examples.

[0083] Figure 1 is the flowchart of the twister inspection path dynamic planning method considering personnel comfort of the present application.

[0084] Figure 2 is the modeling schematic diagram of a single robot in the present application.

[0085] Figure 3 is the modeling schematic diagram of a single worker in the present application.

[0086] Figure 4 is a modeling schematic diagram of the encounter between pedestrians and robots in the present application.

[0087] Figure 5 is a function graph of the MLP with ReLU activation function in the present application.

[0088] Figure 6 is a specific example graph of window path selection in the present application.

[0089] Figure 7 is a network model graph of the improved SARL* algorithm in the present application. DETAILED DESCRIPTION

[0090] The present application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which only show the basic structure of the present application in a schematic manner, and thus only show the components relevant to the present application.

[0091] A personnel comfort-considered inspection path dynamic programming method for a double twister, Figure 1 as shown, comprising the following steps:

[0092] S1: model a map of the inspection environment by laser radar, initialize an obstacle avoidance space model with the double twister inspection robot as the center, and receive end point information;

[0093] S2: generate an inspection path according to the current position of the double twister inspection robot and the end point information;

[0094] S3: the double twister inspection robot detects pedestrians near the production line and obtains the pose information of the pedestrians, including the moving direction, the moving speed, and the position; the specific range of the pedestrians near the production line can be adjusted and selected according to the actual environment.

[0095] S4: model the pedestrians according to the pose information of the pedestrians, judge by combining the obstacle avoidance space model of the double twister inspection robot and the physical model of the pedestrians, and add the personnel comfort as a weight into the calculation network;

[0096] S5: generate a window path according to the obstacle avoidance space and the inspection path;

[0097] S6: identify the pedestrians by the industrial camera loaded on the double twister inspection robot, measure the moving direction, the moving speed, the position of the pedestrians relative to the double twister inspection robot, and the position of the window path relative to the double twister inspection robot, input them into the trained deep reinforcement learning model, make different decisions under different circumstances, and output the linear speed and angular speed for controlling the double twister inspection robot;

[0098] S7: The doubling machine inspection robot navigates forward based on linear velocity and angular velocity;

[0099] S8: Determine whether the doubling machine inspection robot has reached the destination. If it has, navigation ends. If it has not, return to step S2. Since the doubling machine inspection robot may deviate from the inspection path in step S2 or the path planning algorithm can only plan the current path, if the doubling machine inspection robot has not reached the destination, return to step S2 to replan the path.

[0100] In step S1, the obstacle avoidance space model adopts environment modeling based on the stress space, including the following steps:

[0101] S11: The tension space of the twisting machine inspection robot can be represented by a circle and a semi-ellipse. The circle represents the safety space of the twisting machine inspection robot, and the semi-ellipse represents the comfort space of the twisting machine inspection robot.

[0102] S12: The radius of the circle and the minor axis radius b1 of the semi-ellipse are both equal to the radius r1 of the twisting machine inspection robot, and the major axis radius a of the semi-ellipse is... front1 The speed v of the inspection robot for the twisting machine t Related, such as Figure 2 As shown, b1 = r1, and k1 is a parameter related to the front of the doubling machine inspection robot.

[0103] In step S2, the inspection path is autonomously planned in real time using the dynamic programming SRAL* algorithm. The twisting machine inspection robot can flexibly handle various actual working conditions such as employee loading and unloading and temporary equipment shutdowns, ensuring real-time optimization and efficient adjustment of the inspection path. The traditional SRAL* algorithm, as a known existing algorithm, has certain limitations. The modeling is relatively simple, and it can only use a fixed semicircle to represent pedestrians and moving robots, without considering the dynamic interaction between humans and robots. This invention introduces the factor of human comfort, extending the SRAL* algorithm a step further. The basic obstacle avoidance framework of the twisting machine inspection robot still uses the SRAL* algorithm, but a comfort-related penalty term is added to the reward function. When the twisting machine inspection robot encounters a person who is too close, the cost is increased; or the weight of social attention is adjusted so that the robot is more inclined to choose a comfortable path when planning the path. In addition, dynamic interaction may need to be considered, such as how the robot should adjust the path to maintain a comfortable distance when a pedestrian moves, rather than just static obstacle avoidance. When personnel are loading materials behind the inspection robot of the twisting machine, the robot identifies moving pedestrians through the industrial camera mounted on the back and incorporates personnel comfort into its obstacle avoidance strategy. Based on the speed of the worker's movement, the robot dynamically adjusts its own speed to maintain a comfortable distance that does not affect the worker's work.

[0104] In step S3, the pedestrian detection and the acquisition of the pose information include the following steps:

[0105] S31: Obtain the moving speed of the pedestrian through a pedestrian perception module, generate point cloud data of the position of the pedestrian, and locate the position of the worker in the map, wherein the pedestrian perception module is a time-of-flight ranging sensor, the time-of-flight ranging sensor can perceive moving objects within a certain distance within a certain range, and the moving speed of the worker is obtained by combining the time of laser reflection back from the tof (time of flight) laser with the speed of the laser, point cloud data is generated by laser radar scanning, and the position of the worker is located in the modeled map;

[0106] S32: Use the industrial cameras carried on the front and rear of the twister inspection robot to run the target detection program of the yolo algorithm to detect the pedestrians on the 2D images obtained by the industrial cameras, and determine whether the static obstacle in front is a dynamic pedestrian. After the pedestrian perception module detects a moving object, the industrial cameras carried on the front and rear of the twister inspection robot capture the image, and the yolo pedestrian recognition algorithm is run to determine whether the moving object is a pedestrian or a dynamic object.

[0107] In step S4, the physical model of the pedestrian is established, and the construction of the value calculation network is as follows:

[0108] S41: Model the surrounding pedestrians, as shown in Figure 3 , the personal space of the pedestrian is composed of an inner circle and a 1 / 2 ellipse, the inner circle represents the safety space of the pedestrian, and the 1 / 2 ellipse represents the comfort space of the pedestrian;

[0109] a2=[a front2 ,a back2 ] T ,b2=[b left2 ,b right2 ] T

[0110] a p =b p =[r2+c i ,r2+c i ] T

[0111] a2=[a front2 ,a back2 ] T =a p +V h [k2,k3] T

[0112] b2=[b left2 ,b right2 ] T =b p+V h [k4,k5] T

[0113] where a2 represents the tight space boundary of the pedestrian in the front-back direction; b2 represents the tight space boundary of the pedestrian in the left-right direction; a p and b p represent the initial tight space boundary of the pedestrian in the front-back direction and the initial tight space boundary of the pedestrian in the left-right direction, respectively; r2 is the radius of the inner circle; c i is a constant representing the expansion amount of the tight space; V h is a weight vector, k2 and k3 are parameters related to the front-back direction of the pedestrian, respectively; k4 and k5 are parameters related to the left-right direction of the pedestrian, respectively; a frint2 ,a back2 , B left2 ,b right2 represent the tight space boundaries of the pedestrian in the front-back and left-right directions, respectively.

[0114] S42: Interaction between the double-twisting machine inspection robot and the pedestrian: If the comfort space between the double-twisting machine inspection robot and the pedestrian has mutual overlapping space but no collision occurs, the comfort spaces of the double-twisting machine inspection robot and the pedestrian are both corrected, as shown in Figure 4 , V1 and V2 are the motion speeds of the two pedestrians, and ∥O1T∥ and ∥O2T∥ can be calculated by the following formula:

[0115]

[0116] ∥O2T∥ = ∥O1O2∥ - ∥O1T∥

[0117] where V0 is the combined speed of the robot and the pedestrian, ∥O1T∥ is the motion speed of the robot, ∥O2T∥ is the motion speed of the pedestrian, ∥O1O2∥ is the distance between the two, T represents the target point position, O1 and O2 represent the current positions of the robot and the pedestrian, respectively;

[0118] S43: Using an MLP with a ReLU activation function as shown in Figure 5 , denoted as φ e , to learn the embedding representation of each human-robot state, generating n fixed-length tensors e i , i = 1 to n, i represents the number of pedestrians, and n is the total number of pedestrians; then, these embedding tensors are input into another MLP denoted as ρ f to obtain the human-robot interaction feature pair f i , i = 1 to n; at the same time, the embedding tensors are also input into another MLP denoted as to calculate the attention score of each person , i = 1 to n, which is reflects the relative importance of the pedestrians to the twister inspection robot; then, the attention score is taken as the weight, and the human-robot interaction feature pairs are linearly weighted and summed to obtain the crowd feature C t ;

[0119]

[0120] S44: considering the comfort of the personnel, social pressure is introduced, and the social pressure of the pedestrian on the twister inspection robot is denoted as I1, which is directly calculated by the degree of compression deformation of the tension space of the pedestrian:

[0121]

[0122] where S T is the initial tension space of the pedestrian, and S′ T is the transformed tension space of the pedestrian; when the tension space of the pedestrian is not compressed, the front and rear tension spaces do not deform, i.e., S′ T =S T , then I1 is 0 at this time; I1 represents the degree of compression deformation, and the value range of I1 is [0, 1), this variable can be used to describe the distance between the twister inspection robot and the pedestrian;

[0123] S45: the social pressure of the single pedestrian on the twister inspection robot is denoted as I2, and the state of the interaction between the pedestrian and the twister inspection robot is denoted as Φ; when the social pressure on the pedestrian only comes from the twister inspection robot, Φ=1 is recorded, and I1=I2 at this time; when the tension space of the twister inspection robot does not affect the pedestrian, Φ=0 is recorded, and I2=0 at this time; when the pedestrian is simultaneously subjected to multiple social pressures from the twister inspection robot and other pedestrians, Φ=2 is recorded; the I2 formula of the three interactive state conditions between the pedestrian and the twister inspection robot is as follows:

[0124]

[0125] where d r is the distance between the twister inspection robot and the current target pedestrian, is the distance between the other pedestrians 1, 2, …, n and the current target pedestrian, and the value range of I2 is [0, 1); the smaller the distance between the pedestrian and the twister inspection robot, the greater the social pressure exerted by the twister inspection robot on the pedestrian, and the lowest comfort of the personnel at this time; this algorithm not only considers the relationship between the twister inspection robot and the pedestrian, but also considers the relationship between the pedestrians.

[0126] S46: The social pressure exerted by the twister inspection robot on the crowd of pedestrians is recorded as I3:

[0127]

[0128] wherein is the distance between the twister inspection robot and each pedestrian around it, is the social pressure exerted by the twister inspection robot on each pedestrian;

[0129] The calculated I3 reflects the comfort of the personnel, and the smaller the value, the better the comfort of the personnel;

[0130] S47: Introduce the tension space and the comfort demand of the personnel, combine the personnel comfort I3 as part of the environmental state input and the reward function with the SARL* algorithm, add the personnel comfort as an influencing factor on the basis of the SARL* algorithm, and add a comfort-related penalty term in the reward function. When the social pressure I3 calculated through step S46 gradually increases, the personnel comfort decreases, the navigation and obstacle avoidance rate is gradually changed, and finally it is mapped to the speed of the twister inspection robot.

[0131] In step S5, generating the window path includes the following steps:

[0132] S51: Extract discrete points from the inspection path planned by the twister inspection robot at a preset interval of 0.1 meters; discretize the continuous global path for subsequent processing; calculate the distance between each discrete point and the edge of the nearest narrow passage, and load these discrete points into a small top heap priority queue; in a 0.6-meter narrow passage, the small top heap priority queue can efficiently screen the discrete points closest to the edge of the passage, thereby providing the twister inspection robot with the optimal path selection. This data structure can significantly improve the efficiency of path planning, especially in narrow spaces.

[0133] S52: Dequeue the discrete points from the priority queue according to the priority, and perform collision detection. Connect the discrete points with the current position of the twister inspection robot, and if the connection does not pass through the obstacles in the map, it is considered to pass the collision detection; find a discrete point that passes the collision detection and use it as the window path.

[0134] In a 0.6-meter narrow passage, collision detection is a key step to ensure the safety of the twister inspection robot. By selecting the discrete point closest to the edge of the passage as the window path, the twister inspection robot can efficiently navigate in narrow spaces, reduce the tortuosity of the path, and avoid collisions with obstacles. Figure 6

[0135] ​Step S6 is mainly based on two parts of policy network and value network of deep reinforcement learning, including the following steps:

[0136] S61: According to the laser radar data, the moving direction, moving speed, position of the pedestrian relative to the twister inspection robot, and the position of the window path relative to the twister inspection robot are measured and input to the policy network. The linear velocity and angular velocity output by the policy network are executed by the twister inspection robot as actions;

[0137] S62: Social pressure feature processing is introduced into the value network;

[0138] In order to better meet the social needs of safety and comfort of pedestrians, the stress space and the comfort degree of personnel are introduced, and the social pressure exerted by the twister inspection robot on a certain i pedestrian is part of the input state; the social pressure exerted by the twister inspection robot on each pedestrian is processed by the MLP with ReLU activation function mentioned in step S43; and p is obtained through the Sigmoid function; i p i is used as the weight of each human-machine interaction pair, and the feature of human-machine interaction after introducing social pressure is weighted and processed;

[0139]

[0140] h i = p i · e i

[0141] where f s is a multi-layer perceptron (MLP) that extracts features of social pressure ; w s is the learnable weight of MLP f s , used to adjust the feature extraction process; h i is the weighted fusion feature, representing the human-machine interaction feature after introducing social pressure; e i is the tensor mentioned in step S43, representing the original feature of the i-th human-machine interaction;

[0142] S63: Pedestrian attention score acquisition:

[0143] The social pressure exerted by the twister inspection robot on each pedestrian in the obstacle avoidance navigation process is introduced, processed after feature extraction, and p i is obtained through the Sigmoid function; p i is used as the weight of each human-machine interaction pair, and the feature e iThe fusion feature h is obtained by weighting. i The fusion feature h i As input to the function, obtain the attention score for each pedestrian. The attention score of a pedestrian indicates the relative importance of that pedestrian to the twisting machine inspection robot;

[0144]

[0145] Where w α This is the weight matrix in a neural network, used to linearly transform the input features when calculating the attention score. Through the training process, the function... It can more accurately reflect the relative importance of pedestrians to the twisting machine inspection robot.

[0146] This invention utilizes a multilayer perceptron (MLP) with ReLU activation to handle complex input states. Unlike traditional methods that only consider the basic states of the twisting machine inspection robot and pedestrians, this invention incorporates the social pressure exerted by the twisting machine inspection robot on each pedestrian in a human-robot coexistence environment as a crucial component of the input features. By precisely calculating the attention score of each pedestrian and combining these scores with each pair of human-robot interaction features through linear weighted summation, a more refined set of crowd-robot interaction features is successfully obtained. This method enables the twisting machine inspection robot to more intelligently weigh various factors, especially the comfort needs of pedestrians, when making obstacle avoidance decisions. Compared to traditional methods, the proposed method significantly improves obstacle avoidance performance in dynamic crowd environments while better meeting the comfort needs of pedestrians, demonstrating a higher level of intelligence and stronger environmental adaptability.

[0147] Using the attention scores of these pedestrians as weights, a linear weighted sum is performed with each pair of human-computer interaction feature pairs to obtain the crowd-robot interaction feature c. t ;

[0148]

[0149] S64: Obtaining interaction features between the crowd and the twisting machine inspection robot t Then, the value of the crowd-robot interaction feature pairs is calculated as an estimate of the value function V:

[0150] V = f v (s t ,c t ;w v )

[0151] Among them, f v w represents an MLP function with ReLU activation. v These are the weight parameters of the MLP network.

[0152] The ultimate goal of reinforcement learning is to maximize the cumulative expected reward, which requires the state value function to be optimized. When the state value function is optimal, the corresponding policy is also the optimal policy. This policy can maximize the cumulative expected reward starting from the initial state. To achieve this goal, a deep reinforcement learning network is used to decide the best action for the robot based on the current state (a state representation that integrates information from multiple sensors).

[0153] The network parameters are updated through the temporal difference method, which minimizes the temporal difference error between the current value network and the target value network to optimize the policy. To improve data usage efficiency, the algorithm uses a priority experience replay strategy to sample some high-priority samples from the stored previous state-action pairs and their corresponding rewards and next states. This process aims to meet various needs in the robot navigation process, such as accurately reaching the target location, avoiding collisions, and reasonably meeting the comfort needs of the crowd.

[0154] The reward function plays a crucial role in deep reinforcement learning algorithms, serving as a key bridge between the target task and the reinforcement learning algorithm. For the task of obstacle avoidance and navigation of the double-twisting machine inspection robot in a dynamic crowd environment, the design of the reward function is particularly important. Traditional reward functions mainly focus on collision avoidance and reaching the target location during navigation, but often overlook the social comfort needs of pedestrians. To achieve human-robot coexistence, when designing the reward function, in addition to considering traditional obstacle avoidance requirements (such as avoiding collisions and accurately reaching the target location), the social stress experienced by the pedestrian group should also be included in the penalty term of the reward function. This not only ensures the safety of the double-twisting machine inspection robot during navigation, but also improves the comfort of pedestrians, achieving harmonious interaction between the robot and pedestrians. Therefore, when using deep reinforcement learning to solve such tasks, it is crucial to carefully design the reward function to consider navigation goals, safety, and pedestrian comfort. In the obstacle avoidance and navigation task of the double-twisting machine inspection robot, the design of the reward function is crucial.

[0155] The reward function for deep reinforcement learning is:

[0156]

[0157] where d min represents the minimum Euclidean distance between the double-twisting machine inspection robot and the pedestrian, p t represents the position of the double-twisting machine inspection robot at time t, and p g represents the position of the navigation window target, which refers to the window endpoint and the target endpoint after each execution of the strategy; o.w. means "otherwise".

[0158] For deep reinforcement learning, evaluate the state Take action a t The resulting immediate reward, Indicates the joint state of the robot and the pedestrian;

[0159] In order to ensure that the double twister inspection robot does not collide with pedestrians during navigation, when a collision occurs, a fixed penalty value such as -0.5 can be given to the corresponding action during reinforcement learning training as a punishment for deviating from the target task. Conversely, when the double twister inspection robot successfully reaches the target location, a positive fixed value reward can be given.

[0160] Network model combined with personnel comfort:

[0161] The network structure model framework can be divided into four parts: feature extraction part, social stress feature processing, pedestrian attention score acquisition and value function estimation. The overall structure of the network structure model is as shown in Figure 7 The state of the double twister inspection robot and the dynamic pedestrian group obstacle is divided, and the fusion information of the double twister inspection robot not only contains its own state, but also contains the state of the surrounding dynamic pedestrians. In this way, the interaction between the double twister inspection robot and the dynamic pedestrian group can be simulated.

[0162] S65: When the surrounding pedestrians meet the double twister inspection robot, that is, the pedestrian perception module detects dynamic obstacles, the industrial camera recognizes pedestrians, or the comfort space is in contact with each other, the double twister inspection robot will execute the corresponding obstacle avoidance decision;

[0163] When the double twister inspection robot encounters pedestrians in a narrow space, the principle of giving way to pedestrians by the double twister inspection robot is executed, and the double twister inspection robot will stop when it encounters pedestrians during inspection, so as to facilitate the detection of abnormal working conditions by the staff.

[0164] In step S8, the real-time position coordinates of the double twister inspection robot in the world coordinate system are subscribed through ROS (Robot Operating System), and compared with the end point coordinates in the world coordinate system. If the distance between the two coordinates is less than or equal to a threshold value, it indicates that it has arrived.

[0165] The personnel comfort considering double twister machine inspection path dynamic programming method of the application is deployed on a double twister inspection robot, the double twister inspection robot has a chassis, a high-definition industrial camera, a laser radar and a processing unit, the chassis adopts an omni-directional wheel design to ensure flexible movement, the high-definition industrial camera is installed in multiple directions of the robot, and the laser radar is used for accurate ranging and environment modeling.

[0166] The personnel comfort considering double twister machine inspection path dynamic programming method of the application is deployed on a double twister inspection robot, has the following advantages:

[0167] High navigation success rate: for different pedestrian complexity scenes in the simulation environment, the ratio of the inspection robot successfully and safely reaching the target point position is high.

[0168] Low collision rate: when the double twister machine inspection robot passes through a narrow lane of about 0.6m, it can detect the worker and stop in time, which is convenient for the worker to check the abnormal working condition.

[0169] Good trajectory smoothness: qualitative analysis is performed on the smoothness of the motion trajectory of the inspection robot in the obstacle avoidance navigation process, the motion trajectory is very smooth, and the probability of motion freezing of the robot in the navigation process is small.

[0170] High comfort: the personnel comfort is introduced as a reward function into the SRAL* obstacle avoidance navigation algorithm, so that the evaluation of the value network on the current state value is more accurate, the value network will choose more social actions in the training process, which improves the efficiency and success rate of navigation, and the decision of the double twister machine inspection robot can better meet the comfort needs of the workers and is more social.

[0171] Based on the above ideal embodiments according to the application, through the above description, relevant personnel can make various changes and modifications without deviating from the technical idea of the application. The technical scope of the application is not limited to the contents in the specification, and must be determined according to the scope of claims.

Claims

1. A method for dynamic planning of inspection path of a two-for-one twister considering comfort of personnel, characterized in that, The method comprises the following steps: S1: modeling the inspection environment by laser radar to generate a map, initializing a tight space-based obstacle avoidance space model centered on the twister inspection robot, and receiving end point information; S2: generating an inspection path according to the current position of the twister inspection robot and the end point information; S3: the twister inspection robot detects the pedestrians near the production line and obtains the pose information of the pedestrians, including the moving direction, the moving speed and the position; S4: modeling the pedestrians based on the tight space according to the pose information of the pedestrians, combining the twister inspection robot obstacle avoidance space model and the physical model of the pedestrians to make a judgment, and calculating the social pressure and taking the comfort of the personnel as a weight to join the calculation network; S5: generating a window path according to the obstacle avoidance space and the inspection path; S6: identifying the pedestrians by the industrial camera loaded on the twister inspection robot, combining the laser radar data to measure the moving direction, the moving speed, the position of the pedestrians relative to the twister inspection robot and the position of the window path relative to the twister inspection robot, inputting them into the trained deep reinforcement learning model, and outputting the linear velocity and angular velocity for controlling the twister inspection robot, wherein the reward function of the deep reinforcement learning contains the comfort of the personnel based on the social pressure; S7: the twister inspection robot navigates forward according to the linear velocity and angular velocity; S8: judging whether the twister inspection robot reaches the end point, if yes, ending the navigation, and if not, returning to step S2.

2. The method of claim 1, wherein the method is characterized by, In step S1, the obstacle avoidance space model adopts the environment modeling based on the tight space, which comprises the following steps: S11: the tight space of the twister inspection robot can be represented as a circle and a half-ellipse, wherein the circle represents the safety space of the twister inspection robot, and the half-ellipse represents the comfort space of the twister inspection robot; S12: radius of the circle and the minor radius of the semi-ellipse are equal to the radius of the doubling inspection robot , the major radius of the semi-ellipse and the speed of the doubling inspection robot are related, , is a parameter related to the front of the doubling inspection robot.

3. The method of claim 1, wherein the method is characterized by, In step S2, the inspection path is autonomously planned in real time by dynamic programming SRAL algorithm.

4. The method of claim 1, wherein the method is characterized by, In step S3, the pedestrian detection and the obtaining of the pose information comprise the following steps: S31: obtaining the moving speed of the pedestrian by a pedestrian perception module, generating the point cloud data of the position of the pedestrian, and positioning the position of the worker in the map, wherein the pedestrian perception module is a time-of-flight distance sensor; S32: using the industrial cameras loaded on the front and rear of the twister inspection robot, running a target detection program of a yolo algorithm to detect the pedestrians on the 2D images obtained by the industrial cameras, and judging whether the front is a static obstacle or a dynamic pedestrian.

5. The method of claim 1, wherein the method is characterized by, In step S4, the specific steps of establishing the physical model of the pedestrian and constructing the value calculation network are as follows: S41: modeling the surrounding pedestrians, wherein the tight space of the pedestrian is composed of an inner circle and a 1 / 2 ellipse, the inner circle represents the safety space of the pedestrian, and the 1 / 2 ellipse represents the comfort space of the pedestrian; wherein denotes the tension space boundary in the front-back direction of the pedestrian; denotes the tension space boundary in the left-right direction of the pedestrian; denote the initial tension space boundary in the front-back direction and the initial tension space boundary in the left-right direction of the pedestrian, respectively; r 2 is the radius of the inner circle; c i is a constant representing the amount of expansion of the tension space; is a weight vector, and are parameters related to the front-back direction of the pedestrian, respectively; and are parameters related to the left-right direction of the pedestrian, respectively; denote the tension space boundaries in the front-back and left-right directions of the pedestrian, respectively; S42: Interaction between the twister inspection robot and the pedestrian: If the comfort space between the twister inspection robot and the pedestrian has a mutual overlapping space but no collision occurs, the comfort space of the twister inspection robot and the pedestrian is both corrected, and is the motion speed of the two pedestrians, and can be calculated by the following formula: wherein is the combined velocity of the robot and the pedestrian, is the velocity of the robot, is the velocity of the pedestrian, is the distance between the two, T denotes the target point position, and denote the current positions of the robot and the pedestrian, respectively; S43: Using an MLP (Multilayer Perceptron) with a ReLU activation function, denoted as... Learn the embedding representation of each human-machine state and generate n fixed-length tensors. Let i = 1 to n, where i represents the number of pedestrians and n is the total number of pedestrians; then, these embedding tensors are input into another denoted as In MLP, to obtain human-computer interaction feature pairs i = 1 to n; at the same time, the embedding tensor is also input to another denoted as In the MLP, to calculate each person's attention score. For i = 1 to n, the attention score This reflects the relative importance of pedestrians to the twisting machine inspection robot; then, attention scores are used to... Using the weights, a linear weighted sum is performed on the human-computer interaction feature pairs to obtain the crowd features. ; S44: considering the comfort of the personnel, introducing social stress, the social stress of the pedestrian on the inspection robot of the doubling machine is recorded as which is directly calculated by the degree of extrusion deformation of the tension space of the pedestrian: wherein, the initial tension space of the pedestrian, the transformed tension space of the pedestrian; when the tension space of the pedestrian is not squeezed, the front and rear tension spaces do not deform, which means the tension space of the pedestrian at this time, is 0; represents the degree of squeezing deformation, the value range of is [0, 1). S45: The social pressure exerted on a single pedestrian by the twisting machine inspection robot is denoted as... Let Φ denote the interaction state between the pedestrian and the doubling machine inspection robot; when the social pressure experienced by the pedestrian comes only from the doubling machine inspection robot, let Φ = 1, and at this time... = When the confined space of the doubling machine inspection robot does not affect pedestrians, Φ is recorded as 0, and at this time... =0; when a pedestrian is simultaneously subjected to multiple social pressures from the twisting machine inspection robot and other pedestrians, Φ=2; these three interaction states between pedestrians and the twisting machine inspection robot are... The formula is as follows: wherein, is the distance between the twister inspection robot and the current target pedestrian, , , is the distance between the other pedestrian 1, 2, …, n and the current target pedestrian, The value range of is [0, 1); the smaller the distance between the pedestrian and the twister inspection robot, the greater the social pressure exerted by the twister inspection robot on the pedestrian, and at this time the comfort of the personnel is the lowest; S46: the social pressure exerted by the two-for-one twister inspection robot on the group of pedestrians is recorded as : wherein is the distance between the doubling machine inspection robot and each pedestrian in its surroundings, is the social pressure each pedestrian is subjected to by the doubling machine inspection robot. calculated The smaller the value, the better the comfort of the person. S47: Introduce stress space and personnel comfort needs, incorporate personnel comfort as part of the environmental state input and reward function in conjunction with the SARL algorithm.

6. The method of claim 1, wherein the method is characterized by, In step S5, the generation of the window path comprises the following steps: S51: extracting discrete points from the inspection path planned by the twister inspection robot at a preset interval; discretizing the continuous global path for subsequent processing; calculating the distance between each discrete point and the edge of the nearest narrow channel, and loading these discrete points into a small top heap priority queue; S52: Dequeue the discrete points from the priority queue according to the priority, and perform collision detection. Connect the discrete points with the current position of the twister inspection robot. If the connection does not pass through the obstacles in the map, it is considered to pass the collision detection. Find a discrete point that passes the collision detection and take it as the window path.

7. The method of claim 5, wherein the method further comprises: The network model framework combined with the comfort of personnel is divided into four parts: feature extraction part, social stress feature processing, pedestrian attention score acquisition and value function estimation. Step S6 is mainly based on the policy network and value network of deep reinforcement learning, including the following steps: S61: According to the laser radar data, the moving direction, moving speed, position of the pedestrian relative to the twister inspection robot, and the position of the window path relative to the twister inspection robot are measured and input into the policy network. The linear velocity and angular velocity output by the policy network are taken as actions and executed by the twister inspection robot; S62: Introduce social stress feature processing to the value network; social pressure exerted by the twister robot on a certain i-th pedestrian as part of the input state; the social pressure exerted by the twister robot on each pedestrian via the MLP with ReLU activation function described in step S43 is processed as a feature; and is obtained through a Sigmoid function , using as the weight of each human-machine interaction pair, as a weighted processing of the features of human-machine interaction after the introduction of social pressure; wherein is a multi-layer perceptron (MLP) that extracts features from the social stress ; is a learnable weight of the MLP that adjusts the process of feature extraction; is the weighted fused features that represent the human-machine interaction features with the introduction of social stress; is the tensor mentioned in step S43 that represents the original features of the i-th human-machine interaction; S63: Pedestrian attention score acquisition: The social pressure exerted by the introduced double twister inspection robot on each pedestrian in the obstacle avoidance navigation process is extracted as features, and the features are obtained through a Sigmoid function , using as the weight of each human-robot interaction pair, and the features of the human-robot interaction pair after the introduction of social pressure are weighted to obtain the fusion features , and the fusion features are input into the function to obtain the attention score of each pedestrian , i=1…n, the attention score of a certain pedestrian represents the relative importance of the double twister inspection robot to the pedestrian; wherein is a weight matrix in the neural network for linearly transforming the input features when computing the attention score; The human-robot interaction feature is obtained by linearly weighting and summing each pair of human-robot interaction features with the attention score of the pedestrians as the weight ; S64: Obtain crowd-robot interaction feature with the patrol robot of the two-for-one twister Then, calculate the value of the crowd-robot interaction feature pair as an estimate of the value function: ​ wherein, denotes an MLP function with ReLU activation function, are weight parameters of the MLP network; The reward function of deep reinforcement learning is: wherein, represents the minimum Euclidean distance between the twister inspection robot and the pedestrian, represents the position of the twister inspection robot at time t , while represents the position of the navigation window target; o . w . represents "otherwise"; is used in deep reinforcement learning to evaluate the immediate reward resulting from taking action in state , while represents the joint state of the robot and the pedestrian; S65: When the surrounding pedestrians meet the twister inspection robot, or the comfort space contacts each other, the twister inspection robot executes the corresponding obstacle avoidance decision.

8. The method of claim 1, wherein the method is characterized by, In step S8, the real-time position coordinates of the twister inspection robot in the world coordinate system are subscribed and compared with the end point coordinates in the world coordinate system. If the distance between the two coordinates is less than or equal to a threshold value, it means that it has arrived.

Images