A control method of a high-speed railway station inspection robot mechanical arm operation system

Abstract

A control method for the operating system of a high-speed railway station inspection robot's robotic arm. The control method includes: first, teaching the high-speed railway robot to perform tasks in a real-world environment through manual teaching; second, aligning and optimizing the teaching data using Dynamic Time Warping (DTW) based on multiple teaching data obtained in step S1; third, smoothing the teaching trajectory by combining Gaussian Mixture Model (GMM) and Gaussian Mixture Regression (GMR) after aligning and optimizing the teaching data using DTW in step S2; fourth, processing the optimal robotic arm stiffness and contact force data obtained in step S3 using Improved Dynamic Motion Element Methods (DMPs); and fifth, establishing a unified contact-based task learning framework based on the robotic arm stiffness data obtained in step S4, and achieving compliant control of the robotic arm for contact-based tasks through impedance control. This invention's control method enables autonomous, anthropomorphic operation during inspection tasks through simple teaching.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robot control, in particular to a control method of a mechanical arm operation system of a high-speed railway station inspection robot. BACKGROUND

[0002] The high-speed railway station bears the passengers and is an important guarantee for the safety of high-speed train operation, which contains a large number of equipment and devices, a large number of personnel, and a complex environment, and requires high inspection intensity. At present, the inspection and detection of domestic high-speed railway stations mainly relies on manual inspection, which makes simple qualitative judgment and inspection on the equipment in the station, and has the disadvantages of low intelligentization, low detection efficiency, slow response speed, and large investment. The human-robot coexistence inspection operation robot can assist the station operation personnel to complete the high-precision, high-intensity and long-time operation task, reduce the influence of unknown environment on the operation personnel, and effectively improve the operation task completion quality and efficiency, so the development of high-speed railway station human-robot coexistence inspection robot has become a research hotspot.

[0003] At present, the trajectory planning of the inspection robot mechanical arm has the problems of long programming time, poor program reusability and pure mechanical arm position control which is difficult to meet the demand of completing the contact task in a complex environment. SUMMARY

[0004] Therefore, in view of the defects in the prior art, the present application provides a control method of a mechanical arm operation system of a high-speed railway station inspection robot.

[0005] In a first aspect, the present application designs a mechanical arm operation system of a high-speed railway station inspection robot, which comprises a teaching data acquisition system, a trajectory learning system and the trajectory learning system, wherein the teaching data acquisition system is used for multiple drag teaching in a real environment, so that the mechanical arm and the end grabbing device reach the inspection operation task position flexibly, and a plurality of teaching data are acquired.

[0006] In addition, the trajectory learning system is used for dynamic time warping processing of the acquired teaching data to obtain a teaching trajectory, learning of the teaching trajectory by using a Gaussian mixture and Gaussian regression model to obtain an optimal reference trajectory, and learning of the optimal reference trajectory based on an improved dynamic motion primitive to realize the trajectory of the mechanical arm to reproduce the trajectory of the teacher to reach the target position.

[0007] In addition, the compliant control grabbing system is used for constructing a unified contact task learning framework based on the data collected in the contact task operation process performed by the teacher, and guiding the mechanical arm in the compliant control of the contact task based on impedance control.

[0008] In a second aspect, the present application proposes a control method of a mechanical arm operation system of a high-speed railway station inspection robot constructed according to the above control system, which comprises the following steps:

[0009] S1, first, through the method of manual demonstration, the high-speed rail robot in the real environment is demonstrated, and multiple dragging demonstrations are performed; wherein the mechanical arm of the inspection robot is allowed to complete the dragging demonstration and the inspection action, the end gripping device of the mechanical arm of the inspection robot is allowed to reach each task position after the dragging demonstration, and the inspection robot is allowed to complete the inspection operation, and in the process of completing the multiple dragging demonstrations, multiple demonstration data are obtained;

[0010] S2, according to the multiple demonstration data obtained in S1, the dynamic time warping DTW is used to align and optimize the demonstration data, and the method of optimization alignment is:

[0011] Optionally, the time series of two demonstration data is described as:

[0012] ;

[0013] ;

[0014] Among them, And The binary time matrix is stretched in the related vectors in the time signal And The distance matrix between each demonstration data point sequence is obtained, and the formula of the distance matrix is The size of the distance matrix is , and the is the number of sequences in the demonstration data set;

[0015] The distance matrix is expanded and a vector is obtained, the hash calculation is performed on each vector in the distance matrix, each demonstration data point sequence is mapped into the corresponding hash bucket in the hash table, and then the time sequence is converted into a hash string, the hash string is mapped into the hash bucket in the hash table, and the fast indexing of the demonstration sequence is realized;

[0016] S3, after the dynamic time warping DTW is used to align and optimize the demonstration data in S2, the Gaussian mixture model GMM and the Gaussian mixture regression GMR are combined to smooth the demonstration trajectory, and then the best mechanical arm stiffness data and the contact force data of the mechanical arm end gripping device are obtained, and the best stiffness data and the contact force data are extracted;

[0017] S4, according to the best mechanical arm stiffness data and the contact force data of the inspection robot obtained in S3, the improved dynamic motion primitive DMPs is used for processing, and then the best reference trajectory of the best inspection robot mechanical arm is obtained, the trajectory generalization of the inspection robot is performed according to the best reference trajectory, and the robot obtains the learning trajectory.

[0018] S5, establishing a unified contact task learning framework according to the robot arm rigidity data and the robot arm end grabbing device contact force data obtained in S4, then realizing the robot arm contact task compliant control through impedance control.

[0019] Further, in the step S1, the teaching trajectory is obtained by manually dragging the teaching multiple times The teaching trajectory is represented as , ;

[0020] Wherein represents the length of the first trajectory, represents dimensional input time, represents dimensional trajectory variable; the pose of the robot arm end relative to the base coordinate system is represented as , .

[0021] Further, in the step S2, in the hash bucket, all sequences with the same hash value are found according to the fast index, the DTW distance of the sequences with the same hash value is calculated, and the sequence of the teaching data points with the smallest distance is selected as the most similar sequence of teaching data points.

[0022] Further, in the step S3, the joint probability distribution of the input variable and the output variable in the teaching trajectory is modeled:

[0023] ;

[0024] Wherein, ; is the number of Gaussian components in the GMM, represents the prior probability of the first Gaussian component, represents the mean vector of the probability density function of the first Gaussian component, represents the covariance matrix of the first Gaussian component.

[0025] Further, in the step S3, after obtaining the GMM parameters, for any new input , the conditional probability distribution model of the corresponding trajectory is predicted by using GMR:

[0026] ;

[0027] Furthermore, in step S4, given the processed optimal reference trajectory... , Starting point Using local weighted regression as the target point, the basis function is obtained. Weighting coefficients This leads to a new learning trajectory. With teaching trajectory They have similar motion trajectory trends, which enables the robotic arm to learn the instructor's teaching trajectory, and finally allow the robotic arm's end effector to reach the target position of the task by imitating the instructor.

[0028] Furthermore, in step S4, the compliant control method involves establishing a kinematic and dynamic model of the robotic arm and its environment, studying the motion and force analysis of the robotic arm during contact-based task operations, simulating the environment as a spring-damped system, and thus achieving impedance control of the robotic arm. By comparing the actual motion trajectory of the robotic arm with the taught trajectory, a deviation trajectory is obtained. This deviation trajectory is then converted by the aforementioned impedance controller to obtain a force deviation signal. This signal is accumulated to the desired contact force / torque and subtracted from the actual contact force / torque measured by the force / torque sensor at the end of the robotic arm. This signal serves as the input signal for the inner loop of the force control, thereby achieving impedance control of the robotic arm.

[0029] Compared with the prior art, the beneficial effects of this invention are as follows:

[0030] The control method of the robotic arm operating system of the high-speed railway station inspection robot of the present invention can cope with the contact tasks that may occur during the inspection operation of the human-machine collaborative inspection robot, and ensure that the robotic arm can perform complex contact tasks in the environment of high-speed railway stations with high personnel density and complex environment. In order to avoid complex trajectory planning, the robotic arm can quickly acquire motion skills through teaching and learning, effectively improving the planning and control of the robotic arm.

[0031] The control method of the robotic arm operating system of the high-speed railway station inspection robot of the present invention maps the hash string to the hash bucket in the hash table to realize the fast indexing of the teaching sequence. In the hash bucket, all sequences with the same hash value are found according to the fast index, the DTW distance of the sequences with the same hash value is calculated, and the teaching data point sequence with the smallest distance is selected as the most similar teaching data point sequence.

[0032] Firstly, it can solve the accuracy problem, that is, by aligning teaching trajectories of different lengths and deformations, improving their consistency and comparability, and by using hash encoding to achieve approximate matching and searching, thereby improving accuracy and efficiency.

[0033] In the second aspect, the reliability can be improved, i.e., different teaching trajectories can be mapped and modeled to convert them into sequence data of the same shape and length, so that the inaccuracy or loss of control of the robot arm due to the diversity and noise of the teaching data is avoided; more reliable trajectory classification and matching are achieved through fast indexing and similarity matching, so that the confusion and misjudgment between similar trajectories are avoided;

[0034] In the third aspect, the cost and complexity can be reduced, i.e., the cost and complexity of the robot arm teaching and control can be reduced in processing the teaching trajectory data, and since the trajectory alignment and feature extraction are performed, the complex process of manual labeling and data cleaning can be avoided, and the difficulty of manual sampling and feature extraction can also be avoided by mapping each teaching data point sequence into the corresponding hash bucket in the hash table for trajectory classification and matching, so as to help the user to complete the robot arm teaching and motion control more quickly.

[0035] The control method of the high-speed railway station inspection robot arm operation system of the application can, on the one hand, in the complex unstructured environment of the high-speed railway station, enable the teacher to migrate the teacher's behavior to the robot system in a relatively direct way through a small amount of programming operation process, so as to realize the spatial motion planning of the robot in a complex space; at the same time, the traditional dynamic motion primitive DMPs algorithm is improved to improve the learning efficiency of the algorithm for teaching trajectory learning, so as to realize the rapid trajectory planning of the robot arm; at the same time, the robot arm can extract the corresponding motion features from a small amount of teaching samples and generalize the features to new scenes; on the other hand, by collecting the data such as position, speed, interaction force and interaction stiffness during the execution of the task by the teacher, a unified contact task impedance control learning framework is established, and the robot arm is combined with the teaching learning to realize the self-personalized compliant control of the robot arm in the complex unstructured environment of the high-speed railway station. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 The control flowchart of the inspection robot arm operation system of the application is shown in the figure;

[0037] Figure 2 The technical roadmap of the inspection robot arm operation system of the application is shown in the figure;

[0038] Figure 3 The schematic diagram of the drag teaching data of the inspection robot arm of the application is shown in the figure;

[0039] Figure 4 The teaching trajectory learning and generalization flowchart based on the improved dynamic motion primitive DMPs of the application is shown in the figure;

[0040] Figure 5 The impedance control flowchart of the inspection robot arm operation system of the application is shown in the figure;

[0041] Figure 6A specific contact task diagram for the inspection robot mechanical arm of the present application;

[0042] Figure 7 A simulation diagram of the operation system of the inspection robot mechanical arm of the present application. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. The described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0044] Please refer to Figures 1-5 The present application provides an inspection robot mechanical arm operation system for a high-speed rail passenger station, comprising: a teaching data acquisition system, which is used for multiple drag teaching in a real environment, so that the mechanical arm and the end gripping device reach the inspection operation task position in a compliant manner, and multiple teaching data are acquired;

[0045] a trajectory learning system, which is used for dynamic time warping processing of the acquired teaching data, obtaining teaching trajectories, learning the teaching trajectories by using a Gaussian mixture and a Gaussian regression model to obtain optimal reference trajectories, then learning the optimal reference trajectories based on an improved dynamic motion primitive, so that the mechanical arm reproduces the trajectory of the teacher to reach the target position;

[0046] a compliant control gripping system, which is used for constructing a unified contact task learning framework by using the data collected in the contact task operation process performed by the teacher, and guiding the mechanical arm in compliant control of the contact task based on impedance control.

[0047] In addition, the present application also provides a control method of the inspection robot mechanical arm operation system for a high-speed rail passenger station, comprising the following steps:

[0048] S1, based on multiple drag teaching in a real environment, so that the mechanical arm and the end gripping device reach the inspection operation task position in a compliant manner, and multiple teaching data are acquired;

[0049] S2, dynamic time warping (DTW) is used to align and optimize the teaching data (DTW is a time-based trajectory warping, which assigns control points to corresponding time points, so that the two trajectories are as consistent as possible in time. Specifically, DTW can consider variable length and nonlinear shift when calculating the distance between two trajectories, and can use an optimized dynamic programming algorithm for calculation, so that multiple teaching data can be aligned in time to obtain a realistic trajectory matching result, so that the robot can learn more accurate actions in the next stage), and Gaussian mixture model (GMM) and Gaussian mixture regression (GMR) are used to smooth the teaching trajectory and extract features such as robot stiffness data and contact force data;

[0050] Optionally, the time series of two teaching data and , the length is and , one of which is used as a reference template and the other as a test template, and the above time series is described by DTW as:

[0051] ;

[0052] ;

[0053] wherein, and represent the binary time matrix, which stretches the relevant vectors in the time signals and , so that the teaching data time series matrix is aligned in time.

[0054] The joint probability distribution of the input variable and the output variable in the teaching trajectory is modeled as:

[0055] ;

[0056] wherein, , is the number of Gaussian components in GMM, represents the prior probability of the th Gaussian component, represents the mean vector of the probability density function of the th Gaussian component, represents the covariance matrix of the th Gaussian component.

[0057] After obtaining the GMM parameters, for any new input , the GMR is used to predict the conditional probability distribution model of the corresponding trajectory .

[0058] ;

[0059] Further can be expressed as:

[0060] ;

[0061] With the acquisition of the demonstration data, the present application adopts the GMM-GMR method to segmentally model and predict the demonstration data, which can divide the complex demonstration data into several basic motion patterns, and estimate a set of Gaussian distribution parameters for each pattern. Specifically, first, train the GMM model: input the feature data into the GMM model, which will train multiple Gaussian distributions, each of which represents a pattern in the feature data. The EM algorithm can be used to train the GMM model to obtain the weight, mean and covariance matrix of each Gaussian distribution; second, cluster using the GMM model: the trained GMM model can be used to cluster the feature data, and similar data points are divided into the same class. Each data point will be assigned to the Gaussian distribution that is most likely to produce the data point; then train the GMR model: the clustering results can be used as the input of the GMR model, and a GMR model is trained, which can predict the output of the given input feature data. GMR is a generative model that can model the conditional probability distribution of the output given the input data using Gaussian mixture model, and output the mean and variance of the conditional probability distribution. Then use GMR for feature extraction: input the feature data to be extracted into the trained GMR model, and output the prediction result. The output corresponding to the input data is the stiffness data and contact force data of the robot arm and other features; finally, the GMM method can be used to segmentally model the demonstration data, divide the complex trajectory into several basic motion patterns, and estimate a set of Gaussian distribution parameters for each pattern. Then, using the GMR method, the corresponding motion trajectory can be predicted by the model when a new task is performed, realizing the adaptability of the robot arm to complex tasks.

[0062] It can be understood that in another embodiment, the demonstration data of the robot arm is optimized and aligned by the dynamic time warping (DTW) method, and the distance matrix and between the related vectors in the time signal , the size of which is , wherein is the number of sequences in the demonstration data set, and then the distance matrix is unfolded into a vector of ; a set of random parameters and , and a bucket width are selected.DBH is calculated for each vector in the distance matrix,

[0063] ;

[0064] where is a query point, and are random data sampled from uniform distribution, is the distance matrix.

[0065] By projecting the query point to randomly selected rows to approximately preserve arbitrary distance metrics, each demonstration data point sequence is mapped to the corresponding hash bucket in the hash table by DBH binarization, thereby converting the time series into a hash string:

[0066]

[0067] where, is the dot product of the vector and the point of the time series .

[0068] The distance threshold is set to , the number of buckets placed in the hash table is , and the number of each hash function is . For a time series , its hash string representation is , which is mapped to the hash bucket in the hash table to achieve fast indexing of the demonstration sequence:

[0069] ;

[0070] In the hash bucket, all sequences with the same hash value are found according to the fast index, and their DTW distance is calculated, and the demonstration data point sequence with the smallest distance is selected as the most similar demonstration data point sequence:

[0071] ;

[0072] where is the distance, and are random parameters subject to distribution, is the hash bucket width.

[0073] Mapping the hash string to the hash bucket in the hash table achieves fast indexing of the demonstration sequence. In the hash bucket, all sequences with the same hash value are found according to the fast index, the DTW distance of the sequences with the same hash value is calculated, and the demonstration data point sequence with the smallest distance is selected as the most similar demonstration data point sequence.

[0074] The first aspect can solve the precision problem, that is, aligning different lengths and deformation of the teaching trajectory, improving its consistency and comparability while realizing approximate matching and search through hash coding, improving precision and efficiency.

[0075] The second aspect can improve reliability, that is, different teaching trajectories can be mapped and modeled, and converted into sequence data of the same shape and length, avoiding inaccurate or out-of-control movement of the robot arm due to the diversity and noise of the teaching data; more reliable trajectory classification and matching can be realized through fast indexing and similarity matching, avoiding confusion and misjudgment between similar trajectories.

[0076] The third aspect can reduce cost and complexity, that is, in processing teaching trajectory data, the cost and complexity of robot teaching and control can be reduced, and since the trajectory alignment and feature extraction, the complex process of manual annotation and data cleaning can be avoided, and through mapping each teaching data point sequence to the corresponding hash bucket in the hash table for trajectory classification and matching, the difficulty of manual sampling and feature extraction can also be avoided, helping users to complete robot teaching and motion control more quickly.

[0077] As shown in S3, the optimal reference trajectory is learned based on improved dynamic motion primitives (DMPs) and trajectory generalization is realized to obtain a learned trajectory. Figure 4

[0078] A trajectory with a length of is selected , and the motion teaching trajectory is coded:

[0079] ;

[0080] Among them, , respectively, represent the position, velocity, acceleration, and jerk of the teaching trajectory, is a constant, is the length of the teaching trajectory, represents the phase variable, represents the target position of the teaching trajectory, and are the preset diagonal stiffness and damping matrices of the robot arm, is the forcing function of the teaching trajectory.

[0081] In order to make the teaching trajectory closer to the expected trajectory, the above formula introduces a nonlinear forcing function , which realizes the fitting of the trajectory through the normalized linear superposition of multiple nonlinear basis functions, and is represented as follows:

[0082] ;

[0083] Among them, is​ No. One portion, For fitting The number of nonlinear basis functions required As basis functions, These are weighting coefficients. This is the initial position of the teaching trajectory.

[0084] By constructing a loss function, the model parameters of the aforementioned basis functions are learned using the local weighted regression optimization method. For each basis function... corresponding weighting coefficients ,as follows:

[0085] ;

[0086] in, This represents the total number of time steps for the entire teaching trajectory.

[0087] The improved DMPs are an enhancement of traditional DMPs, enabling more effective learning of complex motion trajectories and exhibiting stronger adaptability and generalization capabilities. The improved DMPs use a Gaussian Model (GMM) to segment and model the teaching data, then employ a feedback control mechanism for online adjustment and fine-tuning of the model, thereby achieving precise control of the robotic arm's motion. First, a combined Gaussian kernel function is used to describe the motion trajectory; the kernel function is the fundamental building block for trajectory control in the improved DMPs. The kernel function can be viewed as a transformation of the motion trajectory, converting it from one graphical format to another for control. The Gaussian kernel function can be decomposed into multiple kernels based on the shape and process of the motion trajectory, making the control of the motion trajectory more flexible and natural.

[0088] Secondly, a dynamic adaptive model is trained. This model learns several parameters, obtained through regression analysis. These parameters describe the characteristics of the motion trajectory, such as velocity, acceleration, and the influence of the environment. Then, based on a feedback control strategy, improved DMPs can calculate the robot's trajectory and execute actions at each time step to adjust different motion modes. Through this feedback control strategy, motion information such as acceleration at the target point can be calculated in real time based on the kernel weights and target state at the queried data points, thus forming a human-machine interaction process.

[0089] Finally, by optimizing the model parameters and under feedback control, the robot tracks the end-point trajectory of a series of non-periodic motion generators (DMPs) starting from the target state. Through repeated practice, the model parameters of DMPs are optimized to accurately simulate the motion trajectory of a human instructor and precisely control the robot's motion.

[0090] In summary, DMPs contain two parts: the first part is the learning process, i.e., using GMM segmentation modeling demonstration data, dividing the complex demonstration data into different motion patterns; the second part is the execution process, i.e., using feedback control-based DMPs to control the motion of the end of the robot arm to achieve precise motion control and task execution.

[0091] As shown in S4, a unified contact task learning framework is established according to the collected robot stiffness data and contact force data, and impedance control is used to guide the robot to perform compliant control in contact tasks. Figure 5

[0092] In order to realize the compliant control of the robot in contact tasks, the kinematics and dynamics models of the robot and the environment are established, and the motion and force analysis of the robot in the contact task operation process is studied. The environment is simulated as a spring-damper system to realize the impedance control of the robot. The impedance control three-order differential equation can be expressed as:

[0093] ;

[0094] Wherein: ,, and represent the inertia matrix coefficient, damping matrix coefficient, stiffness matrix coefficient and force matrix coefficient of the target position impedance, respectively; ,, and represent the position, velocity, acceleration and jerk of the desired target trajectory, respectively; ,, and represent the actual position, velocity, acceleration and jerk of the robot, respectively; represents the desired contact force between the robot and the environment; represents the force generated by the contact between the end of the robot and the environment.

[0095] Let The arrangement can be obtained as follows:

[0096] ;

[0097] Wherein, ,, and represent the position deviation, velocity deviation, acceleration deviation and jerk deviation between the desired trajectory and the actual trajectory of the robot and the environment, respectively.

[0098] In order to ensure that the robot has good contact compliance and tracking accuracy during the execution of contact tasks , the above spring-damper system model is optimized to approximate the following linear state space model:

[0099] ;

[0100] Combine it into the model predictive control framework: ​

[0101] ;

[0102] wherein is the prediction horizon, denotes the control sequence , is the discretized model, is the control period, since the discretized model is linear, the cost function is quadratic, the optimal solution is denoted as:

[0103] ;

[0104] By constructing a new impedance model, a variable impedance model based on model predictive control, the variable impedance law design problem is converted into the control law design problem, by collecting the stiffness data of the robot arm, the contact force data to establish a unified contact task learning framework, the next state of the robot arm is predicted, the control model and the target variable are determined, and the trajectory is planned as a series of discrete time step states, the control parameters are calculated, the actual motion trajectory of the robot arm is compared with the teaching trajectory, the deviation trajectory is obtained, the force deviation signal is obtained through the above impedance controller conversion, which is added to the expected contact force / torque and subtracted from the actual contact force / torque measured by the robot arm end force / torque sensor as the input signal of the force control inner loop, and the impedance control of the robot arm is realized.

[0105] It can be understood that, in order to improve the safety of the control method of the high-speed railway station inspection robot arm operating system, the interaction with the environment and the human body needs to be considered during the operation of the robot arm, therefore, the impedance control of the robot arm is required. The impedance control is a control method based on the interaction force between the robot arm and the environment, which measures and processes the interaction force between the robot arm and the environment to realize real-time control of the motion of the robot arm.

[0106] Specifically, first, the kinematics and dynamics model of the robot arm is constructed to describe the motion trajectory and torque feedback information of the robot arm, predict the next state of the robot arm, determine the control model and the target variable, and plan the trajectory as a series of discrete time step states.

[0107] Secondly, a robot control strategy is generated based on the current state and future prediction, and the control parameters are calculated by computer optimization to realize the control planning at the next moment. According to the prediction model, the spatial constraint and the time constraint, the error between the current state and the target state of the robot arm is minimized to adjust the interaction force between the robot and the environment.

[0108] Then, according to the control parameters calculated by the model predictive control optimizer, the variable impedance parameters of the robot arm are dynamically adjusted to realize dynamic control response of the robot arm, and the control model and the variable impedance parameters are adjusted in time according to the observed state of the robot arm to realize real-time feedback control of the robot arm control.

[0109] Thirdly, torque control is performed, the impedance control can feedback the torque information in the environment, and the robot interacts with the environment according to the feedback information to control the force. Through the torque control mode, the robot can better adapt to different environments and motion states, and keep stable motion.

[0110] Finally, motion planning is performed, in the control process, the robot controls the motion of the robot arm end based on the DMPs of feedback control to determine the motion trajectory, and through the motion planning mode, the robot can more effectively perform the task and ensure high quality and high efficiency of task completion.

[0111] Please refer to Figure 6 and Figure 7 are a high-speed rail station inspection robot arm operation scene schematic diagram and a simulation schematic diagram of the embodiment, which are suitable for performing the method described in the application.

[0112] In addition, it should be noted that the technical problem of the application is how to realize autonomous anthropomorphization operation of the high-speed rail station inspection robot arm under inspection operation. The technical problem is that in the real environment, the inspection robot needs to perform inspection operation in the waiting room of the high-speed rail station, and needs to complete different contact tasks, such as opening the door of a certain machine room, shooting a contour map, etc.

[0113] In summary, in the control method provided by the application, the purpose of the multiple formulas involved is to solve the problem of autonomous anthropomorphization operation of the robot arm under inspection operation, and the specific formulas and their functions are as follows:

[0114] Dynamic time warping (DTW) formula: by aligning multiple teaching data in time, the robot can learn more accurate actions. The calculation method of the DTW formula is based on dynamic programming, and the two time series are aligned according to the similarity function, and the aligned path is obtained.

[0115] Gaussian mixture model (GMM) and Gaussian mixture regression (GMR) formula: used for extracting features such as stiffness data and contact force data of the robot arm in the teaching process. GMM uses multiple Gaussian distributions to fit the data distribution, and GMR uses Gaussian mixture model for regression, and by combining GMM-GMR method, the data can be modeled and analyzed in the teaching process, so as to extract the basic motion mode and rule.

[0116] Improved dynamic motion primitive (DMPs) formula: learn the optimal reference trajectory based on the demonstration data and realize trajectory generalization to obtain the learned trajectory. The DMPs formula is a function for describing motion, which enables the robot to imitate the demonstration part to perform the task in a non-demonstration environment.

[0117] Impedance control formula: guide the manipulator to perform compliant control in a contact task. Impedance control is a control method for adjusting the action of the robot by establishing a force and motion model between the robot and the surrounding environment.

[0118] The above formulas are progressive, and through multiple data processing and feature extraction, the autonomous anthropomorphization operation problem of the manipulator in the inspection operation can be solved. Through demonstration learning and impedance control, the robot only needs to perform simple demonstration learning, and can complete the task in a complex environment. The technical scheme of the present application realizes the control method of the manipulator operation system by comprehensively using various formulas, and provides an effective solution for intelligent inspection of high-speed railway stations.

[0119] The control method of the manipulator operation system of the high-speed railway station inspection robot of the present application can cope with the contact task that may occur in the inspection operation process of the human-robot coexistence inspection robot, and ensures that the manipulator can realize complex contact tasks in a high-speed railway station environment with high personnel density and complex environment, so as to avoid complex trajectory planning, and enable the manipulator to quickly acquire motion skills through demonstration learning, thereby effectively improving the planning control of the manipulator.

[0120] The control method of the manipulator operation system of the high-speed railway station inspection robot of the present application, on the one hand, in a complex unstructured environment such as a high-speed railway station, the present application enables the demonstrator to migrate the behavior of the demonstrator to the robot system in a relatively direct way through a small amount of programming operation process, and realizes the complex spatial motion planning of the robot; at the same time, the traditional dynamic motion primitive (DMPs) algorithm is improved to improve the learning efficiency of the algorithm, and realize the rapid trajectory planning of the manipulator; at the same time, the manipulator can extract the corresponding motion features from a small amount of demonstration samples, and generalize the features to new scenes; on the other hand, by collecting the position, velocity, interaction force, interaction stiffness and other data in the process of the demonstrator performing the task, a unified contact task impedance control learning framework is established, and the manipulator is autonomously anthropomorphized compliant control in a complex unstructured environment such as a high-speed railway station is realized by combining demonstration learning.

[0121] Finally, it should be noted that the above only describes the preferred embodiments of the present application, and is not limited to the present application, although the foregoing embodiments of the present application are described in detail, for those skilled in the art, it can still be modified to the technical solutions described in the foregoing embodiments, or some technical features are replaced, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A control method of a high-speed rail passenger station inspection robot mechanical arm operation system, characterized in that, The method comprises the following steps: S1, first, the high-speed robot is taught in the real environment by the method of manual teaching, and multiple dragging teaching is performed; Wherein the mechanical arm of the inspection robot is made to complete the dragging teaching and the inspection action, the end gripping device of the mechanical arm of the inspection robot is made to reach each task position after the dragging teaching, and the inspection robot is made to complete the inspection operation, and multiple teaching data are obtained in the process of completing the multiple dragging teaching; S2, according to the multiple teaching data obtained in S1, the dynamic time warping (DTW) is used to align and optimize the teaching data, and the optimization method is: Optionally, the time sequence of two teaching data is described as: ； ； wherein, and denotes a binary time matrix, the time signal and is stretched to obtain a distance matrix between each sequence of teaching data points , the distance matrix formula is of size , the is the number of sequences in the teaching data set; obtaining a distance matrix The distance matrix is obtained The distance matrix is obtained The distance matrix is obtained S3, after the dynamic time warping (DTW) is used to align and optimize the teaching data in S2, the Gaussian mixture model (GMM) and the Gaussian mixture regression (GMR) are combined to smooth the teaching trajectory, and the best mechanical arm stiffness data and the contact force data of the end gripping device of the mechanical arm are obtained, and the best stiffness data and the contact force data are extracted; S4, according to the best mechanical arm stiffness data and the contact force data of the inspection robot obtained in S3, the improved dynamic motion primitive (DMPs) is used for processing, and the best reference trajectory of the best inspection robot mechanical arm is obtained, the trajectory generalization of the inspection robot is performed according to the best reference trajectory, and the robot obtains the learning trajectory; S5, according to the mechanical arm stiffness data and the contact force data of the end gripping device of the mechanical arm obtained in S4, a unified contact task learning framework is established, then impedance control is performed, and finally the mechanical arm performs contact task compliant control.

2. The control method of the high-speed railway station inspection robot mechanical arm operation system according to claim 1, characterized in that, In the step S1, the teaching is acquired by manually dragging multiple times a bar teaching trajectory, which is represented as: , ; wherein represents the length of the th trajectory, represents dimensional input time, represents dimensional trajectory variable; the pose of the end of the robot arm relative to the base coordinate system is represented as wherein .

3. The control method of the high-speed railway station inspection robot mechanical arm operation system according to claim 2, characterized in that, In the step S2, in the hash bucket, all sequences with the same hash value are found according to the fast index, the DTW distance of the sequences with the same hash value is calculated, and the sequence with the smallest distance is selected as the most similar teaching data sequence.

4. The control method of the high-speed railway station inspection robot mechanical arm operation system according to claim 3, characterized in that, In said step S3, the joint probability distribution of the input variables and the output variables in the teaching trajectory is modeled Modeling: ； wherein , is the number of Gaussian components in the GMM, denotes the prior probability of the th Gaussian component, denotes the mean vector of the probability density function of the th Gaussian component, denotes the covariance matrix of the th Gaussian component.

5. The control method of the high-speed railway station inspection robot mechanical arm operation system according to claim 4, characterized in that, In the step S3, after obtaining the GMM parameters, the conditional probability distribution model of the corresponding trajectory of any new input is predicted by using the GMR. ​ 。 6. The control method of the high-speed railway station inspection robot mechanical arm operation system according to claim 5, characterized in that, In step S4, the processed optimal reference trajectory is given. , Starting point Using local weighted regression as the target point, the basis function is obtained. Weighting coefficients This leads to a new learning trajectory. With teaching trajectory They have similar motion trajectory trends, which enables the robotic arm to learn the instructor's teaching trajectory, and finally allow the robotic arm's end effector to reach the target position of the task by imitating the instructor.

7. The control method of the high-speed railway station inspection robot mechanical arm operation system according to claim 6, characterized in that, In step S4, the compliant control method is to establish the kinematics and dynamics model of the mechanical arm and the environment, study the motion and force analysis of the mechanical arm in the contact task operation process, simulate the environment as a spring damping system, and then realize the impedance control of the mechanical arm.

Images