Inspection robot slip identification and control method based on current prediction

Abstract

An inspection robot slip identification and control method based on current prediction comprises the steps that firstly, an experiment site is arranged, the possible slip condition of an inspection robot in a task is simulated, and current data are collected through an experiment and preprocessed; building a training network model, carrying out prediction training on the collected current data, and carrying out performance evaluation on the trained model; predicting the change trend of n current data in the future by using the trained model, and judging whether the inspection robot slips or not; the corresponding control strategy is adjusted according to whether the inspection robot slips or not, and the influence caused by slipping is reduced. According to the invention, control strategy adjustment is made in advance by predicting the probability of occurrence of slipping in advance.

Description

Technical Field

[0001] The present invention belongs to the technical field of mobile robot travel control, and in particular relates to a mobile robot slip identification and control method. Background Art

[0002] A patrol robot is a machine device that automatically performs prescribed tasks. It is generally used to replace human work and can operate uninterruptedly 24 hours a day. It is often used in dangerous places such as pipeline corridors, chemical plants, and power plants.

[0003] When the inspection robot is performing its work tasks, it is easy for the drive wheel output torque to be greater than the ground adhesion to the tire on some surfaces with low ground friction coefficients (such as muddy and slippery roads), causing the drive wheel to slip. Its actual driving linear speed will be lower than the driving wheel rotation linear speed, causing the inspection robot to deviate from its original working path, affecting its operations. In serious cases, dangerous situations such as rollover may occur.

[0004] In the prior art, there are many commonly used methods for identifying mobile robot skidding, such as: comparing the displacement output by the encoder odometer with the displacement output by sensors such as IMU and GPS, and judging whether skidding has occurred when the difference between the two exceeds a threshold; judging whether skidding has occurred based on a certain rate of change threshold by continuously collecting the drive wheel motor current and encoder speed signals; using multi-sensor fusion to improve positioning accuracy, and judging whether skidding has occurred based on the relationship between the robot's walking trajectory and the planned trajectory. It can be seen that in the prior art, there is a certain lag in the judgment of skidding. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides a patrol robot slip identification and control method based on current prediction, which can predict the occurrence of slip in advance to a certain extent and help the controller make corresponding control strategies earlier.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for detecting and controlling a slippage of an inspection robot based on current prediction comprises the following steps:

[0008] Step S1, obtaining the inspection robot parameters, including the center of mass position, mass and running slope, and calculating the positive pressure on the driving wheel;

[0009] Step S2, arranging an experimental site, simulating the slipping situation that the inspection robot may encounter during the task, collecting current data through experiments and preprocessing it;

[0010] Step S3: building a training network model, performing prediction training on the collected current data, and performing performance evaluation on the trained model;

[0011] Step S4: using the trained model to predict the changing trend of n current data in the future, and determining whether the inspection robot has slipped;

[0012] Step S5: adjusting the control strategy according to whether the inspection robot slips or not, so as to reduce the impact of slipping.

[0013] Furthermore, in step S1, the center of mass position of the inspection robot includes the distance from the center of mass to the front axle, the distance from the center of mass to the rear axle, and the height of the center of mass, which are solved by SolidWorks software.

[0014] Preferably, in step S1, the mass of the inspection robot is obtained by weighing or solving with three-dimensional software, and the running slope is obtained by converting the acceleration signal collected by the IMU sensor.

[0015] In the step S1, after obtaining the relevant parameters of the inspection robot, the vertical load of the driving wheel is calculated according to the robot dynamics model.

[0016] Furthermore, in step S2, the experimental site selects a common operating site for inspection robots, and sets corresponding slopes according to the technical requirements of the robot, simulating the scene of low actual ground adhesion coefficient by adding water or laying mud and sand on different sections of the road.

[0017] Preferably, in step S2, the driving wheel motor data is collected by a servo driver, sent to a vehicle controller via CAN communication, and printed out.

[0018] In step S2, the pre-processing process is:

[0019] (21) Based on the idea of ​​sliding window, the current data is filtered by Gaussian weighted average method;

[0020] (22) The collected current data is a single column of time series data, which is re-divided into m+n columns of data according to the prediction step length m and the prediction step number n;

[0021] (23) The divided data is normalized to facilitate network training.

[0022] Furthermore, in step S3, the neural network structure used for training is a TCN-GRU-ATTENTION network. After the training is completed, the next n data can be predicted based on the first m data.

[0023] In step S3, the performance of the trained model is evaluated by means of root mean square error, mean absolute error and coefficient of determination.

[0024] In step S4, the torque output by the motor is related to the current. The torque constant of the motor can be obtained by consulting the motor manual, and the reduction ratio can be obtained by consulting the reducer nameplate. The motor torque is the product of the parameters.

[0025] In step S4, after the motor output torque is obtained, the angular acceleration information of the driving wheel is collected through the motor encoder, and the longitudinal force can be calculated according to the single-wheel dynamics model. The adhesion coefficient is the ratio of the driving wheel longitudinal force to the vertical load in step S1.

[0026] In step S4, a corresponding current change threshold is set to determine whether the inspection robot is slipping at this time.

[0027] In step S5, when slippage is detected, the magnitude of the motor output driving force is changed by adjusting the robot's running speed. The adjustment is made according to the difference between the encoder wheel speed and the linear speed obtained by the IMU solution. When the difference between the two is within a certain threshold, it can be considered that the slippage situation has improved.

[0028] The technical concept of the present invention is: when the inspection robot slips, since the driving wheel motor adopts speed feedback control, the overall speed performance is relatively stable; when slipping occurs, the friction coefficient decreases, and the load torque of the driving wheel decreases. At this time, the current of the driving wheel motor will also decrease accordingly. The degree of reduction in the motor current will vary depending on the degree of reduction in the friction coefficient.

[0029] The beneficial effects of the present invention are as follows: by predicting the changing trend of the current through a pre-trained model, the occurrence of slippage can be determined earlier, and the controller can intervene and adjust when the inspection robot has not yet seriously deviated, thereby greatly reducing the occurrence of more dangerous situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the overall framework of the present invention.

[0031] Figure 2 This is a schematic diagram of the longitudinal climbing force analysis of the inspection robot in the present invention.

[0032] Figure 3 This is a flow chart of the current-based network model training in the present invention.

[0033] Figure 4 This is a flow chart of the slip state judgment based on current prediction in the present invention.

[0034] Figure 5 This is a force analysis diagram of a single driving wheel in the present invention.

[0035] Figure 6 This is a block diagram of the speed adjustment strategy in the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with the accompanying drawings.

[0037] refer to Figure 1 to Figure 6 A method for detecting and controlling the slippage of an inspection robot based on current prediction is proposed. The idea of ​​time series prediction is introduced to improve the foresight of detecting the slippage of the inspection robot.

[0038] In one embodiment of the present invention, a method for identifying and controlling the slippage of an inspection robot is provided, wherein the driving wheel of the inspection robot is driven by a servo motor, the servo motor is driven by a servo driver, and the servo driver works under a speed loop. When the driving wheel is in a slipping state, the output value of the motor encoder remains unchanged under the regulation of the speed loop, and at this time, the tangential force of the ground on the driving wheel decreases, which is equivalent to a decrease in the motor load torque. At this time, the driving motor current will show a downward trend, and it can be judged whether slippage occurs.

[0039] To achieve the above purpose, Figure 1 As shown, the present invention proposes a method for detecting and controlling the slippage of an inspection robot based on current prediction, comprising the following steps:

[0040] Step S1, obtain the inspection robot parameters, including its center of mass position, mass and running slope, and calculate the positive pressure on the driving wheel. The process is as follows:

[0041] Step S11, in SolidWorks software, assigning materials and masses to various parts of the inspection robot, and obtaining the mass and center of gravity position of the inspection robot through the "mass attribute" function;

[0042] The robot adopts a symmetrical design, and it can be assumed that the center of mass is located at the center of the left and right driving wheel axles. The parameters obtained are mainly the distance a from the center of mass to the front axle, the distance b from the center of mass to the rear axle, and the height h of the center of mass;

[0043] Step S12, obtaining acceleration information through the IMU sensor arranged on the inspection robot, and calculating the slope of the road surface at that time;

[0044] Step S13, longitudinal force analysis of the inspection robot when climbing a slope is as follows: Figure 2 As shown;

[0045] In the figure: G is the robot's gravity; θ is the climbing angle; a is the distance from the center of mass to the front axle; b is the distance from the center of mass to the rear axle; h is the height of the center of mass; F NF 、F NRF is the normal force of the ground on the front and rear wheels; fF 、F fR is the magnitude of the tangential force exerted by the ground on the front and rear wheels;

[0046] The longitudinal dynamics calculation of the inspection robot is shown in the formula:

[0047] F t =F f +F a +F i +F w

[0048] Where: F t F is the driving force; f F is the rolling resistance; a F is the acceleration resistance; i is the slope resistance; F w is air resistance;

[0049] The inspection robot is generally used for indoor operations and has a low running speed, so the air resistance can be ignored. The torque balance is calculated at the contact point between the front wheel and the ground as shown in the formula:

[0050]

[0051] Furthermore, the vertical load of the driving wheel is calculated as shown in the formula:

[0052]

[0053] Step S2: Figure 3 As shown, the present invention proposes a data collection and prediction model training process;

[0054] Arrange the experimental site to simulate the slipping situation that the inspection robot may encounter during the task. Collect the current data through the experiment and pre-process it. The process is as follows:

[0055] Step S21: According to the scenes that may cause the inspection robot to slip during operation, such as water accumulation, oil pollution, mud and sand, etc., a corresponding experimental site is built, and tasks are issued to the inspection robot to make it run along a certain trajectory. The motor current signals of the inspection robot when it runs smoothly and on a slippery road section are collected through the servo driver, and the driver sends them to the vehicle controller VCU through CAN communication. The VCU saves the data in the form of printing log files;

[0056] Step S22: Due to the existence of various interferences, the collected current data needs to be filtered. The Gaussian weighted moving average method is used, with a sliding window size of n and a step size of 1. The process is as follows:

[0057] (221) Define the filter window: its size depends on the characteristics of the signal and the noise level;

[0058] (222) Calculate Gaussian weight: Calculate the weight of each point in the window. The closer the point is to the current sample point, the larger the weight is, and the farther the point is, the smaller the weight is.

[0059] (223) Weighted average: Perform weighted average on the values ​​of each point in the window to obtain the preliminary filtering result of the current sample point;

[0060] (224) Moving window: Move the filter window forward by a step size of 1, and repeat the above steps until all sample points have been processed;

[0061] (225) Output result: Output a smooth signal with reduced noise.

[0062] Step S23, the collected current data is a single column of time series data, which needs to be re-divided before using the network training. Assume that the prediction step is m, the number of prediction steps is n (consistent with the filter window size in step S22), then the sliding window size is m+n, and the sliding step is set to 1. The following are the specific steps: starting from the first data in the time series data, select m+n data as the first row of the new data, then move the sliding window by 1 step, and select m+n data again from the second data as the second row of the new data, and so on, and finally obtain new data of m+n columns, of which the first m columns are used as the input of the network, and the last n columns are used as the output of the network, thereby realizing the prediction of n future data with m historical data;

[0063] In order to improve the performance of model training, the new data is normalized, and the calculation formula is as follows:

[0064]

[0065] x′ is the standardized input data;

[0066] x max is the maximum original data value;

[0067] x min is the minimum original data value;

[0068] By normalizing the input data, we can ensure that the value range of the input original data is within [0,1], eliminating the impact of inconsistent data on the experiment;

[0069] Step S3: Build a training network model, perform prediction training on the collected current data, and perform performance evaluation on the trained model; the network model is a TCN-GRU-ATTENTION combination model, which mainly includes a TCN module, a GRU module, and an ATTENTION module; the process is as follows:

[0070] Step S31, building a training network model, specifically:

[0071] The temporal convolutional network (TCN) is based on a one-dimensional convolutional neural network and integrates multiple mechanisms such as dilated convolution, causal convolution, and residual connection modules. The core of TCN is the dilated causal convolution and residual connection blocks. Causal convolution is a unidirectional structure, that is, the output of a certain time point in the previous level is only related to the input of the next level at that time point and the previous input. The calculation is shown in the formula:

[0072]

[0073] Where: t is the time step; F is the convolution operation; F(t) is the output value at time point t; k is the convolution kernel size; f(i) is the i-th element in the convolution kernel; d is the dilation coefficient; x t-d*i is the historical data in the input sequence;

[0074] The Gated Recurrent Unit (GRU) network has two key gating mechanisms: the reset gate and the update gate. It uses an integrated update gate mechanism that effectively integrates the functions of the forget gate and the output gate in the LSTM. The reset gate determines the amount of information lost in the previous moment, and the update gate controls the amount of state information transmitted from the previous moment to the current moment. These two gating mechanisms can retain important features and effectively capture the long-term and short-term dependencies of the sequence. The calculation is shown in the formula:

[0075] z t =σ(W z h t-1 +U z x t +b z )

[0076] r t =σ(W r h t-1 +U r x t +b r )

[0077]

[0078] Where: z t is the output of the update gate; r t is the output of the reset gate; h tis the output of the hidden state; x t is the input at time t; h t-1 is the state at time t-1; σ is the sigmoid function; Tanh is the hyperbolic tangent function; W z , W r , W h is the weight vector of the recurrent connection; U z , U r , U h are the weight vectors between the input layer, update gate, reset gate and candidate state respectively; b z , b r , b h is the corresponding bias matrix;

[0079] The self-attention mechanism can assign different weights to different features during the training process, reduce dependence on external information, and effectively improve the model's ability to extract features. The attention mechanism can not only capture the global feature information of the data, but also capture the feature information between the same set of data vectors and identify important trends in changes. The calculation is shown in the formula:

[0080]

[0081] Where: Attention is the obtained attention weight; Q is the query matrix length; K is the key matrix length; V is the key vector length; T is the transpose of the matrix; dk is the dimension of the key; Softmax is the normalization equation;

[0082] Step S32, building a network model in MATLAB software, setting corresponding training parameters, such as the maximum number of training times, gradient threshold, initial learning rate, and regularization parameter, etc., using the data preprocessed in step S2 for training, and using anti-normalization to restore the output value to the data;

[0083] Furthermore, the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ) is used as an indicator for evaluating the training effect of the model. The calculation of the effect evaluation indicator is shown in the formula:

[0084]

[0085] Where: Y i is the true value; is the predicted value; is the average value of all true values ​​of samples; m is the total number of samples in the test set. RMSE represents the deviation between the predicted value and the true value, MAE represents the error between the two, and R 2 It represents the model's fit index, with a value range of 0 to 1. The closer it is to 1, the better the model's fit effect.

[0086] Step S4, such as Figure 4 As shown, the present invention proposes a method for judging the slip state of an inspection robot based on current prediction;

[0087] The trained model is used to predict the changing trend of n current data in the future and determine whether the inspection robot has slipped. The process is as follows:

[0088] Step S41: using the trained model, inputting m historical data, outputting n historical data, and obtaining the motor current data prediction set {x1 x2 … x n}, further, the output torque value of the DC servo motor can be calculated by the current value, as shown in the formula:

[0089] T=I A K t i

[0090] Where: T is the motor output torque; I A is the current of the driving motor; K t is the motor torque constant; i is the reduction ratio of the reducer;

[0091] Step S42, performing variance calculation on the current prediction value obtained in step S41, as shown in the formula:

[0092]

[0093] In the formula: var is the variance calculated from this group of data; n is the number of data; x i is the data point of this group; μ is the mean of the data points of this group;

[0094] The reason for choosing variance as the judgment indicator: After trying a variety of time series data characteristic indicators, it is found that the variance value of the current data when slippage occurs shows a rapid upward trend compared with that when the current data is running smoothly. There is an obvious distinction between the two. For this reason, a corresponding threshold is designed. When this threshold is exceeded, it is considered that slippage has occurred (it is necessary to distinguish the sudden change caused by the change of input quantity);

[0095] Step S43, based on the motor output torque prediction value calculated in step S41, combined with the single-wheel dynamics model of the driving wheel (the single-wheel force analysis diagram is as follows Figure 5 The change of the longitudinal force on the driving wheel can be obtained by the formula:

[0096]

[0097] Where: F xis the longitudinal force exerted by the ground on the driving wheel; T is the motor output torque; r is the wheel diameter of the driving wheel; J is the equivalent moment of inertia of the driving wheel; is the driving wheel rotation angular acceleration;

[0098] Further, the friction coefficient μ can be obtained by combining the normal load of the driving wheel obtained in step S13. s The size of is calculated as shown in the formula:

[0099]

[0100] Furthermore, when the friction coefficient is lower than a certain value, the inspection robot will be in an uncontrollable state due to slipping. At this time, applying control may cause the robot to completely lose control. Therefore, a maximum current change threshold Δx is set according to the friction coefficient calculation formula. max When this threshold is exceeded, the robot is considered out of control and should be stopped immediately and an alarm should be sounded, waiting for staff to handle the situation.

[0101] Step S5: Figure 6 As shown, the present invention proposes a control strategy adjustment method after slippage occurs;

[0102] According to whether the inspection robot slips or not, the control strategy is adjusted accordingly to reduce the impact of slipping. The process is as follows:

[0103] Step S51: After determining that the inspection robot has slipped, the controller first reduces the preset speed by a certain value, that is, v c = kv, where v is the original running speed of the robot, v c is the adjusted input speed value, the value of k can be given according to the current change rate in the current prediction value set in step S41, and it should be noted that the left and right driving wheels need to be adjusted synchronously;

[0104] Step S52, after the initial deceleration is completed, the wheel speed signal of the motor encoder and the speed signal obtained by the IMU acceleration integration are collected, and the difference between the two is compared. If the difference between the two is too large, it is considered that the robot is still in a slipping state. The controller needs to continue to adjust the input speed value until the difference between the two is controlled within a certain range, that is, v c =kv+Δv; At the same time, a model predictive controller is designed for motion control so that the inspection robot can still run according to the predetermined trajectory when deviation occurs.

[0105] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. A method for detecting and controlling slippage of an inspection robot based on current prediction, characterized in that: The method comprises the following steps: Step S1, obtaining the inspection robot parameters, including the center of mass position, mass and running slope, and calculating the positive pressure on the driving wheel; Step S2, arranging an experimental site, simulating the slipping situation that the inspection robot may encounter during the task, collecting current data through experiments and preprocessing it; Step S3: building a training network model, performing prediction training on the collected current data, and performing performance evaluation on the trained model; Step S4: using the trained model to predict the changing trend of n current data in the future, and determining whether the inspection robot has slipped; Step S5: adjusting the control strategy according to whether the inspection robot slips or not, so as to reduce the impact of slipping.

2. A method for detecting and controlling slippage of an inspection robot based on current prediction as claimed in claim 1, characterized in that: In the step S1, the center of mass position of the inspection robot includes the distance from the center of mass to the front axle, the distance from the center of mass to the rear axle, and the height of the center of mass, which are solved by SolidWorks software.

3. The method for detecting and controlling slippage of an inspection robot based on current prediction according to claim 2, characterized in that: In step S1, the mass of the inspection robot is obtained by weighing or solving with three-dimensional software, and the running slope is obtained by converting the acceleration signal collected by the IMU sensor.

4. The method for detecting and controlling slippage of an inspection robot based on current prediction according to claim 3, characterized in that: In the step S1, after obtaining the relevant parameters of the inspection robot, the vertical load of the driving wheel is calculated according to the robot dynamics model.

5. A method for detecting and controlling slippage of an inspection robot based on current prediction as claimed in any one of claims 1 to 4, characterized in that: In step S2, the experimental site is selected as a common operating site for inspection robots, and corresponding slopes are set according to the technical requirements of the robot, and the scene with low adhesion coefficient of the actual ground is simulated by adding water or laying mud and sand in different sections; The driving wheel motor data is collected by the servo driver, sent to the vehicle controller via CAN communication and printed out; In step S2, the pre-processing process is: (21) Based on the idea of ​​sliding window, the current data is filtered by Gaussian weighted average method; (22) The collected current data is a single column of time series data, which is re-divided into m+n columns of data according to the prediction step length m and the prediction step number n; (23) The divided data is normalized to facilitate network training.

6. A method for detecting and controlling slippage of an inspection robot based on current prediction as claimed in any one of claims 1 to 4, characterized in that: In step S3, the neural network structure used for training is the TCN-GRU-ATTENTION network. After the training is completed, the next n data can be predicted based on the first m data; the performance of the trained model is evaluated by the root mean square error, mean absolute error and determination coefficient.

7. The method for detecting and controlling slippage of an inspection robot based on current prediction according to claim 6, characterized in that: In step S3, the torque output by the motor is related to the current. The torque constant of the motor can be obtained by consulting the motor manual, and the reduction ratio can be obtained by consulting the reducer nameplate. The motor torque is the product of the parameters.

8. A method for detecting and controlling slippage of an inspection robot based on current prediction as claimed in any one of claims 1 to 4, characterized in that: In step S4, after the motor output torque is obtained, the angular acceleration information of the driving wheel is collected through the motor encoder, and the longitudinal force can be calculated according to the single-wheel dynamics model. The adhesion coefficient is the ratio of the driving wheel longitudinal force to the vertical load in step S1.

9. The method for detecting and controlling slippage of an inspection robot based on current prediction according to claim 8, characterized in that: In step S4, a corresponding current change threshold is set to determine whether the inspection robot is slipping at this time.

10. A method for detecting and controlling slippage of an inspection robot based on current prediction according to any one of claims 1 to 4, characterized in that: In step S5, when slippage is detected, the magnitude of the motor output driving force is changed by adjusting the robot's running speed, and the adjustment is made according to the difference between the encoder wheel speed and the linear speed solved by the IMU. When the difference between the two is within a certain threshold, it is considered that the slippage situation has improved.

Images