1. Introduction
Normally, an airborne heavy-duty mechanical arm has low positioning accuracy. “Airborne heavy-duty mechanical arm” is defined as a heavy-duty mechanical arm installed on a piece of coal mine machine. The coal mine machine is a high-voltage (1140 V/50 Hz) machine which uses hydraulic power to drive its mechanical arm. The coal mine machine can be a mining drill jumbo, a roof bolter, and a road header. The mining drill jumbo built by mining technologies international Inc. (MTI) has an airborne heavy-duty mechanical arm. The positioning accuracy of the MTI equipment is 50 mm [
1]. The manipulators of a dual-arm rock drilling rig in [
2] can also be considered as an airborne heavy-duty mechanical arm. Its positioning accuracy before calibration is 87 mm. In this paper, the heavy-duty mechanical arm of a roof bolter is studied. The positioning accuracy of the arm on the roof bolter is over 100 mm before calibration. Large positioning error not only degrades the quality of task, but also destroys equipment and brings fatal dangers to operators.
It is well known that the error sources of a mechanical arm can be classified into geometric errors and nongeometric errors. The geometric errors refer to nonperfect structure manufacturing, such as nonuniform link geometry and link twist. The nongeometric errors refer to the errors caused by joint and link compliance, gear backlash, and gear wear. Due to the heavy-duty characteristics of an airborne mechanical arm in coal mine, both types of errors have great impact on the performance of a coal mine machine.
Kinematics calibration is one of the most popular methods to improve the positioning accuracy of an industrial robot. By applying kinematics calibration, the positioning error caused by the link deformation can be compensated without changing the structure and links of an industrial robot. Much research on kinematics calibration has been reported in literature. In [
3], Gao et al. proposed a parameter identification method based on Denavit–Hartenberg (DH) model. In [
4,
5,
6,
7], researchers studied the method of kinematics calibration based on the product of exponentials (POE) or improved POE model. In [
8], Li et al. proposed an error model for serial robot kinematics calibration based on dual quaternions. The above mentioned research focused on modeling and geometric parameter error identification. However, these studies ignored the nongeometric errors, which are critical for positioning accuracy. Nongeometric errors cannot be ignored, especially when the equipment is heavy-duty.
There are few publications regarding the nongeometric error calibration. Chen et al. [
9] proposed a rigid-flexible coupling error model for nongeometric calibration of a robot. In [
10], Gong et al. established a comprehensive error model to identify the geometric errors, the position dependent compliance errors and time-variant thermal errors of an industrial robot. Some researchers [
11,
12] also established a comprehensive error model composed of geometric errors and compliance errors, and identified the geometric parameters and joint compliance of industrial robots. Owing to the difficulty of building an error model to include all nongeometric errors, most of the work in literature only took some nongeometric errors into consideration.
Thanks to its principle of data driven modeling, the artificial neural network (ANN) has a promising application in modeling complex systems such as calibration and error compensation of industrial robots [
13,
14,
15,
16]. However, one of the drawbacks of ANN for robot error calibration is the uncertainty of error sources. Therefore, Nguyen et al. [
17] proposed a combination technique, which used extended Kalman filter (EKF) to identify the geometric errors and used the ANN to compensate the nongeometric errors. Similarly, Wang et al. [
18] also proposed a combination method. In [
18], the Denavit–Hartenberg (DH) model was applied to model and identify geometric errors. The nongeometric error sources, which are difficult to model, were compensated using ANN.
However, the training of an ANN is extremely slow. To overcome the limitation, Huang et al. [
19,
20] proposed an extreme learning machine (ELM) algorithm. Compared with the conventional ANN algorithm, ELM algorithm has faster learning speed and better generalization performance. ELM is a nonlinear system and can be used to compensate the unknown nonlinearity of robot nongeometric errors. Yuan et al. [
21] proposed to improve the absolute positioning accuracy of an aviation drilling robot based on ELM. Although in [
21] the implementation was easy, a user cannot obtain the source of geometric errors. Additionally, its performance will be less robust when outliers exist or when there are less training data. Deng et al. [
22] proposed regularized extreme learning machine (RELM) based on empirical risk minimization (ERM) principle and structural risk minimization (SRM) principle. It turned out that RELM provided better generalization ability than ELM and run extremely fast like ELM.
In the current work, a novel calibration method is proposed to improve the positioning accuracy of the heavy-duty mechanical arm of a roof bolter in coal mine. This method is based on the combination of ELGA-based identification of the geometric errors and RELM-based residual errors compensation. The proposed method has the advantages of obtaining geometric error sources, high calibration efficiency, fast training speed, and good generalization performance.
The contributions of this paper are as follows:
A method based on the combination of ELGA and RELM was proposed to solve the positioning error of the heavy-duty mechanical arm of a roof bolter in coal mine. It not only obtained the accurate knowledge of geometric error sources, but also compensates the nongeometric residual errors.
The error-limited genetic algorithm (ELGA) is proposed to more accurately identify the optimal geometric parameters of the heavy-duty mechanical arm of a roof bolter. Compared with the conventional genetic algorithm, ELGA can converge to a better solution.
The regularized extreme learning machine (RELM) is innovatively applied in residual positioning error compensation of the heavy-duty mechanical arm of a roof bolter in coal mine, which improves the speed of the positioning error calibration.
The rest of this paper is organized as follows:
Section 2 provides the forward kinematics model for the identification of the geometric parameters. The error limited genetic algorithm (ELGA) and the principle of geometric parameter identification of the forward kinematics model are presented in
Section 3, which is followed by the errors compensation principle based on RELM and constructs the RELM network in
Section 4.
Section 5 provides experimental setup and related results. The conclusion is given in
Section 6.
2. Overview of the Problem
The actuation mechanism of a typical roof bolter is composed of a boom and a mast. The weight of each mast is about 2000 kg. In this paper, the goal is improving the positioning accuracy of the mast end position of a roof bolter.
Figure 1a shows the heavy-duty mechanical arm of a roof bolter and can function as a typical example of airborne heavy-duty mechanical arms in coal mine.
Figure 1b shows the kinematics model of the airborne heavy-duty mechanical arm with six joints, which defines the coordinate system of each joint. {
B} is the base coordinate system and {
T} is the tool coordinate system.
OB,
O1,
O2,
O3,
O4,
O5,
O6 and
OT are the origin of coordinate system {
B}, {1}, {2}, {3}, {4}, {5}, {6} and {
T}, respectively. The origin
OT of the tool coordinate system {
T} is the mast end of a roof bolter.
Each rotational joint of the mechanical arm takes the z axis as the rotation axis, where θ1, θ2, d3, θ4, θ5, θ6 are the motion variables of each joint around the z axis of the corresponding coordinate system, θ1 is the horizontal angle of the boom, θ2 is the vertical angle of the boom, d3 is the length of the boom, θ4 is the vertical angle of the mast, θ5 is the horizontal angle of the mast, and θ6 is the rotation angle of the mast around the z axis.
A homogeneous transformation from the {
B} to the {
T} coordinate system, which describes the position and pose of the mast end with respect to the base coordinate system {
B}, is computed as Equation (1), where [
Px, Py, Pz] is the position vector of the mast end relative to the base coordinate system {
B}, [
nx, ny, nz] is the direction cosine between the
x axis of the tool coordinate system {
T} and the
x, y, z axis of the base coordinate system {
B}, [o
x, oy, oz] is the direction cosine between the
y axis of the tool coordinate system {
T} and the
x, y, z axis of the base coordinate system {
B}, [
ax, ay, az] is the direction cosine between the
z axis of the tool coordinate system {
T} and the
x, y, z axis of the base coordinate system {
B}. Please refer to
Appendix to obtain transformation matrix
, (2 ≤
i ≤ 6) and
[
23].
4. RELM Compensation Principle
However, the geometric parameters identification using the ELGA does not eliminate the residual errors caused by nongeometric parameters. In this section, the residual errors compensation using the RELM with fast learning speed and good generalization ability is presented.
A single layer ELM network was used. The input weights and biases of the hidden layer are initialized randomly. These parameters do not need to be updated. Only the input weights of the output layer are to be obtained. The input weights of the output layer can be obtained by using Moore Penrose of the hidden layer output matrix [
24]. ELM does not need iterative operation, which greatly improves the learning rate and generalization ability.
A three-layer ELM structure is shown in
Figure 3.
Pr = [
Pxr,
Pyr,
Pzr] is the actual position measured using the laser tracker.
Pm = [
Pxm,
Pym, Pzm] is calculated using the model before ELGA calibration.
P = [
Px,
Py,
Pz] is calculated using the model after ELGA calibration. The error
E = [
ex,
ey,
ez] between the actual position value
Pr and the calibrated position value
P is expressed as
E = Pr − P.
The three elements Px, Py and Pz of the position value P are three input nodes. The hidden layer consists of L neurons nodes. The error E is the output of the ELM. The three elements ex, ey and ez of the error E are three output nodes.
For
N arbitrary samples {
Pi,
Ei |
i = 1, …,
N }, the output function of ELM is expressed as Equation (6):
where
Pi = [
Pxi,
Pyi,
Pzi]
T is the input,
Ei = [
exi,
eyi,
ezi]
T is the output,
βj = [
βxj,
βyj,
βzj]
T represents the weight vector between the jth hidden node and the output nodes,
hj(Pi) is the output function of the jth hidden point as expressed in Equation (7):
where
g is a nonlinear activation function of hidden layer, sigmoid function is selected for
g,
wj = [
wxj,
wyj,
wzj ]
T is the weight vector between the
jth hidden node and the input nodes,
bj is the bias of the jth hidden node.
In Equation (6), the
N equations can be written compactly as Equation (8):
where
represents the weight matrix between hidden layer and output layer.
There are two steps to train the ELM network on the training data set. Firstly, input weights and biases of the hidden layer are generated randomly. The input data is mapped to a new feature space using Equation (7). Then, the weight matrix
between hidden layer and output layer is calculated. When the number of samples is small, empirical risk minimization is likely to produce overfitting. According to the Bartlett’s theory [
25], for the feedforward neural network with small training error, smaller weights norms leads to better network generalization performance. Deng et al. [
22], Huang et al. [
24] studied the RELM. The unweighted regularization objective function is used here, as in Equation (9):
where the first term is the least square error, i.e., the empirical risk, and the second term is the regularization term, i.e., the structural risk,
> 0 is the parameter which balances the least square error and the regularization term. After a sixfold cross-validation experiment,
was selected as 1. When the number of hidden nodes
L is less than or equal to the number of training data, the solution is as Equation (10):
When the number of hidden nodes
L is greater than the number of training data, the solution is as Equation (11):
After is calculated, testing data set is used to verify its effect.
6. Conclusions
In this paper, a novel calibration method based on error-limited genetic algorithm (ELGA) and regularized extreme learning machine (RELM) was proposed to improve the positioning accuracy of an airborne heavy-duty mechanical arm. The novel calibration method has many advantages, such as obtaining geometric error sources, less training time, good generalization ability, and good error calibration effect.
The heavy-duty mechanical arm of a roof bolter was studied in this work to prove the proposed positioning calibration method. We first used the sensors on the equipment, combining with the forward kinematics model, to calculate the nominal mast end position. Then, we used a laser tracker to measure the actual mast end position. Based on the calculated and the measured positions, a root mean square error equation was established. The geometric parameters of the error equation were identified and optimized using the ELGA. Finally, the kinematics model was updated.
Based on the updated kinematics mode, further residual error compensate was carried out using a RELM network. After error compensation with RELM, the RMSEs, MAEs and MAXEs on testing data set were reduced by more than 78.23%, 81.35%, and 58.72% respectively. The maximum absolute errors (MAXEs) in the x, y and z directions were reduced to less than 29.54 mm. It indicates that the method proposed in this paper can meet the support accuracy requirements of the 50 mm × 50 mm mesh for a roof bolter in coal mine. It was further proved in the validation experiment.