Past Work

Physical Modeling and Identification

The cost reduction of the robot manipulators and the resulting increased flexibilities and non-linear behavior implies an increasing model complexity and larger model variations. This inherently requires effective methods to identify the unknown model parameters. Effective methods are particularly important since the parameter variations might require tuning of each individual robot.

Several aspects of this have been treated within the sub-project:

Additional sensors will enable identifiability of more complex model structures, reduced measurement time, and increased model accuracy. Questions regarding optimal sensor location, experiment design, etc. are also covered.
The need to accurately model and identify nonlinearities such as friction, nonlinear stiffness, torque ripple in motors and speed reducers, etc. grows with the continuous cost reduction. Many of these phenomena are also time varying (e.g. temperature dependent), which either require recursive identification or models that describe these variations.

Identification of elastic parameters has mainly been carried out in the frequency domain, and results are presented in the licentiate thesis [Moberg, 2007] and the dissertation [Moberg, 2010] of Stig Moberg. Stig Moberg has also developed two benchmark examples, one a highly flexible four mass model of a single arm robot, and the second a two degree of freedom robot arm with joint flexibilities. In both cases the benchmark models are presented in journal or conference papers and the models are also provided as Simulink-models including the evaluation criteria that will be used. A specification for the control problem is also provided in the papers. In addition an extension to the established joint flexible robot model is presented and referred to as the extended flexible robot model. The extended model includes additional, non-actuated, flexibilities. The extension is motivated from experiments.

Sensor informatics and control

A standard robot controller relies only on sensors measuring the angles of the motor shafts for the feedback control. There is no measurement feedback from the arm structure, which is separated from the motor shafts by compliant bearings and compliant speed reducers with high friction and nonlinear stiffness. In this sub-project the aim has been to get better estimates of the robot tool position and orientation by adding sensors to the arm structure, for example accelerometers and gyroscopes.

Methods to fuse measurements from several sensors are well developed, but have not been applied to industrial robots in any larger extent. Some issues to be studied are:

Comparison of different methods for sensor fusion, including EKF, Particle Filters and complementary filters.
Computational complexity considering model size and sensor fusion method.

The primary reason for using sensor fusion is to enable for example:

Improved robustness in the controller due to errors in the dynamic models.
Better control performance, for example increased stiffness in applications with large tool forces such as material removal, or improved ILC with reduced tool position errors compared to only using motor measurements.
Increased model accuracy during identification.
Enhanced robustness and accuracy of robot diagnosis and fault detection.

Results in this sub-project has been reported in the licentiate thesis [Wallén, 2009] and the dissertation [Wallén, 2011] by Johanna Wallén. Wallén introduced estimation based ILC, a framework where it is possible to analyze stability and performance in the combination of estimation and ILC. The new development was motivated by the typical robot configuration where no sensor measures the arm position which should follow a programmed path. Experiments using estimation based ILC on a parallel kinematic robot manipulator was also included and the work was performed together with Isolde Dressler at Lund University.

Another activity has aimed at developing new model based control methods. Since there is a general interest in improving the position, speed, and acceleration estimates of the robot arm system for control and diagnosis, a general sensor fusion activity was founded within LINK-SIC in 2008. The result from this activity is a patent application, one Licentiate’s thesis [Axelsson, 2011] and a dissertation [Axelsson, 2014]. From 2011 the focus moved from estimation towards combined estimation and control, using robust control and the results are included in Axelsson’s dissertation. Different approaches to achieve robustness and still be able to reach the desired performance requirements are considered and as one result a new, improved design methodology, was developed. During 2014 a Master Thesis work by Johan Norén, now at ABB, investigated and compared different methods for estimation of tool position and orientation using an IMU. The work considered using the IMU only or to combine the IMU measurement with the motor angle measurements and the result showed that, although, the IMU was considered very accurate with very low drift and bias it is necessary to include other measurements to achieve the required accuracy. As estimation method EKF and complementary filter was also compared showing that the complementary filter has similar performance but with much less number of design parameters.

Robot Fault diagnosis

The activities within robot fault diagnosis in LINK-SIC have provided a number of results, and the work can be split into four phases. In the first phase the aim was to develop a mathematical model of friction, including dependence on temperature and load, in addition to the velocity dependence normally present in friction models. The model building was done using data collected at the laboratory facilities of ABB, and the resulting model gives a considerably better ability to predict frictions in robots joints compared to previously published models. The model gives important insight into the nature of friction, and guidance for how to design and build diagnosis functionalities in robot systems. In practice the wear is even closer connected to the condition and reliability of a robot, and the second phase was hence devoted to the development and evaluation of methods to estimate the wear directly. By incorporating the problem in a statistical framework it was possible to derive measures of the accuracy of such estimates, and in particular determine optimal operating conditions in terms of e.g. velocity for estimating the wear. The wear estimation method was designed using the friction model developed during phase one, and it was tested on data collected from tests carried out at ABB. The tests showed that the method is reliable and that the wear estimate is a very good measure of the condition of the gear box. A key feature in many robot applications is that the robots carry out their operation repeatedly. This has resulted in a method that is based on monitoring of the torque distribution. The key element is to compare the distribution from different repetitions of a cycle, using different statistical distance measures related to the distribution, and to diagnose wear changes based on the distance measure. The results are presented in a number of journal as well as conference papers, and they are also included in a licentiate thesis [Bittencourt, 2012] and in a dissertation [Bittencourt, 2014].