Development of a Sensitive Winding Application Based on a Serial Robot and Integrated Torque Sensors

During the winding process, the winding tool always occupies a predefined position and orientation above the winding cylinder as shown in Fig. 5. While the winding cylinder rotates, the last two turns of the rubber seal are pressed against each other. This creates a winding force between the winding tool and the rubber profile, which can be expressed as a force vector in the flange coordinate system. The resulting force will determined by evaluating the relevant torques on the robot axes.

The practiced force is oriented toward the rotation axis of the winding cylinder (Z-axis of the winding cylinder coordinated system). The force measurement in the KUKA LBR iiwa [4] is composed of the measured torques of all seven axes, which significantly increases the uncertainty of measurements received.

Analysis of both the robot’s configuration and the experimental has lead to the conclusion that the seventh axis has always a lever arm with respect to the reaction position of the exanimated force. Thus, the measurement data of the torque sensor integrated in the seventh axis (M7) will be considered as a control variable to the designed feedback controller. The designed external force feedback controller calculates the pose-offset of the flange pose of the robot in order to achieve the desired torque or torque interval. Thus, the offset of the robot pose is considered as control input of the controller.

A comparison between the determined force and the measured torque values will be discussed in Sect. 3.3. When the linear winding is used, a small gap will be developed at the ends of each winding layer due to the dimensions of the winding tool. Thus, it is useful to divide the winding process into two types (see Fig. 6). The linear winding can only be implemented far from the ending of the winding layer. In contrast, winding by rotating the robot hand can be implemented at the ends of each layer to avoid the developing of gaps.

The robot controller represents the orientation of the robot hand using the Euler angles convention, specifically the roll-pitch-yaw convention. These angles are well known in aerospace engineering and computer graphics [5]. These conventions are numerically unstable [6], moreover, they are inconvenient in practice [7].

The interpolation between the two predefined poses will cause an unpredictable path if the roll-pitch-yaw convention is used. Therefore, the interpolation must be implemented by the quaternion, so that the rotation is described by a rotation about only one axis in the space. A program is developed and implemented on the robot controller in order to convert the roll-pitch-yaw angles to the four quaternion parameters and vice versa. The interpolation is firstly solved at the quaternion level. Secondly, the quaternion parameters are converted back to the roll-pitch-yaw angles. Thirdly, the robot movement is implemented.

In terms of control engineering, an attempt was made to solve the complex winding task by means of a simple closed control loop. The controlled system is given by the part of the plant where the control element acts on a physical quantity to be controlled (the controlled variable) [8]. In the presented case, the controlled variable is formed by the contact forces between the last two windings of the rubber seals. The control element in the control loop is often a technical system which influences the controlled system and thus the controlled variable by changing its output variable(s). This function is carried out here by the KUKA iiwa. By moving the robot flange or TCP parallel to the Z-axis of the roll coordinate system, the force between the last two windings can be influenced. The robot pose is thus the manipulated variable for the control process. The Kuka iiwa also functions as a measuring system, which is able to measure and output both the torques in its seven axes and the forces acting on the flange or TCP in a user-defined coordinate system. Therefore, a control based on the torques and the forces is possible.

Since the influence of disturbance variables is significant in the winding process and the exact attainment of the reference variable is irrelevant, it was initially decided to start with a simple P-controller and then to expand or modify it as required. However, initial tests with this system showed early on that the winding process does not function stably with only one simple P-controller. This was therefore coupled with a three-point controller [9].

Three-point controllers are often used for temperature control, but are also used in the field of actuators. Here, they are used to switch between clockwise and counterclockwise rotation [10]. Such a controller is also recommended for processes in which the controlled variable is subject to noise due to disturbances around the set point (reference variable) [10]. This characteristic should also stabilize the presented winding process, which is permanently affected by disturbances due to the rotation of the winding roll. The resulting characteristic of the new controller is determined by multiplying the characteristic of the three-step controller by that of the P controller (see Fig. 7). The resulting dead band around the center (setpoint) of the characteristic serves to raise the response threshold of the controller. Therefore two reference inputs are defined depending of the set point and the dead band. An oscillation of the process due to noise or measurement uncertainties is also prevented by the hystereses included in the three-point controller (see Fig. 8).

However, the process described here also works for the opposite direction. If the winding forces are too low, the controller calculates a correction of the tool position using the P component P2 (<0). In this way, the current winding is again pressed more strongly against the last winding, which increases the winding forces in turn.

In terms of quantity, P1 is significantly larger than P2. This is due to the fact that the disturbance variable (rotation of the roll, winding of the rubber) always occurs on one side only and the winding force is thus constantly increasing. The controller must therefore react faster to reduce the winding forces than to build them up, which requires a larger P component.

Robot systems are usually equipped with measuring systems to realize the required tasks. The measuring tasks within the winding system can be divided into three groups: First, the monitoring of the external force/torque vector in order to implement the required sensitivity; Second, the measurement of the components and tools; Third, the monitoring of the winding process and the 3D inspection of the winding image.

Requirements for the Measuring System (Force/Torque). Sensitive winding requires an exact measurement of the external force/torque. The robot as a component of the process is set as the control unit of the control circuit. The monitoring of the external force/torque will be carried out by the torque sensors integrated in the LBR iiwa. These sensors are integrated in the joints of the robot. The investigation of the LBR iiwa as a measuring robot in the assembly process for the detection of external force and torques is presented in [11]. The wound rubber profile causes a winding force along the rotation axis of the winding cylinder. This results in a torque around the seventh axis of the robot that can be detected by the integrated torque sensor. The resulting force with respect to the flange coordinate system is then determined in the KUKA controller based on the torque measurement of the 7th axis (see Fig. 9). These are model based calculations, in which the sensor data is filtered in the first step. The torque measurement data, on the other hand, is made available as unfiltered raw data.

The robot will first move linearly from the left side of the hole to the right side without the presence of the winding force. This path lies along the winding cylinder and demonstrates a possible process path. Based on the TCP/IP connection, an interface for the transmission of forces/torques and robot data is developed. For this dynamic case, the measured torques and the determined forces are recorded as a function of the robot position. The result can be seen in Fig. 10. This shows that the measured torque values are noisy. Force values, however, are not noisy, but they show peaks in certain poses. The two data sets (forces and moments) are not constant along the path and depend on the robot’s position. This creates a great need for a low-pass filter to smooth the noisy measured torques and for a look-up table in order to compensate the instability in the obtained data.

A Butterworth filter, 2nd order low pass filter, is used to smooth the data. The filter is written in Java and integrated into the Sunrise controller to filter the measurement data. The resulting smoothed trace is shown in Fig. 11.

The peaks of the forces must also be considered. In addition, the dispersion of the determined data is monitored and graphically displayed. This shows that the dispersion at the peaks is very large (up to 14 N) (see Fig. 12). In this pose, the robot is closer to a singularity that is responsible for the inaccuracies of the forces. The calculation of the forces is based on the kinematic and dynamic model of the robot. Thus, singluarities affect the precision of the aforementioned calculations.

The accuracy of the determined data is also considered. For this purpose, a mass is mounted on the tool by means of a cable via a deflection pulley. The robot again moves along the previous path. The forces and torques are determined and compared with the previous measurements regarding the robot position. A difference is formed to determine the external loads. The difference curve is constant in the torque measurement. In contrast, the difference of the forces decreases (see Fig. 13) close to singularities. Therefore, the torque measurements are more stable and can be used as control variable. The external torque is determined as the difference between the filtered measured torque and the relevant torque from the look-up table based on the robot pose. The external force is then determined by multiplying the external torque with the lever arm.

Calibration of the Winding Cylinder. The measurement of robotic tools, sensors and objects with respect to the robot is a basic task that must be performed during commissioning or after reconfiguration of a robot system. The designed tool must be mounted on the robot flange and measured with respect to the flange coordinate system. This is done by applying the 4-point method, which is defined in the KUKA controller. A Tool Center Point (TCP) must be defined on the tool as the mechanical interface between the tool and the working object [12]. The TCP is then moved four times from different orientations of the robot hand to a predefined measuring point. The position of the TCP with respect to the flange coordinate system can then be determined using the least squares methods. The orientation of the TCP coordinate system is then freely selectable.

The measurement of the work object is an essential part when developing robotic applications. The measurement of the winding cylinder with respect to the robot coordinate system is a prerequisite for the execution of the winding process. The geometry of the winding tool and the high accuracy requirements in the winding process do not allow the measuring of the winding cylinder by teach-in procedure. This problem can be solved with the help of non-contact sensors such as a camera or a laser line sensor. In addition, non-contact sensors offer shorter commissioning times than the teach-in method. Furthermore, the resulting winding pattern must be monitored and evaluated. Gaps between two wound profiles and the superposition of two profiles must be detected and localized. This solves the inspection problem and is discussed in Sect. 3.3. Cameras encounter difficulties in solving this inspection task because the winding pattern is a result of several black profiles. Therefore, the laser line sensor is used in the winding process.

The winding cylinder is at first roughly positioned with respect to the robot. Then it is measured precisely. A cylinder is mathematically defined by a length and a circle. The latter is defined in 3D space by a center, a radius and a normal vector. The calibration error of the sensor, the manufacturing tolerances of the cylinder and the robot accuracy lead to the fact that the circle must be over-determined interpolated. This means that at least four positions are required to measure the circle. The points are measured by the laser line sensor. The robot moves towards one of the two ends of the cylinder. The sensor is monitoring the surface and sends a signal to the robot as soon as an edge is detected. This procedure is repeated at least four times, but with different poses. The circle is fitted after collecting the measuring data of the edges. Then the position and the orientation of the circle with respect to the robot base is determined. After that, the robot moves in the opposite direction and stops as soon as the second edge is detected. Then the length of the cylinder is determined. The best fit method for calibration the winding cylinder is briefly discussed in [13].

The translational speed of the profile is determined based on the winding speed and the winding radius. The winding speed can be monitored by the robot controller. The winding radius is calculated from the shortest distance between the robot position and the axis of the measured winding cylinder. The already wound length of the profile between two successive measurements is determined by multiplying the translational speed by the time difference of the measurements. Thus, the 2D measured values can be extended to the 3D space and the winding image can be visualized and analyzed in the 3D space. Since only a section of the rubber profile can be recorded in the static state due to the limited measuring range of the laser line sensor, it is necessary to record the measurement data along the winding cylinder. For this purpose, the data is transformed from the sensor coordinate system into the robot base coordinates during the winding movement. The basic prerequisite for the data transformation is a measurement of the sensor with respect to the robot flange coordinate system. The robot sensor calibration problem is solved in [14]. The entire winding pattern results from the overlapping of the different winding areas (see Fig. 14). The winding pattern of each winding cylinder is documented. The defects can be localized and quantified.