Detection and Monitoring for Anomalies and Degradation of a Robotic Arm Using Machine Learning

Supervised Machine-Learning Methodology for Industrial Robot Positional Health Using Artificial Neural Networks, Discrete Wavelet Transform, and Nonlinear Indicators

“…In addition, the proposed method is helpful in quickly identifying deviating joints (80 s approximately), especially in a serial robot arm, where the error accumulates along the kinematic chain. In contrast with other works that share the use of the same database, such as [ 21 ], where the residual error of the TCP is addressed, or [ 24 ], where a handmade comparison is made between the position and velocity of the joints to detect deviation; the proposed method can identify, in an automated way, in which part of the robot is the deviation, it does not require modeling of the robot or external sensors to perform measurements, and the method is applicable to robots of different degrees of freedom. The early identification of the deviating joint allows the implementation of robot calibration methodologies and the generation of diagnostic strategies for preventive maintenance of the robot, reducing maintenance times and costs, both valuable resources in the industrial field.…”

“…In addition, the requirement that the robot movement has to be the same over long time periods is a limitation for database selection. For example, the method reported in [ 21 ], where the same dataset used in this work is used, addresses the residual error of the robot’s TCP and does not consider the individual analysis of the joints; therefore, only a qualitative comparison can be made. In this sense, the proposed method can identify in which part of the robot there is a deviation that directly affects the precision of the TCP due to the accumulation of the error through the kinematic chain.…”

“…The robot position was measured by means of an artificial vision system, designed for this purpose and presented in [ 20 ]. In [ 21 ], a methodology based on deep long short-term memory (DLSTM) and regression adjustment multivariate (RAM) was proposed, to model and predict (respectively) the displacement of an industrial robot, and estimate the residual error in the tool center position (TCP). A smart target was designed, to be used with the vision system mentioned before, and its feasibility in performing robot accuracy assessment was verified in [ 22 , 23 ].…”

FPGA-Based Methodology for Detecting Positional Accuracy Degradation in Industrial Robots

“…Regarding the results obtained in this work, it should be noted that specifically for joint five of the real case, the results coincide with those reported in [53], where the same dataset was used, and a manual analysis of velocity and position graphs of the joint was conducted without a systematic approach. In another study [17] MRA and DLSTM are used to obtain the residual error of the end-effector of the robot. However, it is not determined in which joint exhibited deviation causing the error in the end-effector.…”

“…On the other hand, there have been works that directly address the problem of positional degradation of industrial robots, e.g., in [3], the dependent errors present in the robot produced by non-ideal deflections of the structure and movement are characterized using polynomials of Chebyshev for an advanced error model. Taha et al [17] use multivariable regression adjustment (MRA) and deep long short-term memory (DLSTM) to predict and model the displacement of an industrial robot and to estimate the residual error on the end-effector. An expert system for detecting a deviation in the joints of a robot is proposed in [18] and is based on DWT analysis, neural networks, and fractal and energy features.…”