Artificial awareness for robots using artificial neural nets to monitor robotic workcells
Abstract
Current robotic systems are unable to recognize most unexpected changes in the work environment, such as tool breakage, workpiece motion, or sensor failure. Unless halted by a human operator, they are likely to continue actions that are at best inappropriate, and at worst may cause damage to the workpiece or robot. This project investigated use of Artificial Neural Networks (ANNs) to learn the expected characteristics of sensor data during normal operations, recognize when data no longer is consistent with normal operation, suspend operations and alert a human operator. Data on force and torque applied at the robot tool tip were collected from two workcells: a robotic deburring system and a robot material-handling system. Data were collected for normal operations and for operations in which a fault condition was introduced. Data simulating sensor failure and excessive sensor noise were generated. Artificial Neural Networks (ANN) were trained to classify operating conditions; several ANN architectures were tested. The selected ANNs were able to correctly classify all valid operating conditions and the majority of fault conditions over the entire range of operating conditions, having {open_quotes}learned{close_quotes} the expected force/torque data. Most faults introduced appreciable error in the data; these were correctly classified. However, a smallmore »
- Authors:
- Publication Date:
- Research Org.:
- Sandia National Lab. (SNL-NM), Albuquerque, NM (United States)
- Sponsoring Org.:
- USDOE, Washington, DC (United States)
- OSTI Identifier:
- 469142
- Report Number(s):
- SAND-97-0957
ON: DE97005328; TRN: 97:002945
- DOE Contract Number:
- AC04-94AL85000
- Resource Type:
- Technical Report
- Resource Relation:
- Other Information: PBD: Apr 1997
- Country of Publication:
- United States
- Language:
- English
- Subject:
- 42 ENGINEERING NOT INCLUDED IN OTHER CATEGORIES; 99 MATHEMATICS, COMPUTERS, INFORMATION SCIENCE, MANAGEMENT, LAW, MISCELLANEOUS; ROBOTS; NEURAL NETWORKS; ALGORITHMS; ARTIFICIAL INTELLIGENCE; TRAINING
Citation Formats
Tucker, S D, and Ray, L P. Artificial awareness for robots using artificial neural nets to monitor robotic workcells. United States: N. p., 1997.
Web. doi:10.2172/469142.
Tucker, S D, & Ray, L P. Artificial awareness for robots using artificial neural nets to monitor robotic workcells. United States. https://doi.org/10.2172/469142
Tucker, S D, and Ray, L P. 1997.
"Artificial awareness for robots using artificial neural nets to monitor robotic workcells". United States. https://doi.org/10.2172/469142. https://www.osti.gov/servlets/purl/469142.
@article{osti_469142,
title = {Artificial awareness for robots using artificial neural nets to monitor robotic workcells},
author = {Tucker, S D and Ray, L P},
abstractNote = {Current robotic systems are unable to recognize most unexpected changes in the work environment, such as tool breakage, workpiece motion, or sensor failure. Unless halted by a human operator, they are likely to continue actions that are at best inappropriate, and at worst may cause damage to the workpiece or robot. This project investigated use of Artificial Neural Networks (ANNs) to learn the expected characteristics of sensor data during normal operations, recognize when data no longer is consistent with normal operation, suspend operations and alert a human operator. Data on force and torque applied at the robot tool tip were collected from two workcells: a robotic deburring system and a robot material-handling system. Data were collected for normal operations and for operations in which a fault condition was introduced. Data simulating sensor failure and excessive sensor noise were generated. Artificial Neural Networks (ANN) were trained to classify operating conditions; several ANN architectures were tested. The selected ANNs were able to correctly classify all valid operating conditions and the majority of fault conditions over the entire range of operating conditions, having {open_quotes}learned{close_quotes} the expected force/torque data. Most faults introduced appreciable error in the data; these were correctly classified. However, a small minority of faults did not give rise to a detectable difference in force and torque data. It is believed that these faults could be detected using other sensors. The computational workload varies with the implementation, but is moderate: up to 2.3 megaflops. This makes implementation of a real-time workcell monitor feasible.},
doi = {10.2172/469142},
url = {https://www.osti.gov/biblio/469142},
journal = {},
number = ,
volume = ,
place = {United States},
year = {Tue Apr 01 00:00:00 EST 1997},
month = {Tue Apr 01 00:00:00 EST 1997}
}