Failure Detection and Classification for Industrial Robots

In industrial robotics, the detection of failures is a key part of the robotsprogram to reach a robust assembly process. However, the setting up of sucha detection is long, very specific to a robotic operation and involves programmingby engineers. In response to this problematic, the thesis presented inthis paper proposes an algorithm which makes it generic and semi-automaticthe creation of a failures detection and a classification method. Based onmachine learning, the algorithm is able to learn how to differentiate betweena success and a failure scenario given a series of examples. Moreover, theproposed method makes the teaching of failure detection/classification accessiblefor any operator without any special programming skills.After the programming of movements for a standard behavior, a trainingset of sensory acquisitions is recorded while the robot performs blindlyoperation cycles. Depending on sensors nature, the gathered signals canbe binary inputs, images, sounds or other information measured by specificsensors (force, enlightening, temperature...). These signals contain specificpatterns or signatures for success and failures. The set of training examples isthen analyzed by the clustering algorithm OPTICS. The algorithm providesan intuitive representation based on similarities between signals which helpsan operator to find the patterns to differenciate success and failure situations.The labels extracted from this analysis are thus taken into account to learna classification function. This last function is able to classify efficiently anew signal between the success and failure cases encountered in the trainingperiod and then to provide a relevant feedback to the robot program. Arecovery can then easily be defined by an operator to fix the flaw