AES_8

tecnica 95 CONTROLLO Automazione e Strumentazione n Novembre - Dicembre 2021 dard time needed to execute basic motions is reported. There are some criticalities. It is known [1], for exam- ple, that 30 to 45 minutes of MTM-UAS analysis are required for every minute of action observed. On the other hand, direct observations imply focusing on few tasks at a time and are usually not well considered by workers, who feel under evaluation and change their normal behavior, therefore. An automatic sampling could lead to lower costs, higher volumes of data per time unit and lower impact on workers. It would also free engineers and techni- cians by low value data entry tasks, letting them focus on higher value activities like identifying the causes of waste and redesigning the processes. One problem emerges when automatizing the sam- pling of manual processes. Every automatic system is designed to detect and react to a certain set of signals, that may assume the same set of values (i.e. come out as ambiguous) for two or more different activities to be sampled. To conduct an automatic detection of manual processes this ambiguity must be overcome. That is why, for example, many camera-based automatic solu- tions use ML (Machine Learning). Engineers train such systems with Supervised approaches, thus making them able to classify the images they acquire. Due to their context-based training, though, this type of solutions lack in flexibility and scalability. Resolving the ambiguity By a theoretical point of view every process can be seen as a finite-state-machine , i.e. a system that may be in exactly one of a finite number of states at any given time. Knowing the list of states and events that can cause a change of state, an automatic system would be able to sample the activities. Given the scope of the analysis (industrial manual operations), events and statuses are assumed to be reported within technical documents like procedures or work instructions. Such documentation is therefore analyzed by a semi-automatic NLP (Natural Language Processing) algorithm to extract the afore- mentioned list of states. The variables used (i.e. the space of representation) to describe each state are: position of the worker within the workstation, position of the tools, functioning levels of the tools (expressed with the most convenient metric, depending on the case, e.g. current amplitude, spindle rotation speed, etc.). Once the list is completed, it is pos- sible to check how many variables need to be consid- ered for a non-ambiguous representation of the process. In other terms, if two or more statuses assume the same value along a specific set of variables, a new variable (i.e. new sensors) must be considered. Each adding variable carries redundant information that reduces the ambiguity. As an example, a set of wearable RFID readers ( Figure 1 ) and passive tags can be used to determine the position of the worker. Each tag identifies a specific check-point within the workstation, i.e. a position that the worker must pass from during the regular execution of his/her tasks. In other terms, once a tag is detected, the sys- tem knows that the position of the worker has changed determining a change in the process’ status. The same concept can be applied as well to the level of current consumption by an electric screwdriver, indicating the transition from ‘in use’ to ‘not in use’ status. Typical scenario Assembly cycles are usually described in very sequen- tial and standardized terms. ‘Pick A and place it in B’ tasks usually cover large part of the cycles, thus mak- ing the analysis much effortless. Since A is the object to be placed and B its final position, it is not difficult to build the list of states and their related values along the ‘position of the worker’ variable. This way the setting of the system reduces to simply placing passive tags on the picking spots within the workstation. An example is given by the set-up implemented at a wire manufactur- er’s assembly bench ( Figure 2 ). Figure 2 - example of a typical set-up. support flexible arms have been used to accurately position and shield every passive tag (here presented as white cards), thus incrementing the set-up reliability