*4.2. Sequence Testing*

In order to validate the proposed methodology, the previous experimental setup is used. The first goal in the experiment was to generate the data set of the force/torque signals from the two sensors. The assembly sequence is programmed in the computer to follow the 5 stages described in Figure 1.

As described in the graphical representation of the sequence in Figure 3, the robot arm that holds the peg moves towards the hole at a very slow speed; when the force limit is reached due to the first contact state, the robot will stop the movement and the contact forces are stabilized in a range of 1 or 2 s. With stable values of force/torque signals, the program gathers the information and saves a data file with the values of the error classification and the force/torque vectors from both sensors.

#### *4.3. Data Set Creation*

The database generated had a total of 2056 contact samples. The data set was analyzed using t-SNE (t- Distributed Stochastic Neighbor Embedded) technique to graphically visualize the separability of the features as shown in Figure 7, where it shows how the data can be separated into the 8 classes. As it can be observed, the classes 0, 6 and 7 show very poor separability characteristics, meaning that most of the misclassifications happen on these corresponding positions, meanwhile the other features seem to be separable.

Additional databases were generated in order to have validation data sets. The number of samples for the validation data sets are reported in Table 4 and start from 35 samples to go to up to 240 samples.

**Figure 7.** Representation of the data set and the 8 classes (0–7) by applying t-SNE technique.
