*2.2. Methods*

Predictive process adjustment is a representative approach in predictive maintenance. The concept of this approach is to automatically perform appropriate feedback actions, which are determined by detecting and diagnosing the current system status or predicting an upcoming fault status through real-time monitoring [31]. Therefore, in this paper, early fault detection and recovery actions are proposed in order to conduct predictive process adjustment in real time, as illustrated in Figure 4.

**Figure 4.** Framework of the predictive process adjustment method based on real-time data acquisition and feedback control.

First, the relevant measurements are acquired from the testbed when a vacuum gripper performs a pick-up operation according to the controller's command. During the training phase, the time-series sensor signals for every cycle are accumulated in a history dataset, which is employed to generate a fault detection model. To develop a fault detection model that can identify an unknown and sudden fault situation in advance, a detection threshold should be determined. Since sensor signals were already collected, it is possible to construct fault detection model using several fault detection/prediction approaches, from conventional statistical models to pattern extraction-based deep learning models. In this study, the simplest statistical-based fault detection model was employed using Equation (1) [41]:

$$\text{Detection threshold} = \overset{=}{X} + \frac{\mathfrak{Z}}{d\_2\sqrt{n}} \times (\overline{\mathbb{R}}),\tag{1}$$

where *X* is (approximately) the mean of the sensor signals a short duration after the suction command is given to an actuator in normal operation; *R* is the mean of the differences of maximum and minimum measurements in the same dataset, which is used for calculating =

*X*; and *d*2 is a constant value chosen by the number of sensor measurements (i.e., *n*).

=

For a statistical process control approach, the detection threshold is computed according to the *X*-chart. The *X*-chart is one of the control charts that is used for identifying whether measurements in the current batch are under the statistical control or not. It can be categorized as unsupervised learning, as history datasets collected during normal

operations are used to develop a detection model, such as those based on *X* and *R* computation. The *X*-chart was utilized in this study because the time-series sensor signals collected during one pick-operation operation are considered as a dataset in one batch. In this regard, a series of sensor signals is collected for a short duration after the suction starts (during one operation cycle) in order to assume that the collected measurements are under identical conditions.

In common process control chart models [41], the upper and lower control limits are applied together to determine the statistical outliers. However, in this study, the direction of the failure status was clearly determined (e.g., a higher value indicates a failure of vacuum generation, and hence, an operation failure); therefore, only the upper control limit was used as the detection threshold. Using the derived detection threshold, the failure status of the pick-up operation was detected when the following condition is not satisfied:

$$
\lambda x\_i \le \alpha \times \text{Detection threshold}\_\prime \tag{2}
$$

=

where *xi* is the current sensor value and *α* is the predefined confidence interval.

If a fault is detected during real-time monitoring based on the above detection model, the appropriate resolution strategy is consequently developed. In this study, the machine parameter was carefully controlled in order to overcome situations with faults.

#### **3. Experimental Results**

#### *3.1. Result: Early Fault Detection for Pick-Up Operation*

Using the testbed described in Section 2, a pick-up experiment was conducted to determine the characteristics of the collected sensor signals depending on the success or failure of the suction operation. The target operation was to lift rectangular boxes using one suction cup. To generate two different states (i.e., success and failure to lift), the following two different types of boxes were used, as shown in Figure 5:


**Figure 5.** Target objects: (**a**) two samples of box type I; (**b**) two samples of box type II. (They do not show any significant difference in 2D images).

Both types of box cover, box type I and II, are considered as normal quality in product quality inspection phase. Although box type II should be used in the manufacturing process from the viewpoint of product quality, because of its very negligible concave on the contact surface, it results in a faulty pick-up operation. Importantly, even if the manufacturer attempts to install a system to remove box type II from the production line, it is highly difficult to distinguish between the two types using the naked eye or a 2D camera, as depicted in Figure 5.

The lifting operation was conducted 20 times for each box type, and the corresponding sensor signals were collected, as illustrated in Figure 6. The generated historical data were recorded with the corresponding timestamps. The duration was approximately 10 s, and the analog signals were recorded at a sampling rate of approximately 10 Hz. In addition, missing or noisy data in the collected raw signals were filtered using adjacent measurements.

Figure 6a describes the pressure signals of the airflow in the Venturi channel for every lifting iteration. Before starting the suction operation using the vacuum gripper, it was ensured that the initial pressure of the airflow in the Venturi line was almost identical to the original air pressure provided by the air service unit and air compressor. After starting the suction operation (before lifting the object), it can be observed in Figure 6a that the monitored air pressure decreased continuously (at approximately 5.6 to 6.0 s). When the suction was sufficiently succeeded, the decreased pressure was constantly maintained during the rest of the pick-up operation, until moving to the destination position, and the decreased level was identical in the 20 repetitions. Although certain fluctuations in the inlet pressure can be observed during the pick-up operation, the decrease in pressure was mostly constant until the suction operation was terminated. Signal collection automatically stopped after the robotic arm was moved with the vacuum gripper to the destination position and the vacuum switch was turned off to release the object.

**Figure 6.** *Cont*.

**Figure 6.** Measured air pressure over time: (**a**) successful lift for box type I; (**b**) failed lift for box type II.

For box type II, the gripper failed to pick-up the box in all repetitions, as listed in Table 3, even though the conditions were kept identical to those for box type I. As depicted in Figure 6b, the air pressure starts to decrease when the suction operation starts, but the level of decrease is different from that for box type I. As a result of the negligible concave curvature on the contact surface, it is not possible to generate sufficient suction, and the amount of outlet air pressure cannot decrease sufficiently for a successful gripper operation. In Figure 6b, the sensor signal for Trial 1 is slightly different from those for the other 19 trials; this is due to an inadvertent early manual control suction start command by the operator.

**Table 3.** Pick-up operation results.


In conventional operations with a vacuum gripper, a vacuum switch determines whether the suction operation was successful after the pick-up operation has been conducted. In other words, although the vacuum switch does not indicate that a sufficient vacuum has been successfully generated, the gripper (and the connected robotic arm) moves to the next position regardless of the completion of the current pick-up operation. However, upon using the information derived from this study, extraneous movements or operations without the target object can be prevented.

In summary, the process can be summarized as follows:


3.2. If the air pressure does not decrease to the desired level, then conduct the predictive process adjustment until the object is lifted by the gripper.

#### *3.2. Discussion: Conducting Appropriate Recovery Actions*

After predicting the outcome of a pick-up operation (as already described in Section 3.1), it is necessary to identify which control parameter(s) can result in a failed pick-up operation. There are several control parameters whose value changes during the gripper operation, such as the initial applied pressure, length of suction time, and suction position [26,27]. In this study, the vertical position of the vacuum gripper at the suction start (hereafter called as the "z-position") was selected as the control parameter for the predictive process adjustment for the pick-up operation. The z-position is illustrated in Figure 7: it represents the bottom position of the suction cup attached on a vacuum gripper. In the case of more than one suction cup, the same z-position is recommended for all suction cups.

**Figure 7.** Suction cup positions for predictive process adjustment: (**a**) axis information for suction cup control; (**b**) diagram for the 0.0 mm z-position; (**c**) diagram for the −0.5 mm z-position.

Experiments were conducted to investigate the effect of the z-position on the pick-up operation, as summarized in Table 4. The pick-up operation consisted of the following consecutive steps: movement to the suction location, suction start, upward movement, movement to the destination, and suction finish.


**Table 4.** Experimental setup to determine effect of z-position on the pick-up operation.

Once the command was sent to the vacuum gripper, the sensor signals were collected in the form of a time series from step 2 to step 3. A total of 40 time-series datasets were recorded. Using mean values corresponding to approximately 5.6 to 6.0 s for each timeseries dataset, a one-way analysis of variance (ANOVA) was conducted to determine the statistical effects of the dependent variable on the independent variable. ANOVA is a typical statistical test to determine the source of measurement differences, that is, whether the differences result from variance between groups (levels) or from variance within groups corresponding to independent variables [42]. It can be treated as the extended version of the t-test for more than three groups corresponding to an independent variable. This statistical test is prevalently employed to determine the effects of controllable parameters on the output [43–45].

The results are summarized in Table 5, where the z-position shows a significant difference in the pick-up operation (*p*-value < 0.05). As can be observed from Table 6, the success ratio of the operation increased from 0% to 100% as the vertical position at the suction start lowered. Specifically, the suction cup failed in nine trials except in the case of Trial 5 (–1.0 mm), whereas it failed in only three trials for box type II (i.e., Trials 1, 2, and 6) with the z-position at –1.5 mm. Finally, when the z-position was set as –2.0 mm, the gripper succeeded in picking the target up in every trial. Clearly, the outlet air pressure decreased by more (approximately 200 to 300 V) than the decrease in the cases of operation failures (approximately 600 to 700 V), and the outlet air pressure was continuously maintained until step 3, as depicted in Figure 8.

**Table 5.** ANOVA results of the effect of the z-position on the pick-up operation of the gripper.


\* SS: sum of squares, df: degree of freedom, MS: mean squares.

**Table 6.** Experimental results of the predictive process adjustment by controlling the z-position of suction start with box type II.


**Figure 8.** Collected air pressure signals depending on the z-position (during pick-up operation with box type II): z-position is set as (**a**) −1 mm; (**b**) −1.5 mm; and (**c**) −2.0 mm.

In summary, the predictive process adjustment with real-time operation monitoring can be conducted as follows:


However, as the z-position is further lowered, the probability of defects on the contact surface increases. A vacuum gripper that lowers excessively may result in a dent or a hole, and consequently, the object will be considered to be defective right before the pick-up operation. Therefore, it is essential to determine the appropriate z-position that does not cause any damage to the surface of the object while maximizing the success rate of the pick-up operation.
