**1. Introduction**

### *1.1. Theoretical Background*

In industrial production, as manufacturing systems have become more complicated, the concept of operations and maintenance (O&M) has become important in preventing unforeseen faults and errors and for smooth operations [1,2]. Lin et al. (2021) found that O&M investments are continuously increasing in the power plant industry owing to concerns regarding the safe and reliable operation of power plants [3]. Accordingly, the authors investigated the major influencing factors in the O&M of power plants by constructing a fishbone diagram and analyzing the gray correlation between the derived factors. In addition, O&M is relevant not only to physical products or systems but to cyber systems as well. In this regard, Furumoto et al. (2020) discussed how to assess and prevent cyber risks to achieve reliable ship operation [4]. In manufacturing processes, the prevention of unknown and sudden faults in real time is critical [5,6].

Grippers play an important role in an industrial manufacturing system [7–9]. They can be classified into two types: mechanical type and vacuum type. Mechanical grippers are also known as robotic grippers [10]. As their shape and operating mode mimic human hands and fingers, they are usually used for performing complex or delicate operations. For example, Vedhagiri et al. (2019) proposed a mechanical gripper with five fingers in order to handle objects with complex shapes, such as a coin, cosmetic cream, and fruit [11]. A mechanical gripper with three fingers has also been proposed to conduct micromanipulation [12]. The authors [12] applied a piezoceramic transducer for effectively handling objects as small as 10 to 800 μm, such as glass hollow microspheres and iron spheres.

**Citation:** Baek, S.; Kim, D.O. Predictive Process Adjustment by Detecting System Status of Vacuum Gripper in Real Time during Pick-Up Operations. *Processes* **2021**, *9*, 634. https://doi.org/10.3390/pr9040634

Academic Editor: Luis Puigjaner

Received: 9 March 2021 Accepted: 1 April 2021 Published: 5 April 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

On the other hand, vacuum grippers utilize suction cup(s), air pipe(s), and Venturi line(s) to handle objects, and they require an external air compressor to supply compressed air, as illustrated in Figure 1. In particular, a discrete manufacturing process is typically performed by several common operations using vacuum grippers, and the grippers are widely used for lifting, transporting, and inspections [13–15] as they are simple, inexpensive to install, easy to operate with fragile objects, and can handle various types of object shapes [16,17].

**Figure 1.** Illustration of a vacuum gripper system (composed with a suction cup and the connected Venturi line).

For example, a vacuum gripper with four suction cups placed in a two by two arrangement has been proposed for handling heavy objects with complex shapes [18]. Amend et al. (2012) developed a positive pressure gripper that uses a balloon instead of a suction cup and can handle objects of various shapes [10]. Since this universal gripper can utilize both positive and negative pressure, it showed good performance in gripping objects of different size, shape, weight, and fragility. Reinforcement materials such as magnetorheological elastomers have also been utilized in suction cups of vacuum gripper systems to increase their performance [19]. As can be deduced from the above, vacuum grippers with one or more suction cups are frequently used to handle different types of objects. For example, Nakamoto et al. (2018) validated that a vacuum gripper can handle objects of 40 different shapes, sizes, and materials without requiring any changes in the gripper system setup [20].

As above mentioned, several studies have focused on how to effectively use a vacuum gripper and how to prevent a faulty gripper operation in advance. Operational parameters that affect a manufacturing process are usually categorized into machine parameters and product parameters. Whereas product parameters pertain to the production of the handled object in the process (such as the three-dimensional lengths of the object), machine parameters are related to the motion control of the installed actuators (for example, the speed of actuator movement). In general, for the effective execution of manufacturing processes, appropriate machine parameters are usually analyzed, instead of product parameters, because product parameters are already optimized before being fed into the manufacturing system. For example, Wang et al. reported that the type of pipe material and operating temperature can determine the suction power during gripper operation [21]. The authors [21] also investigated the effects of cup size and length of suction time on the gripping operation. According to another study [22], the performance of the gripping operation can depend on the shape of suction cups (universal, flat, or deep). In addition,

considerable research has been performed on the effects of machine parameters, including suction cup size [23,24], number of cups [25], and initial applied pressure, on the suction operation [26].

However, these investigated machine parameters are usually "design variables" that are determined in the design phase of the product life cycle; thus, they cannot be easily modified later during production. In other words, it is difficult to adjust these parameters during each gripping operation in real time. Therefore, many studies have alternatively focused on approaches to detect defects in objects, such as dents or ripped sections on the objects, before the gripping operation because the prior detection of defects enables an error-free operation. For box-lifting operations, ultrasonic sensor signals have been utilized to identify the relative curvature of box surfaces [27]. Similarly, convolutional neural network models based on sensor signals have been proposed to classify normal and defective box surfaces [27]. Moreover, surface cleanliness has also been found to be important in the effective use of grippers [28]. The authors determined that the suction performance when using vacuum grippers can be determined by the number and size of contaminants on target surfaces. Since the condition of objects being handled cannot always be changed during production operations, the only feasible solution is to remove the defective objects in advance. In other words, it is still difficult to adjust certain conditions or parameters for recovering a fault state.

Therefore, this paper proposes a real-time predictive process adjustment method for pick-up operations using vacuum grippers. To this end, a pick-up operations testbed equipped with a two-way data monitoring and process parameter control module was developed. To analyze sensor signals in terms of faulty operation detection, a sensing module was devised by installing an air pressure sensor inside the Venturi line of a vacuum gripper, and the significant features that can help distinguish between normal and faulty gripper operation before the completion of the current pick-up operation or start of the next operation were characterized. Finally, an experiment to determine whether it is possible to recover faulty states by adjusting relevant process parameter(s) was performed.

The remainder of this paper is organized as follows: In Section 2, the developed system configuration to acquire appropriate sensor signals and provide feedback commands to the controller is explained in terms of hardware and software. In addition, air pressure sensor signals to obtain meaningful features that can help detect faulty operations in advance are analyzed. The experimental results are summarized in Section 3. The concept of predictive process adjustment in the pick-up operation of a vacuum gripper by controlling a related control parameter is also demonstrated. Lastly, in Section 4, the study is summarized and the scope for future work discussed.

#### *1.2. Literature Review on Fault Detection Methods*

Fault detection and diagnosis (FDD) approaches based on traditional statistical process control charts as well as advanced data mining techniques have been employed in various manufacturing applications [5,29–32]. As summarized in Table 1, statistical chart-based FDD approaches provide a satisfactory detection performance when sensor signals are only collected during the normal operation of a system, and the corresponding measurements are concentrated in one or several clusters. Principal component analysis (PCA) is a popular example of a fault diagnosis method based on traditional statistical process control charts where multivariate analog sensor signals are simultaneously collected from a manufacturing process [33–35]. This method reduces the dimensionality of original historical data according to eigenvectors decomposed via singular value decomposition [33]. Kim et al. (2020) employed PCA to derive a residual control chart to monitor the current system status [36]. Since highly correlated and dimensional sensor signals were fed as input data, the authors were able to extract hidden but meaningful features by combining traditional PCA with functional PCA. Since traditional PCA deals with linear relationships in the given multivariate dataset and functional PCA contains non-linear eigenfunctions and can

handle highly correlated multivariate sensor signals, the derived features were used as indicators of process control charts; this yielded a superior fault detection performance.

**Table 1.** Classification of fault detection and diagnosis (FDD) approaches.


By contrast, data-mining-based FDD approaches are usually utilized when the sensor signals are excessively scattered regardless of the system operation state (normal or fault). Due to this characteristic, it is not easy to develop a robust representative mathematical model for quantifying the state of the system. That is, advanced machine learning techniques have been frequently applied when the collected sensor signals do not have distinguishable features between the normal and faulty system states [37]. In order to detect unknown faulty statuses more accurately, pattern classification with discrete state vectors (DSVs) was improved based on a similarity analysis [38]. Since previous DSV-based fault detection methods showed good performance when every DSV is discernible based on training data, the authors used Naïve Bayes approximation and the Brier score to evaluate unknown DSVs in terms of fault detection power. Based on the proposed technique, the faulty operation of a vehicle engine could be detected more precisely.

As the above approaches explained, most O&M studies have focused on offline fault diagnosis [5,35,36]. Online fault detection is usually performed by comparing the current operation state and the offline fault detection model, and consequently, it gives the user alarm when the current state is considered as a fault state (or, a fault state will occur in near future. For effective manufacturing operations, corrective recovery action(s) should be manually performed by operators to prevent the detected faults. For example, for the effective operation of an automated storage and retrieval system (ASRS), Internetof-things-based controllers and sensors were installed [31]. Depending on the current system status determined by analyzing real-time vibration signals, the controller changes the relevant process parameter (i.e., motor speed in transporters) and, consequently, the ASRS can automatically maintain a failure-free status. In this regard, O&M approaches have evolved from corrective maintenance, through preventive maintenance and conditionbased monitoring, to predictive maintenance [6,39]; however, it is still not easy to combine real-time fault detection/prediction and automatic process adjustment.

#### **2. Materials and Methods**

#### *2.1. Materials: System Configuration for Monitoring Pick-Up Operations in Real Time*

To identify the key machine parameters and test the predictive process adjustment with the corresponding characteristics, a testbed that performs a pick-up operation using a vacuum gripper [27], as depicted in Figure 2 was constructed. Furthermore, the monitoring and control modules were upgraded, as shown in Figure 2b. The detailed information of the system configuration is summarized in Table 2.

**Figure 2.** Testbed for performing the pick-up operation and monitoring the amount of applied air pressure during each operation cycle: (**a**) front view; (**b**) top view; (**c**) closed view for a suction cup.



The outlet air pressure was considered as an indicator to identify the current pick-up operation (i.e., suction operation) in real time, instead of the inlet air pressure, which is most commonly measured. Among various machine parameters, the outlet air pressure is the most appropriate variable because it changes depending on whether the suction is turned on or not. In a vacuum gripper, compressed air is supplied from an air compressor to the suction cup through a Venturi line. The Venturi line is designed to increase airflow speed by changing the size of the air pipe [40]. When the compressed air moves from the entrance to the exit of a Venturi line, a partial vacuum state occurs at the jet nozzle (an additional narrow vertical air pipe between the entrance and exit of the Venturi line). Regardless of whether a partial vacuum state was created, it was assumed that the magnitude of the outlet air pressure changed.

For predictive process adjustment according to real-time operation monitoring, the control modules and the corresponding control software were upgraded, as shown in Figure 3. Due to the upgrade, it was not only possible to acquire sensor signals in real time but also to simultaneously change machine parameters, such as the x-, y-, and z-positions corresponding to the suction start, during the operation. The workflow is as follows:


**Figure 3.** Software to control the testbed, collect sensor signals in real time, and send feedback control messages.
