Experimental Evaluation of Dry and Contactless Cleaning Methods for the Production of Digital Vehicle Dashboards

Abstract

Pillar-to-pillar dashboards have become common in modern electric vehicles. These dashboards are made of liquid crystal displays (LCDs), of which backlight units (BLUs) are an integral part. Particulate contamination inside BLUs can lead to either an aesthetic or functional failure and is in consequence a part of quality control. Automatic optical inspection (AOI) was used to detect particulate matter to enable a process chain analysis to be carried out. The investigation showed that a high percentage of all contaminants originated from the assembly of the edge/side lightguide. The implementation of an additional cleaning process was the favored countermeasure to reduce the contaminants. The objective (cleanliness requirement) was to remove all contaminants larger than 100 µm from the lightguide with contactless (non-destructive) cleaning methods. The preferred cleaning methods of choice were compressed air and CO₂ snow jet cleaning. This work investigates the cleaning efficacy of both cleaning methods under consideration of the following impact factors: distance, orientation (inclination) and speed. The central question of this paper was as follows: would cleaning with compressed air be sufficient to meet the cleanliness requirements? In order to answer this question, a cleaning validation was carried out, based on a Box–Behnken design of experiments (DoE). To do so, representative test contaminants had to be selected in step one, followed by the selection of an appropriate measurement technology to be able to count the contaminants on the lightguide. In the third step, a test rig had to be designed and built to finally carry out the experiments. The data revealed that CO₂ was able to achieve a cleaning efficacy of 100% in five of the experiments, while the best cleaning efficacy of compressed air was 89.87%. The cleaning efficacy of compressed air could be improved by a parameter optimization to 94.19%. In contrast, a 100% cleaning efficacy is achievable with CO₂ after parameter optimization, which is what is needed to meet the cleanliness requirements.

1. Introduction

Technical cleanliness emerged 20 years ago in the automobile industry [1]. Residues from manufacturing processes (e.g., casting, milling, grinding, etc.) were responsible for malfunctions or damages to assemblies/systems—often within the warranty period. Initially, the mechanical elements of the powertrain and others were affected (injection system, crankshaft, ABS, etc.) [2,3]. Around 2015, control electronics (airbag, steering, brakes, etc.) followed [4]. In the case of electric vehicles, traction batteries are predominantly affected, with a potentially higher severity [5,6,7].

However, in the era of electrified vehicles, earlier issues have not become obsolete, as the recent findings by Florian Pillau show [8]. The number of recalls and the number of vehicles within a recall due to cleanliness flaws are nevertheless rather low.

This article is about technical cleanliness of a pillar-to-pillar vehicle dashboard [9,10], or, more precisely, its liquid crystal display (LCD) backlight unit (BLU) consisting of an edge/side lightguide [11,12]. Quality controls in series production revealed particulate contamination (sometimes also referred to as foreign matter or debris) under and on the lightguide—also see Nam et al. [13]—which are mainly of organic origin and are rather similar in its appearance to Figure 4d–4f of Chen et al. [14].

As a result of a process chain analysis—based on VDA 19.2, Annex A.B.2 [15]— various particles were categorized in order to allocate their origin with the aim of deriving suitable countermeasures. The process chain analysis is a technique for identifying risks, as illustrated in Figure A.1 of IEC 31010:2019 [16]. Although IEC 31010:2019 provides a good overview of various risk assessment techniques, the focus of ISO 26262-3:2018 is the application of some of these techniques to ensure the safety of electrical/electronic (E/E) systems [17]. Section 6 (“hazard analysis and risk assessment”) of ISO 26262-3:2018 is a task that needs to be completed during the concept phase. In contrast, the process chain analysis is should be completed during series production because data from produced assemblies/systems are required.

The basic idea of a process chain analysis corresponds to the “data-based process monitoring” described by Kong et al. [18]. However, due to the involvement of a human workforce (e.g., material provision, monitoring, maintenance, etc.), it also corresponds to Nancy G. Leveson’s “sociotechnical system” approach with its “system theoretical process analysis” (STPA) [19]. According to Welker et al., people represent the top two of all contamination sources [20]; therefore, their impact is a vital part of a comprehensive monitoring concept. The link to STPA and its underlying hazard analysis in the context of technical cleanliness can be better understood when exceeding a process-related cleanliness requirement/limit is defined as hazard. It is therefore a question of WHETHER a hazard exists (limit violation) in lieu of WHAT kind of hazard exists.

Countermeasures in the context of technical cleanliness refer very well to the risk mitigation claimed in IEC 61508-5:2010 [21]. Its aim—after the processes with the highest impacts to technical cleanliness have been identified [22]—is a reduction towards a tolerable risk. This is the duty of companies and should be in line with the specified cleanliness requirements (customers), which usually refer to an end-of-line (EOL) requirement only.

An essential risk mitigation within the process chain can encompass the cleaning of assemblies or its components—ideally EOL, as Annex A.B.1 in VDA 19.2 shows [15]. The integration of an additional cleaning process was favored, after the implementation of other countermeasures showed limited success to mitigate the risk. In order to protect the lightguide from being damaged, the selection was therefore limited to non-contact cleaning methods. Finally, the first option was cleaning with compressed air (also referred to as “high-speed air jet removal” [23]). The second option was cleaning with CO₂, as described by Kaller et al. [24] and Robert Sherman [25].

Based on our personal experience, we knew that CO₂ usually achieves better cleaning results. However, we were unable to assess how much better the cleaning performance of CO₂ would be in comparison to compressed air. Moreover, we found only a few studies in which the performance of compressed air was compared with other cleaning methods [26] or with CO₂ directly, but on other substrates and other test particles [27].

Lastly, some empirical data were missing for the desired cost–benefit estimate of both cleaning methods. Therefore, this work was motivated by one central question: does compressed air, under optimal conditions, achieve the cleanliness target (contaminants above 100 µm prohibited) or is CO₂ needed instead?

2. Selection and Application of the Test Contaminants

Firstly, historical data were analyzed from the BLU production lines. Around 9% of all rejects were caused by contaminants. Knowing that, a minimum sample size of $n\ge 119$ was calculated for the survey [28,29].

The number of rejected BLUs reached 144 after a twelve-week production cycle (7200 produced units), out of which 74 rejects could be allocated to the lightguide. In-depth analyses were carried out by light microscopy and FT-IR to categorize the contaminants [30,31]. Particles of PE foam, fibers of textile clothing and cardboard in the size range from approx. 100 µm to 1000 µm were the most frequent causes of production rejects.

Next, test particles were produced from the materials identified in the previous step. The test particles of PE foam were scraped off the original substrate with a knife (see Figure 1b). The cotton fibers could not be obtained in sufficient quantities from the work clothing. Therefore, commercially available cotton fibers were utilized (supplier: KSL Staubtechnik GmbH, Westendstr. 11, 89415 Lauingen, Germany), as shown in Figure 1c. Instead of cardboard, commercially available artificial turf used in model-making was chosen (supplier: Noch GmbH & Co. KG, Lindauer Str. 49, 88239 Wangen im Allgäu, Germany), as shown in Figure 1a,d.

The commercial cotton fibers (c) were not examined in detail by spectroscopy. No information on the material composition was available for the two types of artificial turf (a) and (d), so FT-IR analyses were carried out before the start of the test. The analysis of the light green artificial turf (a) confirmed the expectation that the fibers are made of paper/cardboard. The dark green artificial forest soil (d) was a mix of fibers and flakes, with the former consisting of paper/cardboard and the latter of polypropylene.

The particles were initially applied in a dry state. However, this approach had to be discarded due to strong electrostatic charging, which made it impossible to apply the particles in a defined manner on the referenced field of the test substrate.

In a second trial, the particles were applied in a wet state—with better precision. To this end, a suspension of 2-propanol and the test particles was prepared, which was applied to the substrates by pipetting (0.5 mL per test area), as shown in Figure 2. A target of approximately 110 particles was favored for the application on the referenced field (area: 1764 mm²) in order to avoid errors of the automated particle counting devices. The maximum length of a single light green artificial turf particle (a) was assumed by 4 mm, so that 110 bounding boxes would fit into the referenced field. On the other hand, a lower amount would have had adverse effects for the calculations.

With regard to this target, the quantity of PE foam particles and cotton fibers applied was sufficient for the subsequent evaluation. However, the quantity of artificial turf particles applied was insufficient. Despite prior homogenization, these sedimented faster, so that only a few test particles could be applied by pipetting. The higher adhesion forces [32] of the test particles on the substrate as a result of the wet application were also advantageous during handling. The particles remained on the marked areas even after adjusting the orientation to 45° or 90°. Since the particles were applied wet in each test, this influence between the tests can be neglected in further consideration.

3. Selecting the Measurement Technology

Cleaning efficacy was evaluated by counting the test particles before and after cleaning. For this, optical measuring systems were used, which can automatically count the particles on the surface of the substrate [33,34]. A reflected light microscope and a grazing light measuring device were favored [35]. Original lightguides taken from series production were used as substrates. Lightguides were used in LCD displays, i.e., as digital dashboards for vehicles. The light emitted from the diodes (LED) at the edge/side of the lightguide is focused and directed to its destination by the micro lens arrays on the lightguide [36]. The lightguide has a smooth and a rough (diffuse) side, with micro lenses of approx. 100 µm in diameter (see Figure 3b). The cleaning validation was originally carried out on both sides of the lightguides. The smallest particle size measured (Feret_max [37]) was defined as 100 µm—based on the cleanliness requirement. However, with the chosen measurement systems, it was not possible to differentiate between micro lenses and particles in the size range of 100 µm (see Figure 3b and Figure 4b), so the cleaning validation was only carried out on the smooth side. According to the inline quality control during series production, the lightguides were all free of defects. Under the light microscope, however, microscopic scratches were visible on the smooth side (see Figure 3a). The light microscope identified these micro-scratches as particles, rendering an automated evaluation of the smooth side with the light microscope impossible.

The grazing light measuring device PartSens 4.0 (supplier: OMT GmbH, Schafwäsche 8, 71296 Heimsheim, Germany) proved to be more robust with regard to micro-scratches than the light microscope. The zero measurements (measurement of the smooth side, without test particles) only showed artifacts ≤ 50 µm (see Figure 4a). Thus, selectivity was sufficient to distinguish the smallest particles to be measured (see Figure 4a). This could also be verified after a simulated cleaning step. In this case, all test areas were wiped with a damp cleaning cloth and then subjected to a second zero measurement. The simulated cleaning process corresponded to Step (I)—as described in Section 6.

4. Experimental Setup

The cleaning tests were carried out in a cleanroom with air cleanliness class ISO-1 [38] so as to exclude disturbing environmental influences. The test setup consisted of a (1) movable carriage to which the (2) handle with nozzle for extra clean dry air (xCDA, cleanliness classes [0:1:0] according to ISO 8573 [39]) or carbon dioxide snow (CO₂) was mounted (see Figure 5). CO₂ could be added to the xCDA medium at the touch of a button. The nozzle geometry and the pressure (6.0 bars) were therefore identical in all tests. The (3) lightguide is fixed in a (4) bracket with locking holes for the fixed tilt positions 0°, 45° and 90°. The bracket is fixed to the carriage of the (5) linear axis. There are two (6) line ionizers above the active cleaning zone. The speed of the linear axis is set via the (7) control software.

The three different movement profiles of the linear axis from the tests are shown in Figure 6. Although the speed achieved with Profile 1 was still very close to the set target value, there was an increasing discrepancy with the other two profiles. The decisive factor, however, is that the final speed achieved was constant for all three profiles, allowing the xCDA/CO₂ jet to act equally on all four measurement surfaces.

The four square test areas on the lightguides had an edge length of 42 mm and were situated at approx. 100/200/300/400 mm (midpoint, respectively). Thus, a constant maximum speed was attained on all four test areas.

5. Design of Experiments (DoE)

Initially, eight factors were identified that were expected to influence cleaning efficacy [23]. This would have resulted in 120 experiments for the intended Box–Behnken design [40,41,42]. Therefore, the possibility of eliminating some factors was already investigated during the proof-of-concept phase. Although the factors’ pressure and ionization were classified as relevant, they were not considered as variables in the DoE. In production, these factors are set once and remain constant thereafter. For the experiments, the pressure was therefore set at a fixed 6.0 bars, and the ionization was permanently active. In addition to eliminating these factors, suitable value ranges had to be determined for the variables. The target value (cleaning efficacy) should react sensitively to the various factor combinations. At a distance of 5 cm or 10 cm, for example, it was no longer possible to differentiate between the cleaning efficacy of the two media. Regardless of the other factors, the cleaning efficacy was always 100 %. The levels of the attributes therefore had to be reduced until the desired sensitivity could be achieved.

Table 1. Factor selection for the DoE.

Factor	Type	Attributes
Orientation	Continuous	0°/45°/90°
Distance	Continuous	20 cm/30 cm/40 cm
Speed	Continuous	0.3 ms−1/0.6 ms−1/1.2 ms−1
Media	Category	xCDA/CO2

shows the final factors chosen for the Box–Behnken design (without repetition) with a total of 30 experiments. The complete experimental design (including results) is shown in Figure 8.

6. Procedure

Four marks were placed on each lightguide to reference the PartSens 4.0 measuring head for each measurement (see Figure 7). The zero measurements, as described in Section 3, were carried out before the tests.

(I)

Damp wiping

Since only 10 lightguides were available for the planned 30 experiments, they had to be cleaned before each new loop. For this purpose, the entire surface of the lightguides was wiped with a damp cleanroom cloth (product: Contec Prosat PS 911 BRP).

(II)

Particle application

The respective test contamination was applied to the test areas, with there being four in total for each lightguide.

(III)

Particle analysis

After the contamination liquid evaporated, the measuring head of the PartSens 4.0 was placed on the four test areas one after the other to determine the quantity of test contamination applied (before cleaning).

(IV)

Cleaning

After the lightguide was clamped in the test setup and the test parameters were set, the cleaning step was carried out. In doing so, the linear axis moved from its initial to its end position (see Figure 7).

(V)

Particle analysis

After cleaning, the lightguide was removed from the test setup and—analogous to Step (III)—the quantity of residual test contamination on each test area was determined. The difference between the two quantities indicates the cleaning efficacy.

7. Discussion

The experimental design and the results of the cleaning tests are shown in Figure 8.

Figure 8. Experimental design and results.

Figure 8. Experimental design and results.

The interactions of the main and secondary effects for PE foam and the cotton fibers were investigated. The results of the two artificial turf particles were not evaluated. The particles sedimented faster, which meant that a lower quantity of particles was applied. A higher statistical uncertainty was expected here, having a negative effect on the regression. Although the results with the cotton fibers are based on a better regression model (R² = 90.27 %), the results with PE foam (R² = 76.27 %) will be discussed in more detail below, as these test contaminants are original (i.e., they are identical to the naturally occurring contaminants). In both cases, the regression models were optimized—before evaluation—with regard to the residues in order to achieve a better fit. Quadratic interactions with a p-value > 0.250 were eliminated to increase R². Figure 9 shows which effect has a significant impact on the regression model with the PE foam particles. The cleaning efficacy is influenced in turn by the following effects: A: Orientation, B: Distance, AA: Orientation², BD: Distance×Media. It is noteworthy that the main effect, Distance*Media, has no significant influence according to the regression model. Looking at the primary data (see Figure 8), only CO₂ achieves 100 % cleaning efficacy. The cotton fibers, on the other hand, appear to adhere more strongly to the substrate, which is evident from the lower cleaning efficacy. It seems that the smaller and/or flatter the particles become, the more difficult they are to remove [43,44,45].

The cleaning efficacy decreases linearly with increasing distance from nozzle to substrate, or with increasing speed (see Figure 10). Regarding the orientation factor, there is an optimal range between 45° and 65° for air and between 65° and 90° for CO₂. Contrary to expectations, the statistical model shows that air has a better cleaning effect than CO₂. However, when considering each test individually, this can only be confirmed for the tests (pairwise comparison) 14:27, 17:22 and 40:19. This correlation is also evident in the case of cotton fibers.

However, when looking at the interaction diagrams in Figure 11, the previous statement is put into perspective. At distances > 25 cm (distance between nozzle and substrate), CO₂ has a higher cleaning efficacy. With cotton fibers, CO₂ already achieves better cleaning results at distances > 10 cm. In other words, there is no advantage in using CO₂ if the nozzle can be placed close to the substrate.

If a cleaning efficiency ≥ 99.9997% is required (corresponding to 6σ or C_p = 2.0), this can only be achieved—based on the underlying test conditions—with CO₂ and the following parameters (see Figure 12):

• Orientation: 77.27°
• Distance: 20 cm
• Speed: 0.3 ms⁻¹

This optimization could only be carried out within the limits of the model (parameter sets). However, it can be assumed that an even better cleaning efficacy can be achieved if the distance/speed is further reduced. By comparison, only a cleaning efficacy of 94.19% is achieved with air under optimal conditions (corresponding approximately to 3σ or C_p = 1.00), as shown in Figure 13.

8. Conclusions

The objective of this work was to investigate the cleaning efficacy of compressed air and CO₂ snow jet cleaning in order to assess which of the methods was able to achieve the cleanliness requirement (no particles larger than 100 µm permitted) of the backlight unit (BLU). This work was divided into the following four stages:

• Select representative test contaminants.
• Select a suitable particle measurement instrument.
• Design and build the test rig.
• Experiments.

The underlying Box–Behnken design considered the following impact factors: orientation, distance and speed as representative variables of the BLU assembly. Each experiment was carried out in the following sequence:

• Damp wiping of the substrate (lightguide).
• Particle application.
• Particle count (before cleaning).
• Cleaning with compressed air or CO₂.
• Particle count (after cleaning).

Finally, the cleaning efficacy was calculated as a differential particle count (after and before) and via a percentage. The results show that the cleaning efficacy varies, depending on the level or sort of contamination. With respect to the experiments, for instance, fibers (represented by cotton fibers) are harder to remove than particles (represented by PE foam). This means that, a cleaning performance survey—also known as cleaning validation—requires a representative test contamination.

The impact of the distance (nozzle to surface) on the cleaning efficacy is noticeably higher when cleaning with compressed air. On the contrary, the impact of this factor is rather low for CO₂, although the shell of this jet is formed by compressed air. A similar influence of the distance was expected for compressed air and CO₂, which was not confirmed.

In addition, the impact of the orientation on the cleaning efficacy shall be emphasized as well. A retrofit of compressed air or CO₂ cleaning systems into existing assembly stations is a common practice in the automotive industry. This can become rather challenging due to space limitations inside the station; thus, the optimal orientation cannot be obtained. Cleaning systems are often installed at an angle of 90° to the surface (observation from dozens of shopfloor visits), which potentially reduces the cleaning performance—as the experimental results show.

When installing CO₂ cleaning systems, the highest priority should be given to the orientation. There is an optimal point for the best cleaning performance, with respect to our experimental conditions: 77.27°. On the opposite, when installing compressed air cleaning systems, the following principle applies: the closer they are to the object, the better the cleaning result.

To be able to answer the central question of this work, it is important to distinguish between two types of cleanliness requirements: (a) a maximum permissible particle length and (b) a particle size distribution (maximum permissible particle count per size class).

The underlying cleanliness requirement of this work (no particles larger than 100 µm allowed) corresponds to type (a). In consequence, the cleaning process must be able to remove 100% of all particles larger than 100 µm from the surface. This requirement was only achieved in the experiments with CO₂, when the distance was ≤30 cm and the orientation was 45° (see Figure 8). With compressed air, however, the cleanliness requirement was not achieved in any of the experiments, nor would it be possible after optimization!

Based on the results, someone would expect that a CO₂ cleaning system would have been integrated. In the end, a compressed air cleaning system was integrated, which was justified with some findings of a secondary process chain analysis. The handling and transfer of the lightguide to the next assembly station generated additional particles so that the need for a 100% cleaning solution changed, with a shift towards subsequent processes.

If the cleanliness requirement would have been of type (b), on the other hand, the success of a cleaning method would have depended on how many particles per size class remained on the surface after cleaning. This does not necessarily require a 100% cleaning performance, which could have favored the use of compressed air cleaning.

Let us assume that for the following example, the cleanliness requirement—according to type (b)—would have been as follows: max. 30 particles in size class 50–99 µm and max. 15 particles in size class 100–149 µm allowed. Moreover, let us assume that 110 particles were applied in each size class before cleaning. This would mean that we would need to remove 80/95 particles in size class 50–99 µm/100–149 µm, which would correspond to a cleaning performance of approx. 72.73%/86.36% or better. In the optimal configuration (see Figure 13), compressed air would achieve a cleaning performance of 94.19% and thus fulfill the requirements for both size classes. The configuration from Experiments 23, 26 and 28 would also meet the requirements for the aforementioned size classes.

This example was aimed to demonstrate that an inductive conclusion based on the experimental results obtained in this work is not possible. In other words, due to the dependency of the cleanliness requirement itself, a conclusion from the use case presented here is not generally transferable. For this reason, it was vital for the authors of this study to describe the method and underlying thoughts in detail so that this procedure could be utilized by third parties for their use cases.