Liquid crystal display surface uniformity defect inspection using analysis of variance and exponentially weighted moving average techniques

Display quality is part of the final liquid crystal display (LCD) inspection process before shipping. A 'limited sample' is provided based on the agreement between the manufacturer and the customer. This inspection usually includes an operator who compares the LCD product with the limit sample using the naked eye. The procedure often causes controversy in the manufacturing plant and between the manufacturer and customers. This study attempts to establish a more objective, automatic method to determine the MURA-type defects in LCD panels. A luminance meter is used as the measurement device. An LCD panel is divided into 144 areas. Five points are measured to obtain the luminances. Analysis of variance and the exponentially weighted moving average techniques are applied to determine the existence of MURA defects. Fifty normal LCD panels and 50 MURA defects panels were used to test the inspection method. All 50 LCD panels with MURA defects were correctly identified using the proposed inspection method. The proposed inspection method can help LCD manufacturers reduce the variation in LCD panel inspection results and establish a better relationship with customers through a common inspection mechanism.

Keywords: Liquid crystal display (LCD); Defects inspection; Analysis of variance (ANOVA); Exponentially weighted moving average (EWMA)

The thin-film-transistor-liquid crystal display (TFT-LCD) monitor has recently become the main stream of display devices (Yamazaki et al.[15]) because of its smaller space requirement, high-quality display and acceptable cost. Customers demand high-quality product functions and appearance. The LCD panel surface defect inspections are usually performed manually. This process is a labour-intensive, non-value-added part of the production cycle. As higher product quality is required, manufacturers must improve both manufacturing and inspection capabilities.

Most TFT-LCD manufacturers use images generated from simulated signals with operators inspecting the panel with the naked eye. It is easier to inspect defects induced by electronic signals, such as defect lines or points. Non-uniformity defects, normally called MURA defects, are the most difficult to recognize. A 'limit sample' is usually used as the basis for quality judgement. However, controversy often occurs due to the vagueness of the image.

Generally, automatic inspection systems for LCD panel have an optical- or an electrical-based approach. Several electrical or optical-based inspection techniques have been developed for LCD manufacturing. Nevertheless, most existing methods of automatic inspection systems for LCD are based on conventional electrical methods to detect the surface potential. That electrical-based techniques work well for functional verification of an LCD panel. However, they can only be accomplished after the fabrication is completed. In-process inspection may not be applicable to the functional test approach. There is not much research in the literature related to TFT-LCD defect inspection using a vision-based approach. Lu et al. ([7]) applied a singular value decomposition (SVD) algorithm and developed an image reconstruction scheme for inspecting LCD panel surface defects, such as pinholes, scratches, particles and fingerprints. The SVD algorithm could eliminate the normal texture on the inspected surface, thus concentrating on the defect areas. Lin et al. ([6]) established a fast image segmentation technique for inspecting the defects on an LCD spacer. Fuzzy set theory was used to determine the threshold for image segmentation. A set of specially designed operation masks was then applied to abstract the features on a spacer. Hwang and Lin ([2]) suggested using Houng's transformation technique to calibrate the positional accuracy of an LCD. The defects on the colour filter were thus identified (Wu et al.[14]). Kim et al. ([5]) developed a parallel-operation architecture on a multi-processor vision system for large sized TFT-LCD inspection. A line scan camera was used to obtain a higher resolution image. Four digital signal processors were used to reduce image-processing time. Saitoh ([12]) developed a machine vision system for inspecting LCD brightness unevenness. Saitoh first used an edge detection algorithm to detect discontinuous points that included brightness unevenness and noise pixels. A generic algorithm was then applied for detecting the brightness unevenness defects. Previous research did not deal with the LCD panel display non-uniformity problem. The objective of the present study is therefore to establish an automatic inspection process for LCD panel surface uniformity defects.

A major difficulty in TFT-LCD panel non-uniformity inspection is that a good-quality image is difficult to obtain using a regular charge-coupled device (CCD) camera because the camera's grey level resolution is limited. Therefore, a CCD camera used for inspecting a printed circuit board (PCB) (Jiang et al.[4], [3], Wang and Jiang [13]) would be difficult to use for LCD panel non-uniformity inspection. In this research, a luminance meter, a light-sensitive device, was used to measure panel surface brightness. A data-sampling plan based on the smallest defect area to be detected was designed to collect data effectively for analysis. The analysis of variance (ANOVA) (Montgomery and Runger [10]) and exponentially weighted moving average (EWMA) control chart (Montgomery [9]) concepts were used to detect the non-uniformity areas in a panel. ANOVA was used because the non-uniformity problem between panels is not an issue. The major concern is the non-uniformity (e.g. MURA defects) existing in a given panel; therefore, this technique could be used to identify areas significantly different from other areas in a panel. The EWMA was applied to detect the small differences in various panel areas.

The TFT-LCD manufacturing process produces numerous product types and defects. The product requirements and defect acceptance levels vary according to customer requirements. This research was conducted under the following conditions:

• 1. Product was limited to 15-inch TFT-LCD modules.
• 2. Manufacturer normally separates these products into higher and lower classes to fulfil various customer needs. Research focused only on the higher class products.
• 3. Manufacturer negotiates with the customer using a 'limit sample', which is used to determine the lower product acceptance level limit. In this research, the limit samples used were samples commonly accepted by general customers.
• 4. According to a common specification, a MURA defect less than 5 cm2 can be ignored. This served as the minimum recognized defect size.

There are three types of TFT-LCD display defects: (1) visual appearance defects, (2) electricity induced defects and (3) non-uniform display defects. The first type is caused primarily by the existence of strange articles in the panel, non-uniform colour filter distribution, an uneven space between the upper and lower glass panels, and poor polarizer or back light panel quality. The visual appearance defects mainly include strange particle, spacer clog, white spot, black spot, scratch, operational rubbing fail, rubbing string, strange article, fibre-type article bubble and white spot. The electricity-induced defects are usually caused by poor electrical transmission. These defects appear more regular in shape such as lines or dot defects such as S-line, G-line, SS shot, GS shot, S open, G shift, G open, GS open, luminance defect, black spot defect, continuous spots defect, spot distance, IC poor activation, IC force deviation, poor grey level adjustment, adjustment line defect, feedback circuit defect and power source problem.

Non-uniform display defects appear as an area in which the colour (or grey) is different from the other areas in the panel (figure 1). These defects could appear with a different colour background. These defects might occur due to uneven exposure or other steps in the manufacturing process.

Graph: Figure 1. Common MURA defects in an TFT-LCD.

Among the three defect types, the visual appearance defects can be detected using a vision system (Lu et al.[7]), the electrically induced defects can be adjusted using conventional electrical methods such an electro-optic modulator (Ryan [11]). There is no systematic method for determining the third defect type, the non-uniformity display defect. This research therefore attempts to establish a method for detecting and inspecting non-uniformity display defects.

A luminance meter (BM-5A) was used to measure the luminance level of a defined area on an TFT-LCD panel. A 1000-pixel area was used for each measurement. Data collection was conducted in a dark room with the panel switched to a white background. The equipment set-up is shown in figure 2. A schematic diagram of the experimental set-up is shown in figure 3.

Graph: Figure 2. Equipment set-up for data collection.

Graph: Figure 3. Schematic diagram for data collection.

The relationship between the measurement distance (WD), measurement diameter (D) and the meter aperture (θ) is:

Graph

Using a 1000-pixel measurement area, the measurement diameter is D =

Graph

35.6 pixels. In general, the spacing between two pixels for a 15-inch LCD panel is 0.28 mm. Therefore, D ≈ 10 mm. The BM-5A aperture was selected as 1°. WD can be calculated as follows:

Graph

As noted, a MURA defect less than 5 cm2 is usually ignored in the field. A 15-inch TFT-LCD panel was then divided into 16 × 9 blocks at 4.28 cm2 each. Five sub-areas with 1000 pixels each were measured in each block. The five sub-areas were the four corners plus centre locations. The data collection design is shown in figure 4. A 15-inch TFT-LCD has 1024 × 768 = 786 432 pixels. This is equivalent to 5461 pixels in each of the 144 blocks. The luminance meter measures luminance by collecting an area of 200 pixels of data. Five sub-areas with 5000 pixels of data were collected. This meant that 92% of the pixels were measured in each block. This data collection plan was thought sufficient to detect MURA defects on a TFT-LCD panel.

Graph: Figure 4. The 144 blocks and five sub-areas on each panel designed for data collection.

The ANOVA and EWMA techniques were used for data collection and analysis for determining the non-uniform areas and their locations. ANOVA was used to detect if a MURA defect exists on a panel. ANOVA was used because in the LCD non-uniformity problem (such as MURA), there is more concern with non-uniformity on a given LCD panel and there is less concern with the display variation among different panels. Therefore, this technique could be used to identify areas significantly different from other areas in a panel. The EWMA was applied to detect the small differences in various panel areas and to detect position and size of MURA defects.

The major purpose for applying ANOVA was to evaluate the difference between the means from several populations (Montgomery and Ruger [10]). One-way ANOVA was used to establish a procedure for evaluating the non-uniformity in a given LCD panel.

Let xij be the ith block and the jth measurement of a brightness, i = 1, 2, ... , 144; j = 1, 2, 3, 4, 5. The statistical ANOVA model can be expressed as:

Graph

where μi is the ith block mean and εij is the random error term. It is also assumed that the error terms form a normal distribution with mean of 0 and standard deviation of σ (Montgomery and Runger [10]). The following hypothesis test was established according to the defined ANOVA model:

Graph

If H0 is rejected, the non-uniformity problem exists on the tested LCD panel. As part of the ANOVA theory, the total sum of squares (SSTO) for the measured values in a block are dissembled into the sum of squares for the treatments (SSTR), the variance within different blocks, and the sum of squares for the random errors (SSE). The mean square treatment (MSTR) and the mean square error (MSE) are obtained by dividing the SSTR and SSE by their degree of freedoms, respectively. If the hypothesis H0 holds true, both MSTR and MSE represent the unbiased estimates of σ2. While H0 is not true, the MSTR will overestimate σ2, and the MSE is still an unbiased estimate of σ2. According to this, the MSTR/MSE ratio can be used to judge the degree of difference among the mean measurement block values. It can be statistically proven that the MSTR/MSE is an F distribution and, therefore, under a decision risk α, normally set as 0.05, the MSRE/MSE can be compared with the standard F-value. When F(143,576,α)<(MSTR/MSE), i.e. accept H0, indicating the panel is non-uniform under 100(1 − α)% confidence.

The set α level will affect the ANOVA confidence conclusion. Three assumptions are held when conducting ANOVA applications for the validity of a conclusion. The three assumptions are normality, equal variance and independence. The residual analyses with graphical and statistical tests are normally used to verify these three assumptions. Box et al. ([1]) showed that the F-test is more robust to the normality assumption if the statistical model is fixed. That is, the ANOVA conclusion is less sensitive to data that violates the normality assumption. The independence nature is critical to the ANOVA application. It is important to collect random data to avoid the data dependency problem.

The procedure for applying ANOVA is summarized as follows:

Brightness data were collected from an LCD panel without the MURA defect as previously described, with five points from each of the 144 blocks. Residual analysis was conducted to check the normality, equal variance and independence assumptions. The residual data plots are shown in figure 5. The normality assumption is accepted from the 'normal residual plot' and the 'residual histogram'. No particular trend is observed from the 'Residual I chart' and 'Residuals vs. fits'. The equal variance and independence assumptions were also acceptable. A set of statistical tests was conducted to serve as a reference for accepting the three assumptions. The Kolmogorov–Smirnov test, Bartlett and Durbin–Waston tests were used to test normality, equal variance and independence, respectively (Ryan [11]). The results showed that the normality assumption was marginally acceptable and the other two assumptions were confirmed. Because the ANOVA is more robust to the normality assumption, the analysis was continued.

Graph: Figure 5. Residual analysis plots for an LCD panel without MURA defects.

The ANOVA table is shown in table 1. The p-value is 1.000. This is much higher than the set α of 0.05, indicating that the panel has no MURA defects.

Table 1.  One-way ANOVA for an LCD panel with MURA defects.

The box plots for the tested panel are shown in figure 6. Note that under the 95% confidence level, the 144 data blocks have no significant differences.

Graph: Figure 6. Box plots for the LCD panel with MURA defects.

An LCD panel with MURA defects was selected for the same procedure. The residual plots are shown in figure 7. From the 'Residuals vs. fits' plot, the data are clustered into two bands of areas. A statistical test might be helpful for determining the validity of the equal variance and independence assumptions. Bartlett's test was used to test the equal variance property (Ryan [11]). The p-value was 0.674 and thus the equal variance was accepted. The Durbin–Waston test was used to test the independence assumption. The Durbin–Waston value was 1.92, which is very close to 2, an indication of being independent. Therefore, the ANOVA table was established as shown in table 2. The p-value is close to 0, and it is concluded that the panel has significant MURA defects.

Graph: Figure 7. Residual plots for an LCD panel with GAP MURA defects.

Table 2. ANOVA table for an LCD panel with MURA defects.

The box plots for the tested data are shown in figure 8. Under the 95%, confidence level, at least one block has luminances significantly different from the other blocks.

Graph: Figure 8. Box plots for an LCD with GAP MURA.

Control charts are normally used for detecting assignable causes of variation (Montgomery [9]). In this study, the control charts concept was used for detecting the position and size of MURA defects, shown as non-uniform luminance on an LCD panel. The measurement data from the 144 blocks were plotted onto the control chart. Because the luminance variation might be slight, the EWMA control chart was selected for this study.

The tth exponentially weighted moving average zt is defined as:

Graph

where λ is a smoothing constant, or called weighting constant, 0<λ ≤ 1,

Graph

is the average of the tth group data, zt is the weighted average for the past data, the initial value

Graph

, or an estimate of the mean can be given. The upper and lower limits of the EWMA control chart are as follows (Lucas and Saccucci [8]):

When t<5, the upper and lower control limits are:

Graph

When t ≥ 5, (1 − λ)2t approaches zero, and the upper and lower control limits are

Graph

where

Graph

can be estimated as

Graph

.

The EWMA detection ability is affected by the parameters k and λ. As a common practice, k is set as 3. λ is determined using the trial-and-error method. One normal and one defective LCD (with MURA defects) panels were used to determine λ. The MURA position of the defective panel is shown in figure 9. The detection rate for various λ values is shown in figure 10. The detection rate is highest when λ is 0.8. Therefore, λ was set at 0.8 in this study. The resulting EWMA control charts are shown in figure 12.

Graph: Figure 9. Position of the GAP MURA of an LCD panel.

Graph: Figure 10. Detection rate for various λ values.

Graph: Figure 11. EWMA control charts for normal and MURA LCD panels (λ = 0.8).

Graph: Figure 12. MURA area position.

When λ = 0.8, the blocks were detected as abnormal through the EWMA control chart: 28·29·41·42·43·44·45·46·56·57·58·59·60·61·62·63·64·74·75·76·77·78·91·92·93·94. The MURA area is shown in figure 12. The area is in agreement with the area shown in figure 9.

To verify the effectiveness of the developed procedure, 100 LCD panel sample data were collected from a local LCD manufacturer. Among the 100 samples, 50 were normal and the other 50 were defective with six types of MURA defects. The 50 MURA defect samples include 15 GAP MURA(I), 12 Surrounding GAP(II), 9 Half-moon GAP(III), six Tr GAP(IV), four Non-uniform exposure(V) and four Non-uniform side angle exposure(VI).

The size distribution of the MURA areas in the 50 defective panels is shown in figure 13.

Graph: Figure 13. MURA defect area distribution.

The luminance data were collected using BM-5A equipment. The previously described procedure was followed. The ANOVA was conducted and the p values were obtained. The EWMA control charts were then used to identify the defective areas. These results were confirmed by an experienced LCD inspection team. The results are shown in table 3. The results showed that all 50 defective LCD panels with six types of MURA defects were identified using the proposed procedure. The defect type, size and location information can then be fed back to manufacturing engineers for process improvement.

Table 3. ANOVA and EWMA inspection results.

An automatic inspection procedure was developed in this study for detecting 15-inch TFT-LCD panel MURA defects. Luminance measurement equipment (BM-5A) was used to collect the needed data for analysis. The first analysis phase involves conducting an ANOVA to detect the existence of MURA defects. The second phase involves plotting EWMA control charts for determining the location and size of the defects. A flowchart of the detection procedure is shown in figure 14. One hundred samples were used to test the effectiveness of the proposed procedure. All of these charts were compared with the results from an LCD panel inspection team currently conducting inspections without the proposed automatic procedure. The advantages of the proposed procedure are: (1) the proposed method uses equipment currently used for other purposes in the plant. Additional equipment is therefore not needed. (2) Defective LCD panel determination is more objective compared with naked eye inspection. (3) The defect type, location and size can be readily obtained. This information can be provided to manufacturing for process improvement. The proposed procedure was limited to 15-inch LCD panels as produced by a local manufacturer. The disadvantage of the proposed procedure is that the λ value is determined through trial and error, and a defective area crossing two of the 144 testing blocks is difficult to identify. The limitations and disadvantages of the results from this study point toward future research areas.

Graph: Figure 14. Flowchart of the proposed LCD MURA defect detection procedure.

Research was supported by the National Science Council of Taiwan, Project Numbers NSC 91-2213-E-155-034 and NSC 92-2213-E-131-011.