A data-driven manufacturing support system for rubber extrusion lines

A better control of extrusion processes offers clear advantages in the manufacturing of rubber profiles for the automotive industry. This work reports our experience in developing a support system aimed to ease the work of the extruder machinist while improving the quality of the profiles obtained. In order to build the system, an approach based on facts was adopted, following ISO 9000 standard quality principles. The data warehouse service available provided a wealth of information on the conditions of the running processes. The collected data, after being analysed with the appropriate data-mining techniques, allowed us to gain a better understanding of the process and to identify the main causes of variance. In particular, principal components analysis, Sammon projection and several classification techniques were applied for exploratory purposes. Different behaviours could be described for the extrusion process, allowing for the definition of a control strategy and, eventually, the development of a manufacturing support system. The estimates displayed by the system greatly improve the responsiveness of the machinist when the process departs from expected behaviour. The results of using this system in a local factory proved highly satisfactory and encouraging.

Keywords: multivariate statistics; neural network applications; neural networks; new technology management; software engineering

1. Introduction

Rubber profile manufacturing is an important economic activity linked to the automotive industry. These profiles are commonly found in windshields, windows, and rear view mirrors, where clients impose high specifications on the finishing and the geometrical dimensions of the sections. The manufacturing process is generally based on the extrusion of rubber. This is an intrinsically complex non-deterministic process (Raddatz et al. [28]) affected by multiple variables, e.g. the temperature in the different extruder areas, pressure in the transfer blocks and the rotational speed of the spindles, to name a few. Not surprisingly, it also has the added difficulty of increasingly higher quality demands.

The process is typically operated following an open loop actuation scheme. Current sensing techniques only allow measurement of the profile features by taking samples for analysis in the laboratory. The resulting products at the output of the extruder are quite unstable, and thus the measurement cannot be performed until the end of the line. Considering that the length of the line may be more than 60 m and that the production rate can be close to 15 m min−1, it will take no less than 4 min for a sample to be obtained, and additional time for the sample to be processed in the laboratory. Success rests in the hands of experienced operators who must drive the process, simultaneously paying attention to a number of variables and not knowing the results until after 4 min have elapsed.

There are additional disadvantages of the process: incorrect profiles cannot be recycled and the raw material is quite expensive, compelling reasons to take advantage of the excellent opportunities offered by optimisation of the extrusion process. Factory owners would certainly aspire to something better than 60 m of wasted rubber for every error. This would only appear to be possible by improving the responsiveness of the operators. A manufacturing support system may be useful in such a situation as long as it provides estimates of the status of the process. A support system would ease the work of the operator, providing real-time information to determine whether the process has strayed outside the production constrains, thus improving the quality of the profiles.

A brief search of the literature shows that this is an active field of research where different approaches can be followed to improve our understanding of the extrusion process, either by means of classical models (Mackerle [25], Sutanto et al. [37]), computational fluid dynamics (del Coz Díaz et al. [10], Dai et al. [8], Müllner et al. [26], Zheng et al. [42]) or case-based reasoning (Raddatz et al. [28]). Among the different techniques, a data-mining approach offers ample opportunities to gain unique insights into the rubber extrusion process.

Despite the successful application of data-mining techniques to manufacturing processes (Lian et al. [22], Chen [5], Kuriakose et al. [20], Ordieres Meré et al. [27], Agard and Kusiak [1], Shi et al. [34], Chaudhry and Luo [4], Cunha et al. [7], Dengiz et al. [11], Feng and Kusiak [12], Hsu et al. [16], Jin and Ishino [17], Kusiak [17], Liao et al. [23], Raheja et al. [29], Shahbaz et al. [32], Shao et al. [33], Tseng et al. [38], Wang et al. [40], Braha et al. [3], González Marcos et al. [14], Wong et al. [41], Al-Ahmari [2], Chen and Wang [6], Kao and Li [19], Schreck et al. [31], Shiue [35], Dean et al. [9], Ho et al. [15], Liu et al. [24], Song and Kusiak [36]), many companies still ignore the true potential of data-mining tools as an aid to improve their processes. Nevertheless, many companies are already in a position to initiate analyses, as information systems and databases that store the records of significant production variables are becoming more and more common.

This paper reports a successful experience in applying data-mining techniques to the rubber extrusion process in order to provide a support system that will aid extruder operators to reduce the waste of time, material and money involved in defective production. The project was aimed at keeping all the critical variables involved in the process, i.e. pressures, temperatures and speeds, at constant values so as to ensure a uniform extrusion and, thus, a good-quality product with constant section. A support tool was developed to help control the steady state of the extrusion process by visualising onto a projected map the areas where the process stabilises and the track of its evolution. Such a tool will help the extruder operator approach stable production regions.

2. Data analysis

2.1 Data preprocessing

The factory where we developed this project (Metzeler Automotive Profile Systems Ibérica, S.A.) already had monitoring systems in place for the production processes supported by databases. These databases recorded the values of the variables from different processes, including the extrusion process. In order to provide a general model, the data sets corresponding to 12 different rubber profiles and five extrusion lines were considered. This caution prevented the obtained model from being overfitted to a specific profile or line. In total, more than 1,250,000 patterns were considered. Samples from this data set were taken every 100 ms, averaged per minute and stored in the database during 2005 and 2006. The data from 2005 were obtained from only one line from February to November. The data from 2006 came from five different lines from January to May. Fifty-two variables were considered, 25 corresponding to commanding action values and 31 were related to various measurements. These variables were both categorical, e.g. 'plate shape', and quantitative, e.g. 'target temperature at the head of the first extruder'.

In order to reduce the amount of redundant information and, in general, useless data, those records corresponding to starting and stopping transients, negligible running periods or outliers, e.g. records registered while some sensors were out of order, were removed. Removing the transients allows the analysis to focus on the steady production state (see Figure 1). After removing the outliers, the variables were rescaled (Sarle [30]) in the range [0, 1]. This prevented numerical difficulties when running the different analyses.

Graph: Figure 1. Line graph for a single variable, such as the temperature of die machine number 1. (a) Start of production and outliers. Data not preprocessed. (b) Preprocessed data and stable production data set related to this temperature.

2.2 Exploratory data analysis

Exploratory data analysis (EDA) techniques, such as principal component analysis (PCA) (Jolliffe [18]) and Sammon's projection (Venables and Ripley [39]), were applied in order to visualize the structure of the data sets and to provide a first suggestion for the number of clusters. Figure 2 shows the variance corresponding to the principal components obtained by applying PCA analysis to one of the lines. As can be seen, the first three principal components have variances significantly larger than the rest.

Graph: Figure 2. Variance of the principal components of one of the analysed lines.

From the weights of the change in the basis matrix associated with the PCA results, we can highlight those variables whose contribution to the first two principal components is greater than 25% (see Table 1). These results show that the temperature in the different areas of the extruder is the variable with the greatest influence on the variance. Thus, a greater control over that variable is mandatory.

Table 1. Influence of the dominant variables in the first two principal components.

Variable	First PCA component (%)	Second PCA component (%)
E1_C1_T_R	31	3
E1_C2_T_R	32	17
E1_C3_T_R	31	25
E1_CAB_T_R	33	18
E2_C1_T_R	36	51
EP_C1_T_R	29	35
EP_C2_T_R	30	25
EP_C3_T_R	31	19
EP_CAB_T_R	20	52
EP_HUS_T_R	29	11

PCA analysis provides additional information. Besides indicating those variables with a greater influence on the variance observed in the data set, a map can be obtained, considering the projection of the data set according to the change in the basis matrix determined by PCA. In this case, the projection of the data over the PCA map (see Figure 3) suggests the presence of six different clusters or stable states.

Graph: Figure 3. Parameters of two principal components of the analysed line.

In order to gain an alternative perspective, we applied the Sammon's projection technique to the data set. The results are shown in Figure 4. Figures 3 and 4 are similar. The same clusters are represented in both maps, although rotated. Figures 3 and 4 also show a few samples that could be considered outliers according to their distant positions. These outliers are due to temperature variations and their influence on viscosity. This effect is a major cause of deviations from the client's quality requirements. Thus, it is mandatory to identify them before they are sent to the client.

Graph: Figure 4. Sammon projection of the data set considered in the text.

Figure 5 shows the track of five production sequences corresponding to the line and plate considered in the above analyses. It shows that the evolution of the variables as the process progresses is not limited to only one cluster, but different clusters can be crossed. This is a reflection of the instability of the extrusion process, and an encouraging reason to establish an adequate control strategy that allows the operator to identify these fluctuations in order to take action and drive the process to the desired working point.

Graph: Figure 5. PCA of different production sequences of a profile type.

2.3 Unsupervised and supervised classification

Both PCA and Sammon's projection suggested the presence of different clusters. In order to characterise these clusters, different clustering techniques were applied: the AGNES and CLARA (Venables and Ripley [39]) and MCLUST (Fraley and Raftery [13]) algorithms. The results obtained by applying these techniques confirmed what was first suggested by PCA and Sammon. Figure 6 shows the results obtained by applying these three techniques to the data set considered. The clustering techniques identify four groups instead of the four of five suggested by PCA and Sammon's projection, as some of them are sufficiently close to be statistically considered as only one group. The congruency amongst the results of the different clustering techniques is remarkable.

Graph: Figure 6. Groups observed in a line and a profile type using three different models: AGNES, CLARA and MCLUST.

As a complementary visualisation technique, linear discriminant analysis (LDA), a particular supervised classification technique, provides a projection that gives the best separation of the clusters (Fraley and Raftery [13]). Figure 7 shows the LDA projection of the data set considered. Again, only four clusters are suggested.

Graph: Figure 7. Grouping of data of a line and a profile type applying the LDA technique.

2.4 Control strategy

We now use the PCA map to draw the projection of the commanding actions (PR) and the real measurements (R) corresponding to the data set considered in the above analyses (see Figure 8). The results obtained show little differences between the projected values of PR and R. This reflects the fact that the control system drives the process according to the commanding actions without significant inertia. The real values show greater variance, which emphasises the importance of the control strategy. In order to obtain the best quality, it is mandatory to control the process parameters in a more precise manner.

Graph: Figure 8. PR values projected on the PCA plane of the R values, in two types of line profile.

In general, of the set of variables studied, the temperature, both in the auxiliary and tri-component extruder and in the body of the main extruder, is the variable that appears, in most cases, to be mostly responsible for the variance in the two first principal components of the data set considered. It is remarkable the influence that the temperature has on the viscosity of the polymer. It also has an effect on the characteristics of the surface of the extruded material and its degree of crystallisation. One of the mechanical properties most effected is the impact resistance of the products, a very important feature in the automotive industry as these profiles will undergo constant use.

We have presented a probability density estimation (p.d.e.) of the data set on the plane defined by the first two principal components so as to provide the extruder operator with a bi-dimensional picture of the status of the process and its location according to the different working points. Figure 9 shows the p.d.e. picture corresponding to the line and plate shown in Figure 3. The p.d.e. confirms our previous observation that, of the six clusters suggested by PCA, one of them lacks the necessary density to be considered a proper cluster and two appear to be sufficiently close to be embraced by common equiprobability curves, thus providing a visual interpretation of the cluster analysis results. Once the position of the process has been determined, the track of the process can be updated on the map, thus providing the operator with the essential information to determine whether or not the process is outside the desired working area. When a departure is identified, the pattern is analysed to determine which variables are out of range, and the operator can then focus their attention on those variables.

Graph: Figure 9. Equiprobability curves for a profile type of the analysed lines.

Once the process has been characterised, a control strategy can be established. The purpose of this strategy is to improve the conditions under which extrusion is performed. This is accomplished by a continuous advanced supervision system (Figure 10) with a set of warning signals that alert the operator when a departure trend has been identified in the projected space. The advantages of simplicity and reduction in time needed to detect an anomaly has fostered the implementation of this strategy in production lines.

Graph: Figure 10. Projector of the region of space of the PCA used in the production lines.

3. Conclusions

The insights gained from analysis of the rubber extrusion process using a data-based approach have lead to the development of a useful tool that allows the extruder operator to drive the process in a more confident manner, providing tangible benefits to the factory by reducing the waste of raw material and endowing better quality on the manufactured profiles. The application of these techniques to the extrusion process has a direct impact as an inestimable aid to the extruder operator. As the operator does not have a particular fixed location with respect to the machine, he(she) is alerted as soon as a departure is identified, with complementary information concerning the variable mainly responsible for the departure. With such a support system, the operator can initiate corrective actions sooner, preventing the waste of many meters of rubber profiles. From our experience, responsiveness is improved by 2 min with a significant reduction (0.85% of the total production) in wasted material.

Acknowledgements

This work was partially supported by the Ministerio de Ciencia y Tecnología de España through the Dirección General de Investigación (DPI2007-61090 and DPI2006-14784 project grants), and the Asociación Nacional de Universidades e Instituciones de Educación Superior (ANUIES) de México through a SUPERA fellowship granted to the professors of the Sistema Nacional de Educación Tecnológica. The authors would also like to acknowledge the support of Metzeler Automotive Profile Systems Ibérica, S.A. and staff who made this project possible.

1 Agard, B and Kusiak, A. 2004. Data-mining-based methodology for the design of product families. International Journal of Production Research, 42(15): 2955–2969. 2 Al-Ahmari, AMA. 2008. Prediction and optimisation models for turning operations. International Journal of Production Research, 46(15): 4061–4081. 3 Braha, D, Elovici, Y and Last, M. 2007. Theory of actionable data mining with application to semiconductor manufacturing control. International Journal of Production Research, 45(13): 3059–3084. 4 Chaudhry, SS and Luo, W. 2005. Application of genetic algorithms in production and operations management: a review. International Journal of Production Research, 43(19): 4083–4101. 5 Chen, MC. 2003. Configuration of cellular manufacturing systems using association rule induction. International Journal of Production Research, 41(2): 381–395. 6 Chen, Z and Wang, L. 2008. Adaptable product configuration system based on neural network. International Journal of Production Research, doi: 10.1080/00207540802007571 7 Cunha, CD, Agard, B and Kusiak, A. 2006. Data mining for improvement of product quality. International Journal of Production Research, 44(18): 4027–4041. 8 Dai, Y, Zhou, C and Yu, W. 2007. Inverse designing simulation of extrusion die of auto rubber seal and verifications. Plastics, Rubber and Composites, 36(4): 141–148. 9 Dean, PR, Xue, D and Tu, YL. 2009. Prediction of manufacturing resource requirements from customer demands in mass-customization production. International Journal of Production Research, 47(5): 1245–1268. del Coz Díaz, J. 2007. Computer simulation of the laminar nozzle flow of a non-Newtonian fluid in a rubber extrusion process by the finite volume method and experimental comparison. Journal of Non-Crystalline Solids, 353(8–10): 981–983. Dengiz, O, Smith, AE and Nettleship, I. 2006. Two-stage data mining for flaw identification in ceramics manufacture. International Journal of Production Research, 44(14): 2839–2851. Feng, JCX and Kusiak, A. 2006. Data mining applications in engineering design, manufacturing and logistics. International Journal of Production Research, 44(14): 2689–2694. Fraley, C and Raftery, A. 2002. "MCLUST: software for model-based clustering, density estimation and discriminant analysis". In Technical report 415, Seattle: Department of Statistics, University of Washington. González Marcos, A. 2007. A neural network based approach to optimize rubber extrusion lines. International Journal of Computer Integrated Manufacturing, 20(8): 828–837. Ho, GTS. 2009. Development of a co-operative distributed process mining system for quality assurance. International Journal of Production Research, 47(4): 883–918. Hsu, PL. 2006. Regression trees approach for flow-time prediction in wafer manufacturing processes using constraint-based genetic algorithm. International Journal of Production Research, 44(24): 5327–5341. Jin, Y and Ishino, Y. 2006. DAKA: design activity knowledge acquisition through data-mining. International Journal of Production Research, 44(14): 2813–2837. Jolliffe, I. 2002. Principal component analysis, Berlin: Springer. Kao, Y and Li, YL. 2008. Ant colony recognition systems for part clustering problems. International Journal of Production Research, 46(15): 4237–4258. Kuriakose, S, Mohan, K and Shunmugam, M. 2003. Data mining applied to wire-EDM process. Data Mining Applied to Wire-EDM Process, 142(1): 182–189. Kusiak, A. 2006. Data mining: manufacturing and service applications. International Journal of Production Research, 44(18): 4175–4191. Lian, J. 2002. Application of data mining and process knowledge discovery in sheet metal assembly dimensional variation diagnosis. Journal of Materials Processing Technology, 129(1): 315–320. Liao, TW, Ting, CF and Chang, PC. 2006. An adaptive genetic clustering method for exploratory mining of feature vector and time series data. International Journal of Production Research, 44(14): 2731–2748. Liu, C. 2009. Uncertain association rule mining algorithm for the cell formation problem in cellular manufacturing systems. International Journal of Production Research, 47(3): 667–685. Mackerle, J. 2006. Finite element modelling and simulation of bulk material forming. A bibliography (1996–2005). International Journal for Computer-Aided Engineering and Software, 23(3): 250–342. Müllner, HW, Wieczorek, A and Eberhardsteiner, J. 2007. "Numerical prediction of the dimensioning of tools for the extrusion process of rubber profiles". In American institute of physics conference seriesEdited by: Cueto, E and Chinesta, F. Vol. 907, 986–994. Ordieres Meré, J. 2003. Intelligent methods helping the design of a manufacturing system for die extrusion rubbers. International Journal of Computer Integrated Manufacturing, 16(3): 173–180. Raddatz, M, Schlüter, M and Brandt, S. 2006. Identification and reuse of experience knowledge in continuous production processes. 9th IFAC Symposium an Automated Systems Based on Human Skill and Knowledge. Nancy, Raheja, D. 2006. Data fusion/data mining-based architecture for condition-based maintenance. International Journal of Production Research, 44(14): 2869–2887. Sarle, SW. 2002. Comp. Ai. neural-nets FAQ, Part 2 of 7: Learning, Should I Normalize/standardize/rescale the data?, Avilable form: http://www.faqs.org/faqs/ai-faq/neural-nets/part2/section-16.html Schreck, G, Lisounkin, A and Krüger, J. 2008. Knowledge modelling for rule-based supervision and control of production facilities. International Journal of Production Research, 46(9): 2531–2546. Shahbaz, M. 2006. Product design and manufacturing process improvement using association rules. Proceedings of the I MECH E Part B Journal of Engineering Manufacture, 220(2): 243–254. Shao, XY. 2006. Integrating data mining and rough set for customer group-based discovery of product configuration rules. International Journal of Production Research, 44(14): 2789–2811. Shi, X, Schillings, P and Boyd, D. 2004. Applying artificial neural networks and virtual experimental design to quality improvement of two industrial processes. International Journal of Production Research, 42(1): 101–118. Shiue, YR. 2008. Data-mining-based dynamic dispatching rule selection mechanism for shop floor control systems using a support vector machine approach. International Journal of Production Research, doi: 10.1080/00207540701846236 Song, Z and Kusiak, A. 2009. Optimising product configurations with a data-mining approach. International Journal of Production Research, 47(7): 1733–1751. Sutanto, P, Picchioni, F and Janssen, L. 2006. Modelling a continuous devulcanization in an extruder. Chemical Engineering Science, 61: 7077–7086. Tseng, TL. 2006. Applying a hybrid data-mining approach to prediction problems: a case of preferred suppliers prediction. International Journal of Production Research, 44(14): 2935–2954. Venables, WN and Ripley, BD. 2002. Modern applied statistics with S, 4, Berlin: Springer. Wang, CH, Wang, SJ and Lee, WD. 2006. Automatic identification of spatial defect patterns for semiconductor manufacturing. International Journal of Production Research, 44(23): 5169–5185. Wong, JT, Su, CT and Hsieh, HT. 2007. A framework for determining MIMO process parameters by a neuro-DM & ACO approach. International Journal of Production Research, 45(15): 3505–3520. Zheng, H. 2007. Computer-aided optimization of the extrusion process of automobile rubber seal. Journal of Macromolecular Science, Part A, 44(5): 509–516.

By Claudia Barreto Cabrera; Joaquín B. Ordieres Meré; Manuel Castejon Limas and Juan José del Coz Díaz

Reported by Author; Author; Author; Author