Abstract
This research-in-progress paper describes a study of data mining using historical manufacturing test data at an electronic circuit board assembly firm. The purpose of this data mining is to determine, if in a multi-step manufacturing process at small and medium manufacturing firm downstream production problems can be identified and prevented by evaluating up stream data. While quantitative analysis and data mining have been used in manufacturing, the scope of their application is generally limited to a specific step or process that is a first order analysis of the data.
Introduction
The manufacturing field has long been aware of and has applied quantitative methods in the manufacturing process. Yet, the use of data mining techniques in manufacturing is generally not heard of. For example the term manufacturing appeared in the title of only 4 of the 216 articles that were reviewed in a survey of data mining applications from 2000 to 2011 (Liao, Chu, and Hsiao, Citation2012). There are several possible reasons for this low occurrence. One contributing factor is that before the term data mining became popular, there were many quantitative tools that have been applied and accepted in manufacturing. However, the scope and domain of the techniques that have been applied in manufacturing are generally narrow in their application.
In the domain that we studied of electronic circuit board assembly the most common application is the use of quantitative tools/techniques at particular steps in the manufacturing process. Research has often been done in these narrow domains of specific step in the manufacturing process. Examples would be using a modified k-means algorithm, and neural networks to improve the performance of the surface mount technology (Ng, Citation2000; Vainio et al., Citation2010). The reporting of more holistic approaches has been limited in the literature (Harding, Shahbaz, and Kusiak, Citation2006). As manufacturing departments at small and medium firms are often very sensitive to costs and since there is a cost associated with doing data mining of a more holistic nature, there is a natural reluctance to engage in it without hearing about other firms’ success in using it. The purpose of this research-in-progress is to examine a more holistic approach to data mining. Specifically, the objective is to study whether the current process capabilities are sufficient or whether data mining should be used to seek improvements on a wider portion of the manufacturing process.
Data mining in manufacturing electronic circuit boards
Data mining in manufacturing has been applied at specific steps and process in manufacturing environments, e.g. the creation of rules for scheduling of products in manufacturing (Kim, Citation2015). In the area of quality management (Köksal, Batmaz, and Testik, Citation2011) provides a survey of the literature from 1997 to 2007. Another quality related study involving automating of inspection using optical equipment is reported by (Benedek, Krammer, Janóczki, and Jakab, Citation2013). (Liukkonen, Havia, and Hiltunen, Citation2012) provides a good survey of data mining techniques used in the development of methods and algorithms of the optical inspection of solder joints. In practice the substantial engineering and developments costs in the specialised equipment to perform optical inspection has caused the creation of 3rd party vendors to create and sell equipment for this specialised function to small and medium manufacturers. This 3rd party equipment has software built into it that incorporates some of the results and methods learned from those studies. They do so by allowing the specification of some the inspection parameters to be set by product by the manufacturer.. Hence this 3rd party inspection equipment is purchased by manufactures allowing for a first order quality analysis evaluation at those points in the process where the equipment is used. This type of equipment often allows this inspection data to be extracted and accessed for further analysis. This data is generally used to evaluate that specific step in a manufacturing process. However if captured and stored by parts that have unique serial numbers as part of the inspection data the inspection results that may be of value to mine across the manufacturing process. The need to incorporate data mining tools and AI algorithms across the manufacturing process has been suggested by many researchers. (Di Orio, Cândido, and Barata, Citation2015; Guha, Citation2015). However, it should be noted that even some of these researchers acknowledge that this needs to be proven. In their own words they state:
The key assumption is that a deeper use of context awareness and integrated data mining techniques applied to the manufacturing production systems will allow (on-line) identification of current dynamically changing context in which the manufacturing production system operates and adaptation of process parameters according to the detected changes in context. (Di Orio et al., Citation2015)
Overview of manufacturing electronic circuit boards
The basic process used in assembling electronic circuit boards is broken down as follows in an assembly line process.
| (1) | Bare printed circuit boards are placed at the beginning of an assembly line. | ||||
| (2) | A machine applies solder paste to one side of the board at the locations where components will be soldered. | ||||
| (3) | A machine optically inspects the paste on the boards. Any product that does not conform to a configured inspection criterion is removed from the line. | ||||
| (4) | Electronic parts are robotically inserted by a specialised piece of equipment | ||||
| (5) | The boards with the inserted parts are placed through a machine that does the actual soldering. | ||||
| (6) | The boards are then optically inspected to make sure the solder joints are good. | ||||
| (7) | The boards are tested electronically | ||||
It should be noted that at steps 3, 6, and 7 above the results of the test or inspection are compared to predefined parameters and rejected and removed if not within the specified tolerance. The other side of the board then goes through the same basic process. Figure shows some of the data that is collected at each of these steps. A more detailed explanation and review of some of the related quality issues can be found in (Tsai, Citation2012; Wang & Le, Citation2015).
Figure 1. Example of data generated at each step.
While there may be additional steps that may be included in a manufacturing setting, this is the basic process that is generally followed. While the process is largely automated, there is still a small component of human intervention needed. Hence, one aspect of data in manufacturing that is different from data mining applications in other domains, is that the variability is primarily can be attributed to variability in machine related aspects as opposed to human behaviour.
The minimization of human involvement does not mean there are no levers that can be used to change the process. Indeed, at just about every step that is automated there are parameters that have been configured that could be set at a different level. The configuration often requires the person doing them to possess a high level of knowledge and expertise in that area. The appropriate configuration of the parameters is not part of this study, since in practice they are seldom changed and would not contain the variability needed to provide any insight.
In general, the farther along the production process that a product is found with defects, the greater are the chances of having to scrap the circuit board. Hence, the early inspections that are done are critical. These inspections however are one of the steps that are configured. There are many trade offs that have to be made. If the inspection criteria is set too rigidly, then the product may be rejected and subject to unnecessary rework. However, inspection criteria that is not rigid enough may cause the product be rejected later, or worse still end up as a defective product shipped to the customer.
The inspection that takes place at particular points in the manufacturing process is what we refer to as first order analysis and quality control. They tend to be the main points where the majority of production problems are found and identified.
In addition to the existing first order analysis, data mining of the detail test data can potentially provide the following: (1) guidance of when parameter for inspection may be too lax, (2) potential identification of non-linear or interaction problems between steps, and (3) some insight into manufacturing processes that might potentially be changed. This is consistent with but a perquisite first step to the vision quoted earlier from (Di Orio et al., Citation2015).
Case study
The firm studied produces electronic circuit board assemblies. This particular study involves examining data that was captured in producing over 600,000 circuit boards. Due to customer requirements, the firm assigns a unique serial number to each board produced and tracks data at over a dozen points in the manufacturing process. Among other items, this data consists of detailed inspection-data captured by serial number from equipment that is used as part of the quality and manufacturing process. The existing quality controls and process have resulted in this firm in maintaining a very high quality. The return rate from the field has been less then 30 boards out of over 600,000 boards shipped.
The data that has been accumulated is occasionally used in a retrospective analysis of an issue such as when a circuit board is returned from the field. The firm does not know if there is any additional value of real time analysis beyond the first order quality control checks that are in place. Two areas that might be impacted by real time analysis of the data are returns and scrap. Given the small return rate of units from the field one could argue that the cost of developing and maintaining a system would not be worth doing any real time analysis beyond the existing system. There may however be some value in the ability to reduce the amount of product that is scrapped. The firm has partnered with the researchers to evaluate utility of data mining.
Data mining process
The CRISP-DM methodology (Shearer, Citation2000) is being followed in the following manner:
| (1) | Business Understanding: Discussions with the firm has provided a basic understanding of the project and its goals. | ||||
| (2) | Data Understanding: Understanding the meaning of the data and ensuring its quality is essential. One of the researchers on this project has extensive knowledge of this data, as he was involved in designing the data collection system and database. To ensure data quality a quality audit is being performed. | ||||
| (3) | Data Preparation: The project is currently at this stage. This involves many steps: initially we are identifying, for each board that has been assigned a serial number, its ultimate status (shipped, scrapped, in repair, unknown). This has caused some revisiting of the steps 1 and 2. Some of the issues currently being worked through are (1) repair information on the boards has been manually kept in an excel spreadsheet. As such, the repair data contains more errors than the data from the manufacturing process. (2) There are thousands of boards, which were started but never shipped that we have yet to account for. (3) The processing of the volume of data within the firms infrastructure is a concern. Decisions as to how much data to analyse at a given time will need to be addressed. We have not at this time identified the specific software tool that we will be able to use. The firm does not want us to take data off site so we are looking at the pragmatic issues of what our software licence allows. | ||||
| (4) | Models: The problem domain of interest will primarily be related to prediction and classification. The limitation mentioned above of working within the firms’ infrastructure might restrict some of our modelling choices. | ||||
| (5) | Evaluation: The resulting findings will be documented and presented to management and written up for the academic community. | ||||
| (6) | Deployment: Any deployment as a result of the findings will be left up to the firm, which provided access to their data. | ||||
Expected contributions
| (1) | Providing some empirical data in an area where there is a paucity of studies and in turn provide more insight into the domain | ||||
| (2) | Raising awareness of the need for research in this area of multiable-step manufacturing. | ||||
| (3) | Providing guidance to firms engaging in data mining across multiple steps in the manufacturing process. | ||||
| (4) | Benefits to the specific firm studied will be additional knowledge as to whether upstream data can capture downstream problems beyond their current first order controls. | ||||
Limitations
This study proposes to the determine if by data mining historical test data if in a multi-step manufacturing process downstream production problems can be identified and prevented by evaluating up stream data. The study results will be limited in that the number of ways of analysing the data is quite large and only a subset would be used.
Disclosure statement
No potential conflict of interest was reported by the authors.
References
- Benedek, C., Krammer, O., Janóczki, M., & Jakab, L. (2013). Solder paste scooping detection by multilevel visual inspection of printed circuit boards. Industrial Electronics, IEEE Transactions on, 60, 2318–2331. 10.1109/TIE.2012.2193859
- Di Orio, G., Cândido, G., & Barata, J. (2015). The Adapter module: A building block for self-learning production systems. Robotics and Computer-Integrated Manufacturing, 36, 25–35. 10.1016/j.rcim.2014.12.007
- Guha, S. (2015). Process monitoring and predictive analytics in multi-stage manufacturing processes using partial least square methods ( Unpublished master’s thesis), The University of Texas at San Antonio, San Antonio, TX.
- Harding, J., Shahbaz, M., & Kusiak, A. (2006). Data mining in manufacturing: A review. Journal of Manufacturing Science and Engineering, 128, 969–976. 10.1115/1.2194554
- Kim, D. (2015). Learning dispatching rules via an association rule mining approach ( Unpublished master’s thesis), Iowa State University, Ames IA.
- Köksal, G., Batmaz, I., & Testik, M. C. (2011). A review of data mining applications for quality improvement in manufacturing industry. Expert Systems with Applications, 38, 13448–13467. 10.1016/j.eswa.2011.04.063
- Liao, S.-H., Chu, P.-H., & Hsiao, P.-Y. (2012). Data mining techniques and applications – A decade review from 2000 to 2011. Expert Systems with Applications, 39, 11303–11311. 10.1016/j.eswa.2012.02.063
- Liukkonen, M., Havia, E., & Hiltunen, Y. (2012). Computational intelligence in mass soldering of electronics–A survey. Expert Systems with Applications, 39, 9928–9937. 10.1016/j.eswa.2012.02.100
- Ng, M. K. (2000). A note on constrained k-means algorithms. Pattern Recognition, 33, 515–519. 10.1016/S0031-3203(99)00057-6
- Shearer, C. (2000). The CRISP-DM model: The new blueprint for data mining. Journal of data warehousing, 5, 13–22.
- Tsai, T.-N. (2012). Development of a soldering quality classifier system using a hybrid data mining approach. Expert Systems with Applications, 39, 5727–5738. 10.1016/j.eswa.2011.11.097
- Vainio, F., Maier, M., Knuutila, T., Alhoniemi, E., Johnsson, M., & Nevalainen, O. S. (2010). Estimating printed circuit board assembly times using neural networks. International Journal of Production Research, 48, 2201–2218. 10.1080/00207540802572574
- Wang, L. W., & Le, Q. L. (2015). Six-sigma and taguchi approaches to the printed circuit board quality improvement. Research in World Economy, 6, 29–52.