What You Should Know about Data Mining

Data mining is a tool to guide the speedier development of better products.

Article From: Time Compression, Rob Meyer, D. Sc. and Stephen Langdell Ph.D.

Posted on: 1/1/2004

The modern world, design engineers included, is awash in data. Rather than be swept away in this tidal wave, more forward-thinking product developers are beginning to pay attention to how data mining tools can be used to their advantage.

What is data mining? A standard definition is the nontrivial extraction of implicit, previously unknown and potentially useful knowledge from data. Another definition is a variety of techniques used to identify nuggets of information or decision-making knowledge in bodies of data. For the design engineer, data mining is a potential tool to guide the development of better products and a way to get at these new product specifications faster.

Data that is relevant to the design engineering process may come from many sources. Microprocessor-controlled machines generate much data in their normal course of operation. Many manufacturing processes similarly create data that is used to monitor quality, throughput, etc. Repetitive cycles of prototype testing or simulations of complex products also have potential to create vast quantities of data. For example, more indirect sources of data might come from marketing department-driven studies of customer satisfaction. Whatever the source, the promise of data mining is to discover new information on product function that is not available from a more routine statistical analysis to guide the engineering of a better machine and to get to these product definitions faster.

In machine design, there are both hardware and software design issues. Data mining can speed things up in two ways—first, analysis of operating data, including failures, on existing machines can reveal potential improvements in terms of quality, uptime and throughput. Secondly, most computerized machinery has a significant component of control software. This means that as operating data is analyzed, even after prototype and production machines are built, data mining results can more quickly identify the need for software changes.

David Harris, president of David Harris Design Associates (Chicago, IL), which specializes in fast-track product development, describes the opportunity that data mining provided to his group's design of a motion-controlled operating room table, "Previously, we just had a hydraulic machine. When we added robotic control, we saw an immediate benefit. Unlike its dumb predecessor, the robotically-controlled machine was generating operational data. With data mining tools, we were able to work with this data and mine it for a great deal of valuable information, such as how to develop an optimized service protocol and research actual use in creating the specification for the next generation product."

"This is by no means a unique example. While machine development is dependent on many variables, the one that is often overlooked is the quality of the design specification. These documents usually are the end result to applying collective wisdom to an initial specification, averaged from a list of drivers that may include the following:

Old machine specifications

Competitive machine specifications

Customer requests

Service insights

New technology capabilities

You know you are making a faster smaller machine, but can you be sure that you are designing the right machine? Machine operating data can add another perspective. Initially, this data, whether from one machine or a large population of machines may look like a large, undecipherable database. However, using data mining tools, it can provide key insights into how machines are used, and therefore how they may be improved. Machine data does not lie and it is not anecdotal. It is just a different language," explains Harris.

While most design engineers will rely on mathematical experts to implement data mining routines, a general understanding of how data mining proceeds will empower a product development team to rapidly guide data mining exercises to meaningful economic results.

Generally speaking, data mining protocols include three discrete stages:

1. Data cleaning
2. Data transformation
3. Model building.

The glossary below defines many of the terms used in data mining that are important for machine designers to know.

Glossary of data mining terms

1. Data cleaning
The methods used to replace missing or incorrect values in data.

2. Data transformation
Creating new variables from original variables, e.g., principal component analysis to reduce the dimensionality of data.

3. Clustering
A popular technique that can be used either to reduce the number of data or for data exploration.

(a) k-means clustering
The clustering technique known as k-means is used to cluster data into a number of groups. The number of groups, k, must be specified in advance by a user.

(b) Hierarchical clustering
Hierarchical clustering is used to explore the number of clusters that occur naturally in a set of data. Typically, hierarchical clustering assigns each datum to a group of its own and then iteratively merges groups until only one group remains. Patterns created by the merging of groups are analysed by users who then decide how many groups are appropriate for the data. This highly informative approach is limited to lower numbers of data.

4. Classification
Classification is the process of assigning data to one of a set of known classes. The following data mining techniques can be used for classification.

(a) Decision tree
Decision trees recursively split data into groups by using tests derived from the data. By tracing from the root node of a decision tree to each leaf node, a set of rules can be derived.

(b) Logistic regression
A logistic regression can be used when the outcome response is binary, such as yes or no. For example, the outcome might be: will the customer buy a particular product or will an insurance client make a claim? Logistic regression relates the probability of a yes answer to the values of a number of explanatory variables.

5. Prediction
Not all data can be thought of as belonging to one of a number of classes (e.g., the value of products sold. These data sets can be mined by using prediction models).

6. Classification or prediction
The following data mining models can be used for either classification or prediction.

(a) k-nearest neighbors
The k-nearest neighbors algorithm returns a value based on the k most similar data. For classification problems the majority class is returned, whereas the average value is returned for prediction problems. The measure of similarity is usually a distance function.

(b) Multilayer perceptron neural networks
A multilayer perceptron (MLP) is a very flexible model suitable for complicated problems. Values for the parameters in the model are found by iterating an optimization algorithm and, therefore, the MLP is suitable only for small data sets.

(c) Linear regression
A technique used to build simple models from data. Values for the parameters in a model are found by using only one pass through the data, making linear regression suitable for large data sets.

What You Should Know about Data Mining

Glossary of data mining terms

Zones

Supplier Categories