|
|
|||
|
|
|
|
|
|
||||
|
Company History Key People News Newsletter Customer Profiles Contact Us Employment |
Category: Engineering Tracking the Right Clues with Exploratory Data Analysis By John V. James Statistical analysis of data generally aims to uncover the essential characteristics of and relationships among measured variables. In most analyses, traditional statistical tools and methodologies are central, but the benefits of a somewhat different way of examining data- exploratory data analysis- are being discovered by more and more engineers. The analyses here were derived with the aid of Data Desk, by Data Description, Ithaca, N.Y., which is designed especially for exploratory data analysis. Many statistical software companies are now adding features from this approach to their own, more traditionally oriented packages. The use of these tools is illustrated by an engine analysis project in which the author was involved at Ford Motor Co., Dearborn, Mich. The project got its start from recent additions to Federal and state automotive regulations that mandate on-board monitoring of all engine functions affecting exhaust gas emissions. One new requirement is the ongoing detection of combustion failures within the cylinder. Such misfires let fuel vapors pass through the engine unburned, and a misfire rate of just a few percent may increase emissions significantly. But detecting misfires under virtually all operating conditions of the engine, as required by law presents quite a problem. One reason, ironically enough, is that the speed loss due to a misfire is tiny, due to the inertia of the engine's flywheel. Then, too, crankshaft speed normally fluctuates continuously, and the signals of this normal behavior mask the signals from those other decelerations due to misfires. Besides the signals from the crankshaft-position sensor, Ford's analysis team used data acquisition equipment (unavailable on commercial vehicles) so that they might develop and test misfire detection algorithms off line. Broadly speaking, we looked at every "firing event"- both normal firings and misfires. The data samples were not spaced evenly in time, being gathered at every firing interval, which of course varies with engine speed. From here on, exploratory data analysis will be shown being used in four ways to develop and evaluate various misfire-detection algorithms [and to thus simulate the misfire process]. Using exploratory software on one's own data, interactively, is the best way to appreciate its value. This is not to devalue traditional statistics. Rather, some tools free us to follow the analysis suggested by our own observations of the data. Careful observers formulate questions, software programs do not. But without exploratory data analysis tools, many productive WHAT IF questions could go unasked. |
Name: John James Company: Ford Motor Company Location: Dearborn, Michigan |
|||
|
|||||
|
|
|||||
