Page:Advanced Automation for Space Missions.djvu/27



There are no standard definitions for compaction and compression, so the following usages are adopted for the purposes of the present study:

Compaction of data - any technique that reduces the size of physical data representation while preserving the relevant information.

Compression of data the application of some function to elements of the database. If x is a specified element of the database then the compression of x is v, where v = /fx). Usually /is invertible, which means that the original information may be recovered whole from the compressed data.

Compaction techniques other than compression involve elimination of information deemed superfluous, in order to decrease overall storage requirements. One such method is abstraction. Abstraction is accomplished by processing data over important common image features (in the case of photographic information) by using, for instance, a world model. After abstraction it is not possible to recover the original image.

Mission data such as are received daily or are already stored in various NASA facilities (e.g., the EROS data center at Sioux Falls, South Dakota) and slated for compaction may be classified within two broad categories - continuous and noncontinuous depending on timing and event characteristics. Classes of continuous data include:

? Periodic data When the same event appears again and again, only one copy need be saved.

? Trendless data If the data are random continuous, a sample should be taken to check for trend. If no trend is found, the data may be represented by a histogram updated regularly as more information accumulates. Only histogram parameters need be saved.

? Data with trend If a trend is detected, multiple regression and curve-fitting are best to record the feature. It may be possible to correlate variation in one time series with that in another (e.g., how sea level is affected by temperature or pressure). Data compression is achieved in this case by fitting polynomial segments, possibly straight lines, to the data.

? Data with turning points Turning or "inflection" points, such as where an upward trend suddenly shifts downward in a time series graph, may require different models to be fitted to the two parts of the series. Only the parameters of fitted polynomials and a few statistical abstractions (e.g., maxima and minima, mean, variance, and several others) in any particular range need be saved.

Suggested classification and processing techniques for non- continuous numerical data are summarized in figure 2.7.

2.3 The World Model

The world model is a crucial element in the achievement of specific goals. Without a sophisticated model two serious problems are encountered with remote Earth-sensing data, particularly images:

? It is very difficult, if not impossible in many instances, to accurately separate interesting from non- interesting observations.

? It is difficult to comprehend raw sensor data in terms readily understandable by human beings.

The first of these leads to the collection and retention of great volumes of data, simply because there is no practical way to perform an appropriate selection of the useful subset of information applicable to a user request. The second problem results in gross underutilization even of potentially useful data. The lack of a world model in present-day spacecraft makes necessary a voluminous and costly stream of highly redundant data which must be transmitted and collected on the ground before any useful information is retrieved, leaving a huge reservoir of unprocessed data in expensive storage facilities. It is the world model which transforms IESIS from a collection of remote cameras into an entity able to perceive the planet in a manner interesting and informative to humans. This world model is a compact representation of persistent spatial and temporal characteristics of the Earth (its land, oceans, and atmosphere), and algorithms for use of the model.

2.3.1 World Model Structure

The IESIS world model has two separate components. The first is the state component, which defines the physical status of the world to a predetermined level of accuracy and completeness at a specified time. Second is the theory component that allows derivation of the following information from the state component: (1) Values of parameters of the world state not explicitly stored in the state component, and (2) a forecast of the time evolution of the state of the world. The theory component gives the world model a predictive capability, in that the model can predict facts about the world not explicitly retained in the database.

The disparity between predicted information and reality generally increases with increasing computational distance separating the starting information and the derived result, increasing time in the case of forecasting, and certain other factors. The world model requires a continual influx of new observations to remain temporally current. A major research goal for efficient IESIS operation is to develop the AI capability to construct an effective real-time world model which can act as a database for answering questions