This technique has applications in intelligent signal analysis, data mining in aerospace manufacturing, military target recognition, and intelligence gathering.
NASA’s Jet Propulsion Laboratory, Pasadena, California
The communication between NASA space mission operations teams and their respective spacecraft in outer space is accomplished via the Deep Space Network (DSN). To ensure proper operations in returning telemetry data to mission operations, sending commands to spacecraft and providing radiometric data for navigation purposes, the DSN equipment generates a large set of self-monitor data. System noise temperature (SNT) and link margin are two of the key metrics of system performance for telemetry data return. The ability to detect the signal is affected by the SNT; the lower the noise, the better chance the system can detect the signal. Thus, there is a strong interest in monitoring and classifying the SNT. However, there are many causes of SNT behavior patterns that are still not known and difficult to classify by simple logic. There is a benefit to use a more sophisticated pattern-recognition method to detect and classify abnormal SNT behaviors.
In addition to classifying the pattern of abnormal SNT, it is also important to understand the cause of such instances. High SNT, which results in a lower link margin, is often caused by bad weather. A correlation among these three parameters would establish whether such causal effect indeed occurs in a given track. In this data analysis, a reasonably good correlation is detected among high SNT, low link margin, and bad weather. The proposed new method could provide a more reliable and more intelligent monitoring of the DSN operations.
A fuzzy logic function trained by an artificial neural network was developed to classify the SNT of antennas in the NASA DSN. The SNT data were classified into normal, marginal, and abnormal classes. In order to capture various irregular patterns of SNT data, a set of features of the SNT curves was designed: mean values, standard deviations, peak numbers, peak-to-valley variations, and slope of the peaks. The SNT data from various tracking passes were processed to extract the SNT curve feature vectors; a feature vector represents each pass. Since the criteria were not simple Boolean operations and there was possible need for adding non-threshold criteria in later analysis, a simple threshold approach was deemed not suitable for classifying the SNT patterns. Instead, a fuzzy logic was designed to classify the SNT patterns, and a neural network was then used to train the fuzzy logic.
A three-layer feed-forward neural network was constructed with six neurons in the input layer to represent the six features, six output neurons for the six categories, and six hidden layer neurons were chosen to accommodate nonlinear boundaries, as shown in Figure 1. In the neural network training process, a domain knowledge expert picked a set of training SNT data and assigned the SNT data (training inputs) into correct classes (target outputs). The output result from neural network was compared to the target output; the output error was used to back-propagate through the network to tune the weights. This learning process was repeated many times until the output error of the neural network was less than a pre-defined value. Figure 2 shows an example of the results of neural net classification of SNT data quality in three categories: “Good,” “Marginal,” and “Bad” on a data set collected from all 12 DSN antennas (Deep Space Station (DSS)) that tracked the Mars Reconnaissance Orbiter (MRO) spacecraft in the first 2 months in 2007.