Fault Tolerant Neural Predictors for Compression of Sensor Telemetry Data

When dealing with remote systems, it is desirable that these systems are capable of operation within acceptable levels with minimal control and maintenance. In terms or transmission of telemetry information, a prediction-based compression scheme has been introduced. This paper studies the influence of some typical transmission and network errors on the encoded residue stream produced by a number of predictors used in the scheme, with the intention of identifying the more fault tolerant architecture that may be preferred as predictors. Classical linear predictors such as FIR and lattice filters, as well as a variety of feedforward and recurrent neural networks are studied. The residue streams produced by these predictors are subjected to two types of commonly occurring transmission noise, namely gaussian and burst. The noisy signal is decoded at the receiver and the magnitude of error, in terms or MSE and MAE are compared. Hardware failures in the input receptor and multiplier are also simulated and the performance of various predictors is compared. Overall, it is found that even small low- complexity neural networks are more resilient to faults due to the characteristics of their parallel architecture and distributed storage/processing characteristics.