RNN Failure Prediction Using Memory Depth Segmentation

The patent segments the long inter-failure gap prediction problem into multiple smaller sub-problems by dividing the time gap into segments. Each segment is processed by the RNN separately, avoiding the need to process the entire long sequence at once. This segmentation allows the standard RNN to handle long gaps effectively by breaking down the vanishing gradient problem into manageable chunks that can be processed sequentially.

The patent performs preliminary actions by pre-processing the failure history data and preparing segmented sequences before feeding them to the RNN. The system pre-calculates and stores intermediate results for different segment lengths, which are then reused during prediction. This preliminary preparation enables the standard RNN to efficiently handle long inter-failure gaps without requiring specialized architecture modifications.

The patent makes the standard RNN architecture universal by demonstrating its ability to handle both short and long inter-failure gaps through the segmentation approach. The same basic RNN structure and training procedure are used regardless of the gap length, eliminating the need for different specialized architectures. This multi-functionality allows a single simple model to serve multiple prediction scenarios effectively.

The patent changes the parameter of sequence length segmentation rather than modifying the fundamental RNN architecture. By adjusting how the input sequence is divided into segments and the corresponding processing parameters, the system adapts to different gap lengths using the same underlying model. This parameter-based adaptation avoids the complexity of architectural modifications while maintaining prediction accuracy.