Continual learning by gradient monitoring for remaining useful lifetime estimation

Sequential addition of tasks in a learning regime leads to catastrophic forgetting of previous knowledge when learning is focused solely on new tasks. This complication can be alleviated by preserving the old data and retraining the network or introducing new networks for corresponding tasks. However, this leads to computation and storage overhead, rendering these solutions impractical in real industrial applications. We introduce a novel Gradient Monitoring (GM) approach to learning incoming tasks and simultaneously preserving old knowledge for predicting the remaining useful life (RUL) of industrial systems. A criterion matrix is formulated to evaluate the relationship between the weights of a network with their corresponding gradients generated during backpropagation. The mean of the criterion, multiplied by a learning factor, generates the threshold for creating a binary masking matrix. We propose two sub-methods of GM namely, Vanilla (VGM) and Momentum (MGM). The binary mask is utilized in VGM whereas, MGM uses the masking memory from previous training iterations along with the binary masks. We perform joint training of task-specific heads using a distillation loss and eliminate the dependency on old datasets. Using theoretical proofs, we demonstrate that GM can greatly reduce backward forgetting of old tasks. We focus our experiments on the RUL estimation of industrial systems and compare our methods with existing frameworks in continual learning (CL). For our CL experiments in RUL estimation, we select the CMAPSS Dataset and the NASA Li-ion Battery dataset, as these datasets offer subsets with varying operational features and diversity in data distribution. • A novel gradient monitoring-based continual learning framework is proposed for predicting remaining useful lifetime estimation. • The framework utilizes a shared encoder network with task-specific decoder heads for all tasks. • Gradient monitoring alleviates catastrophic forgetting in the shared encoder by actively masking gradients based on a criterion-based threshold and masking memory. • Theoretical argumentation is presented to prove that the proposed vanilla and momentum gradient monitoring reduces backward forgetting of old tasks. • Empirical analysis and comparisons are provided for the proposed gradient monitoring framework and existing continual learning methodologies.