Uncertainty Assessment-Based Active Learning for Reliable Fire Detection Systems

Deep learning technologies, due to their advanced pattern extraction and recognition of high-dimensional data, have been widely adopted into multisensor-based fire detection systems. Since deep approaches can generate erroneous predictions incomplete training datasets, a retraining process over unseen observations is needed. However, storing large amount data from continuous multisensor streams labeling them create dataset are costly time-consuming. In this paper, we propose an active framework based on informative experience memory that populated with meaningful by assessing the uncertainty data. proposed framework, model predicts occurrence estimates taking advantage Bayesian neural network using Monte Carlo dropout. By only higher uncertain points fixed-size querying system managers, storage costs minimized while improving performance. To evaluate our different structures, develop three networks conventional classification networks, including feedforward network, fully convolutional long short-term memory. We further investigate various assessment scoring methods for tasks such as entropy, BALD, variation ratios, mean STD. Experiments real show FCN BALD method has highest performance gain F1 score 0.95, improvement 24% 700 points.