The paper proposes an approach to improve the efficiency of machine learning models used in monitoring tasks using metric spaces. To solve this problem, a method is proposed for assessing the quality of monitoring systems based on interval estimates of the response zones to a possible incident. This approach extends the classical metrics for evaluating machine learning models to take into account the specific requirements of monitoring tasks. The calculation of interval boundaries is based on probabilities derived from a classifier trained on historical data to detect dangerous states of the system. By combining the probability of an incident with the normalized distance to incidents in the training sample, it is possible to simultaneously improve all the considered quality metrics for monitoring - accuracy, completeness, and timeliness. One approach to improving results is to use the scalar product of the normalized components of the metric space and their importance as features in a machine learning model. The permutation feature importance method is used for this purpose, which does not depend on the chosen machine learning algorithm. Numerical experiments have shown that using distances in a metric space of incident points from the training sample can improve the early detection of dangerous situations by up to two times. This proposed approach is versatile and can be applied to various classification algorithms and distance calculation methods.

Keywords: monitoring, machine learning, state classification, incident prediction, lead time, anomaly detection