Formation and validation of the feature alphabet in a comparator model of preventive software maintenance

Abstract

Modern software systems operate in dynamic environments where performance degra-dation may gradually accumulate and eventually lead to service disruptions or system fail-ures. Preventive software maintenance aims to detect early symptoms of degradation before critical incidents occur. However, reliable early diagnosis remains challenging because in-ternal system states are not directly observable and must be inferred from operational teleme-try such as logs, monitoring metrics, and user interaction data.
Recent research in software reliability and software aging detection primarily focuses on statistical analysis of telemetry data and machine learning techniques for anomaly detec-tion and predictive maintenance. Despite significant progress in these areas, considerably less attention has been paid to the formal construction of interpretable feature systems that can be used for operational state diagnosis and preventive maintenance decision-making.
The purpose of this study is to develop and formalize an approach to forming and vali-dating a feature alphabet for a comparator model of preventive software maintenance. The proposed model represents software system states using a finite set of observable features de-rived from operational metrics. These features are transformed into Boolean predicates that describe the current operational state of the system. Based on these predicates, a comparator identification model is constructed that classifies system states into operational response classes such as normal state, degradation state, and problem state.
The research method is based on the comparator identification approach, where system states are represented as vectors of binary features and classified using logical predicates expressed in disjunctive normal form. This representation enables interpretable state classifi-cation and allows engineers to understand which combinations of operational indicators cor-respond to different system conditions.
To evaluate the proposed approach, an experimental study was conducted using a syn-thetic event log simulating the operation of a distributed mobile application system. The ex-perimental environment modeled normal operation as well as controlled degradation scenar-ios, including performance degradation, server overload, user experience deterioration, and security anomalies. Operational events were aggregated using sliding time windows, and a set of operational indicators was calculated for each window.
The results of the experiment demonstrate that the comparator model can effectively identify degradation symptoms based on operational metrics. The model correctly classified approximately 93% of system states and was able to detect degradation symptoms before crit-ical incidents occurred. In addition, the model identified previously unseen operational states, indicating the possibility of adaptive refinement of the feature alphabet during system opera-tion.
The obtained results confirm that the proposed comparator model and the feature al-phabet formation approach can serve as an effective instrument for preventive software main-tenance. The model provides interpretable diagnostics, supports early detection of degrada-tion processes, and can be extended through iterative refinement of features and logical rules during system operation.