Machine Learning for Predictive Maintenance in Banking Infrastructure Services: A Data-Centric Approach

Bharath Somu

Abstract: The banking industry is undergoing a significant digital transformation, leading to the introduction of new microservices and software for infrastructure services. Although these changes can increase the efficiency of a bank?s infrastructure, they can introduce new risks related to system operation. To mitigate risks effectively, an institution must properly monitor system operation and facilitate rapid troubleshooting. One key aspect of ensuring the high operability of banking infrastructure is managing the ?invisible zone of uncertainty? that lies between operational monitoring and reliable failure detection, prioritization, and prediction. The growing microservice architecture in contemporary banking infrastructure systems leads to an exponential increase in number and complexity of the monitoring events being generated. Current approaches to predictive maintenance do not scale adequately. Serve Allocation Problem is an NP-hard one which attempts to allocate costs with a solution working in any kind of model, making it adaptable to any data format. Moreover, few of the applied approaches are focused on the context of banking services. Recent advances in the data-centric field of machine learning and natural language processing with Large Language Models has led to better contextualization of data, making it more understandable to humans. Aiming to mitigate the gap between operational monitoring and failure prediction, this study proposes the first data-centric approach to predictive maintenance on banking infrastructure services. In particular, a working proof-of-concept solution is presented that suggests mapping production monitoring events to predictive maintenance onboarding data using large language models to create and enrich monitoring data contextually. Three unsupervised approaches are implemented to identify false positive monitoring events at the second level of a three-level hierarchy on monitoring event criticality. These methods utilize clustering, dimensionality reduction visualization, and modelling with a probabilistic graphical model to achieve interpretability of ?black box? algorithms and contrast false positives and true positives with an illustrative example. Ultimately, the human-centered nature of banking infrastructure development and operation is acknowledged. Further development paths are suggested in the latent area between the increasing demand for monitoring and deeper contextualization of the monitored systems through AI techniques.

Keywords: Machine Learning, Predictive Maintenance, Structural Health Monitoring, Banking Infrastructure Services, Data Centric Approach