Predictive Maintenance for ATMs using AI

Training a machine learning model to predict target operating values for monitored devices like meters, and then using the model to identify anomalies in the actual operating values. The model is trained on a combination of real data and fake data with noise, and learns relationships between successive time-series values. By comparing predicted values to actual values, anomalies can be detected. The approach involves adding noise to some training data points to teach the model to handle variance, and ranking anomalies based on patterns to prioritize remediation.

Source

Proactively detecting power supply degradation in ATMs before failure to prevent downtime and customer inconvenience. The method involves training a machine learning model using historical maintenance records to predict scores indicating likelihood of power supply degradation based on preprocessed ATM event data. The scores are reported to the ATM's enterprise, allowing proactive action like closely monitoring or replacing the power supply if scores exceed thresholds.

Source

A data analytics system for automated insight generation in domains like predictive maintenance of devices. The system uses self-contained insight providers, each with a domain-specific model and configurable parameters, to analyze asset data. The providers can operate in filter or enrichment modes. A pipeline of providers can be chained. The system provides a user interface to interact with the providers and drill into analytics results. It aims to enable broader use of data analytics by providing specialized, easily configurable, and chainable insight providers instead of complex data science.

Source

Preventative maintenance and customer service platform that uses AI to predict future events and determine preventative solutions. It processes real-time and historical data to find future events, extracts features to prioritize solutions, and generates an intelligent assistant to perform them. This reduces service requests by addressing issues proactively instead of waiting for customer complaints.

Source

Automated self-learning ATM security system that enables an ATM to decide whether its security is under threat and learn to fight such threat in real time using means made available to the ATM when the threat arises. The system involves extracting data points from the ATM's transition flows between states and generating rules for each transition based on the extracted data. If a new transition's data doesn't match existing rules, it's flagged as potentially unsafe. This allows the ATM to learn from its own mistakes and adapt its security without external intervention or static rules.

Source

Determining asset health scores using digital twin resources to provide more accurate and trustworthy asset health predictions. The method involves retrieving a digital twin with similar characteristics to a physical asset and using its health formula to predict the physical asset's health score. This leverages the digital twin's detailed data and modeling to provide more reliable health assessments for physical assets.

Source

Optimizing ATM replenishment schedules using machine learning to analyze user data and generate more accurate forecasts. It retrieves unstructured user data like social media posts, generates user personas, then combines with structured transaction data and deep learning to create optimal replenishment schedules. The user personas provide insight into user behavior and consumption patterns beyond just transaction history. This allows predicting when an ATM will run out based on user characteristics and generating optimized schedules for replenishment.

Source

Automated device issue resolution using machine learning to predict and prevent configuration-based issues in devices. The system learns from historical data to identify device configurations that lead to issues and recommends resolutions. It then automatically applies those resolutions to similar devices to prevent issues before they occur. The machine learning models determine device issues and similarity, and recommend actions based on learned configurations. This provides proactive device self-healing by predicting issues and resolving them across similar devices.

Source

Prioritizing maintenance schedules for hardware devices like ATMs using predictive failure scores that are prioritized by business impacts. The method involves generating predictive failure scores for each device, setting windows for servicing, determining missed services, prioritizing devices by business impact factors, and generating schedules with lower priority devices having missed services. This maximizes business impact by focusing maintenance on devices with the highest predicted failures and the greatest business impact if unserviced.

Source

Simultaneously monitoring multiple system indicators using a long short-term memory (LSTM) model trained on historical data. The LSTM model can learn co-movement relationships between indicators from historical data. By inputting historical data into the LSTM and comparing it to real-time data, anomalies in co-movements between indicators can be detected. This avoids false alarms and missed anomalies from single indicator monitoring.

Source

Cash demand prediction system that improves accuracy and interpretability by dynamically selecting the prediction formula based on whether it's a day with expected cash inflow/outflow. The system generates predicting data with a flag indicating if it's a day like payday or pension payment. A learned model with prediction formulae for different explanatory variable values is used. The prediction device selects the formula based on the flag value in the predicting data, then applies it to predict cash demand. This allows tailoring the prediction to day-specific factors for better accuracy.

Source

Intelligent maintenance recommendation system that uses machine learning to predict equipment failures and optimize maintenance schedules. The system learns failure prediction models from historical data and continuously evaluates and switches between models based on cost functions to minimize false alarms and missed failures. It applies the best model to equipment for a period, then re-evaluates and switches based on feedback. This adaptive model selection improves prediction accuracy compared to fixed models.

Source