Multi-Instance Attention Network for Anomaly Detection from Multivariate Time Series

Abstract

Anomaly detection and state prediction research using multivariate data is being actively conducted in various industrial fields. However, since most dynamically operating industrial machines perform different operating conditions, they contain different types of abnormal conditions, making it difficult to detect anomalies and predict the remaining life. This white paper proposes a condition diagnosis model based on multi-sensor data prediction for estimating the remaining lifespan of equipment while solving two complex problems of detecting four typical abnormal conditions and sensor omissions of industrial machines. First, we use a multi-sensor data generation model to learn relationships between sensors, and second, we use a sensor data prediction model to learn sensor-specific feature information. In order to extract the temporal and spatial characteristics of sensing information and to derive the relationship between the sensors, we propose an attention model with three types of cases. Finally, the state of the device is diagnosed through the difference between the model predicted value and the actual value, and future state information of the device is predicted through the accumulation of error information. In order to prove the robustness of the proposed model, extensive experiments were conducted focusing on the case where sensor omission occurred due to data from equipment with more than 4 types and conditions. Our model produces missing sensor data with about 92% accuracy and detects anomalies with about 88% accuracy, even if parts of the sensor are missing or the operating environments have been changed. The proposed model has improved anomaly detection accuracy compared to the comparative model, and has been proven to be applicable to real industrial problems.