From Condition Monitoring to AI-assisted Anomaly Detection: An Industry 4.0 Paradigm
From Condition Monitoring to
AI-assisted Anomaly Detection:
An Industry 4.0 Paradigm
Author: Ioannis Meintanis
12th December 2022
Nowadays, digitalization and automation based on the interconnection of heterogeneous systems have become an integral part of our daily lives.
Behind all these buzzwords the main goal is to use technology and data to increase productivity and efficiency.
Industry is keeping pace with the latest developments by integrating new technologies that can boost its productivity and flexibility exponentially. Manufacturers are adopting new technologies, including Internet of Things (IoT), cloud computing, data analytics and Artificial Intelligence (AI) throughout their production operations.
Condition Monitoring
Condition monitoring (CM) systems are an essential element of this transformation. Predictive maintenance is based on the condition monitoring of the assets to determine whether they will fail during some future period and on taking appropriate actions to prevent the consequences of such failure. Monitoring machines with live sensor data enable humans and intelligent systems to use vast amounts of available data to extract useful information from their production processes, making it possible to reduce costs, keep downtime to a minimum and respond early to possible malfunctions or safety risks.
This of course requires that machines are equipped with instrumentation sensors that collect data and distribute them either locally or via the cloud. Typical sensorial data include temperature, power, torque, pressure, acceleration/vibration signals and more.
This transition is not cost free, but rather require investments on multiple fonts: from provisioning infrastructure material (HW/SW) to training machine operators to perform investigative work based on the new information provided by the CM systems. However, the benefits from adopting condition monitoring systems are multiple:
Planning and implementing predictive maintenance schedules
Reducing maintenance costs through effective maintenance planning
Maximizing production output through outage prevention
Reducing downtime by prolonging equipment life
The traditional way condition monitoring is used, is to observe the monitoring signals and impose hard-coded alarm limits that trigger further maintenance actions. However, such procedure is effective only in simple tasks and requires expertise and good knowledge on the process itself. In complex processes when there are hundreds of sensor measurements, which is often the case in practical applications, it is impossible to assess the health of an asset just by analyzing each measurement on its own.
AI-assisted anomaly detection
This is where AI and deep learning techniques come into play. Unlike a human operator, AI algorithms have no problems analyzing historical datasets for hundreds of sensors over a period of several years. Recent advances in deep learning have shown great performance in modelling complex tasks in various domains.
The basic idea is to utilize an autoencoder network to compress the multidimensional sensor data to a lower dimension representation, which captures the rich interactions between the various signals. It has been observed that as the monitoring machine degrades, the interaction between these measurements is changing which also makes the autoencoder reconstruction error to increase. This can be used as an indicator of machine health which allows us to move from the simple identification of a malfunction to the proactive correction of the underlying problem.
Despite the active research in the field there are still some open challenges related to the low anomaly detection recall rate, the anomaly detection in high-dimensional and/or not-independent data, the noise-resilient anomaly detection and the explainability mechanisms for predictive maintenance (PdM).
