Aku Karhinen: Improving the Generalisability of Deep Learning Models in Vibration-Based Condition Monitoring

Improving the Generalisability of Deep Learning Models in Vibration-Based Condition Monitoring

By Aku Karhinen

Overview

Rotating machines are a fundamental part of the modern world: they enable power transmission in transport systems, serve as key drivers in industrial processes, and generate a large share of global electricity.

The operation of these machines relies on components designed to wear over time, such as bearings and gears, whose condition is typically monitored using vibration measurements. Based on vibration signals, the condition of a machine can be assessed and maintenance actions scheduled proactively and cost-effectively.

Unexpected failures, however, can lead to significant economic losses. For example, condition monitoring of large industrial internal combustion engines is critical for ensuring plant availability, safety, and energy efficiency.

Challenge

Although the automation of vibration analysis using machine learning has been studied extensively, practical applications are often limited by the high cost of measurement equipment and the scarcity of real fault data.

Research & Innovation

In our article, “Data-driven torsional vibration-based fault diagnosis of large internal combustion engines without real fault data” (Engineering Applications of Artificial Intelligence), we present a method developed with funding from the Walter Ahlström Foundation.

The approach enables cylinder-specific fault diagnosis in a 20-cylinder industrial engine using data from a single measurement point, reducing the required number of sensors from dozens to just one.

A central challenge is shared across many industrial deep learning applications: reliable, labelled fault data is not available in practice, and faults cannot be safely or cost-effectively generated for measurement purposes.

Our solution is based on simulation-driven training and bridging the gap between simulation and real-world data by combining various transfer learning techniques.

Results

The results demonstrate that, with carefully selected alignment and regularisation methods, it is possible to achieve very high fault detection accuracy while maintaining strong classification performance under limited real-world validation.

This makes the approach viable for industrial environments.

Impact

The research presented in this publication is the result of several years of collaboration with Wärtsilä.

The starting point was a problem long considered nearly unsolvable: how to build reliable fault diagnostics without real fault data. The grant from the Walter Ahlström Foundation made it possible to address this challenge in a sustained manner, explore multiple approaches, and develop a solution that combines simulation with limited real-world measurements—evolving from concept to a fully functional method.