AI in Predictive Maintenance Software: How It Works

In both digital services and manufacturing, the modest profitability of the average delivery pipeline makes downtime expensive. In the energy industry, operation and maintenance costs for offshore wind turbines eat up 20-35% of all revenue for generated electricity¹, while in the oil and gas sector they represent somewhere between 15-70% of total production costs².

Therefore it's no surprise that these industries, as well as others, are keen to adopt more intelligent and informed maintenance approaches through predictive maintenance software and intelligent analytics with the help of AI developers.

In this article, we'll take a look at some new developments related to this growing field, as well as some of the challenges in taking maintenance management out of the steam era and into the industry 4.0 age.

On the factory floor, a bearing assembly typically comprises a ring that houses spherical metal balls designed to help machine components or speed them on their way to the next stage in manufacturing.

Since the general principles of bearing assemblies apply also to vehicles and other types of machine than factory installations, bearing analysis is a pivotal area of research in AI-based preventative maintenance.

Prior to recent advances in GPU-driven machine learning, a wide range of 'classical' machine learning solutions were trained on bearing assembly PdM over the years. These include artificial neural networks (ANNs)⁴, support vector machines (SVMs)⁵, principle component analysis (PCA)⁶ and k-NN classification⁷, among others.

These older methods depend on feature engineering⁸, where the shape of the equipment and rotation/movement rate is manually calculated against other user-specified input characteristics prior to machine analysis. This approach is quite brittle and often cannot be re-used or adapted to similar cases, making it labor-intensive and costly in itself.

Newer deep learning techniques are able to observe a greater range of contributing factors that are hard to include in a feature engineering model. These include:

• The influence of heat⁹
• Variations in lubrication quality and flow¹⁰
• External vibrations, where multiple vibration sources can reach damaging frequencies¹¹ not anticipated in either individual source

Deep learning approaches incorporate automated feature engineering, provide transferable solutions, and, according to Google's most prominent contributing AI architect, Andrew Ng, offer an exponential improvement in failure prediction as the input data volume rises:

The first notable deep learning application for bearing PdM came from Ghent University in 2016¹² in the form of a convolutional neural network (CNN) with automated feature extraction and vibration analysis.

Neural network-based approaches to this challenge are also able to calculate 'imponderables' such as degradation of lubricating oil and other contributing factors that are difficult to model in traditional feature engineering.

As a further aid to research in this sub-field, dedicated data resources for bearing-based PdM have come up online in recent years, including the Case Western Reserve datasets¹³, the Mendeley vibration datasets¹⁴, and the bearing dataset in NASA's PCoE collection¹⁵.

One 2020 paper²⁴ out of the Politecnico di Torino in Italy proposes a method of data mining to enable a regression-based PdM system that can predict vehicle faults across a range of possible data availability: a) vehicle-specific data, the ideal scenario; b) analogous data from similar types of vehicle; and c), a generalized vehicle model, which can be used when the vehicle is new, and no model-specific data yet exists.

Another Italian-led research initiative²⁵ offers a 'generalized' machine learning model for fault detection in centrifugal pumps for the oil and gas industries with approaches using support vector machines (SVMs) and multilayer perceptron (MLP).

The University of Bologna leads a recent paper²⁶ offering a fuzzy-rule-based approach to predictive maintenance for the Tier-1 data center supporting the Large Hadron Collider (LHC) at Geneva. The generalized methodology used by the study's Gaussian fuzzy classifier (eGFC) offers potential to transfer across to other similar data center and computing systems.

A Spanish-led paper²⁷ has proposed a real-time monitoring system for wind turbines wherein high volumes of historical data are fed into an always-on fault-tolerant network, with dashboard access.

Michigan State University recently published details of a methodology designed to develop 'universal diagnostics' across different types of vehicle by leveraging in-vehicle IoT sensors, as well as the media capabilities and GPU power of modern smartphones²⁸.

In many cases, the costs of establishing a PdM framework can be lowered by using cloud-based services and open-source frameworks, typically applied in enterprise-grade AI. The downside is that the attack surface for the system will be greatly increased — not a minor consideration when assaults on industrial frameworks such as Supervisory Control and Data Acquisition (SCADA) have grown in 2020³³.

Outright Denial of Service (DoS) attacks are damaging but usually lead to quick resolution and long-term countermeasures afterwards. A more effective type of incursion against PdM systems is a False Data Injection (FDI) attack³⁴.

In an FDI, the mechanisms of the targeted framework are used against themselves, either to cause unnecessary expense and downtime through premature maintenance requests or to permit equipment to fail beyond the threshold where the PdM system would normally have intervened. In critical cases, such an attack can threaten lives³⁵.

Typically, sensors are the weak link in the chain³⁶, since AI-based PdM systems are likely to use IoT-based information feeds from affordable or specialized sensors. The evolving state of long-range and near-signal network connectivity systems increases the opportunity to explore vulnerabilities in a growing number of network protocols:

Research out of the University of Missouri in 2019³⁷ tested the efficacy of three deep learning techniques to predict Remaining Useful Life (RUL) and to protect against FDI attacks: long short-term memory (LSTM); gated recurrent unit (GRU); and a convolutional neural network (CNN).

The project tested its system against a model of a turbofan engine in NASA C-MAPSS Aircraft Engine Simulator Data³⁸. GRU is a gating mechanism in a recurrent neural network that's capable of more refined training and higher specificity than an LTSM unit.

Other studies confirm^39,40 that a GRU-based approach seems to be the way forward in defending industrial systems against FDI attacks, and generally requires less training time. However, GRU sacrifices a certain degree of accuracy to LTSM in favor of greater resilience against FDI incursions.

In this sector, as elsewhere, sustainability is an intensely political topic^45,46. The growth of predictive maintenance is a potential threat to suppliers that have become financially dependent on after-market demand for parts and repairs⁴⁷ and that have already re-evaluated their core economic models, at some expense, to accommodate the 'disposable' culture that emerged in the late 20th century⁴⁸.

Though this incumbent burden is a challenge to the evolution of PdM, an ongoing shift in political and public thinking towards more ecological and economical manufacturing processes offers hope that the suppliers will increasingly adopt — and eventually champion — AI-based PdM.

In the meantime, the challenge remains to distil essential principles of industrial damage data into models that are transferable across systems and across a broader range of equipment, until the sector reaches a better level of commoditization, generalization and accessibility.