Implementation of Predictive Maintenance

In today's production systems, machine breakdowns are very severe. They not only generate unscheduled downtimes with related expenses and dissatisfied customers, but they also wreak havoc with staff scheduling (additional shifts) and warehousing (additional storage), as well as quality deterioration (scrapped components) and subsequent machine warm-up periods.

Most programs for maintenance management have been outlined by a variety of authors. According to Bateman [1], there are three basic types of maintenance programs, including reactive, preventive, and predictive maintenance.

Many organisations have traditionally used a reactive maintenance policy, simply fixing machinery when they break down.Improved technology and the skill of maintenance employees have caused several organisations to abandon this reactive strategy in recent years. Improved technology and the skill of maintenance employees have lately forced several organisations to abandon this reactive strategy and opt for a more proactive strategy.

Specified by Bateman, one such proactive strategy is preventive maintenance which is currently being used in the industrial sector to avoid unexpected machine outages and the subsequent corrective repair. In this case, a machine is scheduled for maintenance, usually from its manufacturer, by predetermined calendar days and/or machine usage. Although this method helps to decrease unplanned downtime, it remains inefficient since the equipment is blindly maintained independent of its real condition, increasing production expenses.

Even though with this strategy, maintenance expenses are reduced by the price of a machine breakdown, maintenance nonetheless incurs large expenditures such as additional staff, wasted remaining useful life, and tooling. Another form of proactive strategy is the so-called predictive maintenance. The difference between preventive maintenance and predictive maintenance lies in how the data is being analysed. Preventive maintenance focuses on monitoring and analysing data from the actual, present condition of the equipment in operation, whereas predictive maintenance depends on historical data which is being analysed and sophisticated algorithms to predict the remaining useful life, thus forecasting when maintenance actions are expected to be necessary.

The three basic maintenance strategies

This trade-off of each strategy can be summarised with the following qualitative diagram [3]:

zhaw 2

According to the illustration above, it would be better to apply predictive maintenance based on the asset's actual and historical sensory data, emphasising that the total maintenance cost curve is kept to a minimum. This observation isn't new, and it comes as no surprise that predictive maintenance is a growing field of study that spans various disciplines, including statistics, physical-based modelling, machine learning, and others. The recent increase in the number of techniques and tools is due to technological advancements and the diversity of purposes (costs, reliability, risk, safety, and feasibility).

Knowledge-based systems, classical machine learning systems, and systems based on deep learning technologies were defined as three key groups in a recent taxonomy of practical solutions [3]. This investigation revealed the need for more advances, including:

  1. The standardisation of new technology
  2. The ability to work with minimal data
  3. The architecture of hybrid systems
  4. Comprehensive methods that go beyond the level of single components

There are several intriguing scenarios from a scientific standpoint, but they lack a systematic method [4]. On the other hand, a purely data-driven approach with general-purpose ML solutions could deliver promising insight into the remaining useful life of an asset. Data-driven solutions allow a reduction in sensors and in addition, can handle machine diversity well. However, Data-driven systems are frequently based on ambagious, uninterpretable models (e.g., neural networks), which are incompatible with Current Good Manufacturing Practices (CGMPs), which mandate that all functions are affecting a product a predetermined and repeatable behaviour.

In addition, data-driven approaches highly depend on very large datasets in which the data itself suffers from the so-called "class imbalance" [5] caused by the nature of manufacturing because the failure of events from an asset seldomly occurs since it is usually being maintained before the failure completely occurs. Consequently, it might take years for adequate data to become available for training accurate and reliable models [6].

On the other hand, models based on physical models [7] offer a promising alternative. However, these need professional machine operating expertise and therefore scale poorly with the number of machines or the complexity of the systems [8]. Furthermore, developing a thorough model for a standalone station takes a few days and requires dealing with several mathematical assumptions as well as the adjustment of several physical characteristics. Physical models that have been constructed are fixed in stone, and as a result, they cannot adapt to changes.

The following table illustrates the pros and cons of each approach:

Approach for PdM



Data-Driven (Machine Learning)

Little domain knowledge


Lots of data required

(effort, time, infrastructure)

Physical Modelling


Little data required

Sophisticated domain knowledge required


To attain the best of both worlds, hybrid techniques could be utilised. Hybrid techniques are based on data-driven solutions, which are enhanced by physical models, enabling to achieve significant increases in accuracy with sparse training data. Such strategies have been developed effectively in research and have proven to be useful in practice [9].

When considering that almost every industrial facility is equipped with a large variety of machines supplied by different machine manufacturers, the implementation of predictive maintenance isn't a trivial task.

One possible implementation process for predictive maintenance would be to pick the simplest and most critical asset as a pilot.


If the chosen asset is a legacy machine without any sensors, it needs to be retrofit with sen-sors to record and store data.

zhaw 4

In a second step, a model should be designed and trained based on the collected data. Lastly, the model should be deployed and provide forecasts for the remaining useful life of the given asset. These steps can then be applied again to the next compliant asset.


The current state-of-the-art in predictive maintenance implies that the machinery is equipped with several sensors. Physical-based models and/or data-driven models are used to anticipate the machine's remaining useful life based on sensor data. While these methods are highly accurate, they come with at least the following scalability issues for practitioners:

  • Sensors can malfunction. The more sensors you have, the more likely you are to have a failure. As a result, we are forced to install "sensors monitoring sensors" indefinitely.
  • Due to additional validation, installation work, and infrastructure needs like bandwidth, mounting many sensors can be costly.
  • It's difficult to create complex physical models. A sophisticated model requires a significant amount of physical labour and highly specialised knowledge.
  • The sheer number of modelling approaches and sensors available is overwhelming. Practitioners are commonly puzzled, especially if they lack a thorough understanding of modelling and machine learning.
  • For data-driven model's large datasets are required, yet historical data is scarce or non-existent.
  • Gathering large datasets in most cases result in impractically lengthy recordings and a heavy burden on the infrastructure.

In many practical situations, however, such precision is not required. This is especially true for production sites without predictive maintenance, where the expense or complexity of the installation exceeds the necessity for precise forecasting. Therefore a minimal setup could already yield valuable insight.

[1] J. Bateman, Preventive maintenance: Standalone manufacturing compared with cellular manufacturing, Industrial Management 37 (1) (1995) 19}21.

[2] Swanson, Laura. "Linking maintenance strategies to performance." International journal of production economics 70.3 (2001): 237-244.

[3] "A Survey of Predictive Maintenance: Systems, Purposes and Approaches", Y. Ran et Al., arXiv:1912.07383, 12 Dec 2019.

[4] "A System Engineering Approach to Predictive Maintenance Systems: from needs and desires to logical architecture", J. J. Montero Jiménez and R. Vingerhoeds, IEEE-ISSE, 2019.

[5] "A Data-Analytics Approach for Enterprise Resilience", D. Xu et Al., IEEE Intelligent Systems, vol. 34, no. 3, pp. 6-18, 1 May-June 2019.

[6] "A systematic literature review of machine learning methods applied to predictive maintenance", T. P. Carvalho et Al., Computers & Industrial Engineering 137, 2019.

[7] "Predictive Maintenance in Dynamic Systems", Edwin Lughofer and Moamar Sayed-Mouchaweh, Springer International Publishing, 2019.

[8] "Physical Model-Based Prognostics and Health Monitoring to Enable Predictive Maintenance", T. Tinga, R. Loendersloot, In Predictive Maintenance in Dynamic Systems. Springer, Cham, 2019.

[9] "A new hybrid prognostic methodology.", Eker O, Camps F, Jennions IK., International Journal of Prognostics and Health Management", Volume 10, March 2019.

Partner Name:
zhaw logo