Meet the Predictive Maintenance Squad
1. Who we are
The Predictive Maintenance squad (PdM) is focused on detecting early stages of degradation on the process and industrial plants providing a window of opportunity to act before a failure occurs. Therefore, our main goal is to prevent downtime and unnecessary maintenance cost. Having that in mind, we have developed solutions for equipment and processes by using different machine learning techniques.
We have a team with expertise in both data science and engineering. For developing our tools, we interact closely with our final users to build the most suitable approach for the issue we want to avoid.
2. What do we do?
2.1. Context
Predictive Maintenance is the most recent type of management program. Considering how this area has evolved over the years, there are three main types: corrective, preventive, and predictive maintenance.
The first is the most reactive one, since it works in a run-to-failure logic. In other words. the maintenance is only performed when a machine breaks down. Although it feels reasonable at first, not performing any maintenance before a system of the machine fails to operate tends to be the most expensive technique used.
Preventive maintenance, on the other hand, has the objective of anticipating maintenance, as the name says. It works based on a time schedule, i.e., based on the hours of operations, maintenance activities are defined. Therefore, there is an assumption that machines and systems degrade according to a typical time frame. This can be a disadvantage since it does not take into consideration the specificities of the equipment and process conditions, which can lead to either unnecessary maintenance activities or unexpected failures.
This problem was the motivation for the predictive maintenance development. This approach nowadays can be understood in a variety of definitions. However, the premise is that by monitoring the specificities of operation from each equipment, it is possible to provide the optimal time interval between maintenance activities. [1]
a. Types of Predictive Maintenance
Predictive Maintenance is widely used as a generic term to describe very different maturity stages of analytics applied to asset reliability. In Table 1 we can see different approaches inside its scope. It is important to mention the paradigm shift from PdM 2.0 to PdM 3.0, as data-driven techniques increase its feasibility.
b. The impact of Industry 4.0 on PdM
One of the biggest changes from PdM 2.0 to 3.0 and 4.0 is that computer science and artificial intelligence has become increasingly present complementing solutions in which engineering was the dominant expertise[2]. This is due to the connectivity, amount of data, and new devices that we are experiencing in this so-called Industry 4.0 that has enabled autonomous decision-making processes, monitor assets and processes in real-time.[3]
To achieve an effective PdM monitoring system, enough data from all parts of the industrial process is crucial. Nowadays, this is possible due to the IoT technology that enables the communication between machines, software solutions, and cloud technology. Condition-monitoring sensors are installed on the equipment and send real-time data to predictive models that deliver failure predictions to diminish maintenance costs and downtime, and improve productivity and quality as well[4].
Our PdM squad works on the PdM 4.0, in which we unite the latest data science techniques with engineering knowledge in a structured monitoring center. Therefore, PdM aims not only for a fit-for-purpose solution but actually for a cultural shift on how to use data to bring actionable insights driven to impact.
3. How we work?
There are plenty of strategies to capture different types of failure modes — the best approach for your problem will depend on what question you want to answer. So, below we have listed some approaches:
● Survival models to estimate risk probability and time before failure
● Classification models to predict failure within a given time window
● Flagging anomalous behaviour models
Survival Models
Survival analysis focuses on estimating time to events, so in other words, we can estimate the expected duration of time until one event happens. In our case, the main goal is to find out “how many days are left before the equipment fails?” and in predictive maintenance this kind of model is well known as Remaining Useful Life (RUL).
It requires labeled data as a health index that records the current condition of the equipment and identifies the failure. So, the model will learn the relationship between features and the condition of the equipment over time.[5] So the model will fit the data and generate a Survival Probability Curve as shown below:
Keplan-Meier
The survival curve (also called survival probability function) can be estimated using Kaplan-Meier (KM) estimation technique, as follow:
In which:
dtFi: it’s the number of equipment that failed at time tFi
YtFi: it’s the number of equipment which are at risk at time tFi .
By “at risk”, we mean the number of equipment that did not fail before tFi .[6]
Which will give us an estimated curve similar to the previously shown.
Classification models to predict failure within a given time window
On the other hand, we may want to find out if the equipment will fail in the next N days, and in that case, we train a classification model to predict it. The idea is the same as in RUL, but instead of regression, we have a classification problem that also requires labeled data.
On the paper “A systematic literature review of machine learning methods applied to predictive maintenance” [7] there is a table which contains an overview of the most recent papers for PdM (between 2009 and 2018), where each line is related to a paper and there’s a description about what kind of model was used, including Random Forest, SVM and Neural Networks models to solve classification problems.
Flagging anomalous behaviour
Depending on the context, you may not be able to obtain sufficient labeled data classified as failure or, even worse, there won’t be any failure tag in your dataset. So, in these cases modeling the normal behavior of your system is a good way to overcome this type of lack of data and when the model detects some anomalous behavior it will flag as a failure.
One downside of this approach it’s that anomalous behavior doesn’t always mean failure on the system and even if the model is correctly labeling failures, it will not give you any information in advance about the failure.
The evaluation of this type of model is also challenging due to the lack of labeled data. But, in this case, you can ask for experts to provide feedback on data that your model tagged as anomalous to understand if your model is making sense or not.
Written by Thais Stilck and Rodrigo Hamacher
Let’s keep in touch!
Wrapping up all we have discussed, soon we’ll explore in detail some predictive maintenance techniques — in the meantime, check our page on LinkedIn.
Stay tuned! See ya!
References
[1] An Introduction to Predictive Maintenance, MOBLEY, Keith, 2002;
[2] Predictive maintenance in the Industry 4.0: A systematic literature review, Z., Tiago, C., Cristiano, R., Rodrigo, 2020;
[3] Industry 4.0: the fourth industrial revolution — guide to Industrie 4.0”, https://www.i-scoop.eu/industry-4-0/#:~:text=Industry%204.0%20is%20the%20current,called%20a%20%E2%80%9Csmart%20factory%E2%80%9D
[4] “A Complete Guide To Predictive and Predictive Maintenance”, LEVITT, Joel, 2010.
[5] Kang, Z.; Catal, C.,Tekinerdogan, B. Remaining Useful Life (RUL) Prediction of Equipment in Production Lines Using Artificial Neural Networks, 2021
[6] Remaining useful life prediction using prognostic methodology based on logical analysis of data and Kaplan–Meier estimation, RAGAB, Ahmed, YACOUT, Soumaya, 2014
[7] A systematic literature review of machine learning methods applied to predictive maintenance, CARVALHO, Thyago, SOARES, Fabrízzio, 2019.
