Predictive maintenance (PdM) is unarguably the most cost-effective maintenance strategy available to plant operators today. It can bring a 6-8% uplift in the overall production, and can reduce the costs of maintenance by 18-25% depending on the industry in question. However, not all PdM implementations are able to achieve these results.
While numerous factors can influence the impact of a PdM solution on maintenance KPIs and the achieved cost savings, the technology that sits at the heart of PdM is usually the foremost determinant of its success. This technology is machine learning (ML). Each PdM solution exploits ML to achieve different end-goals, and the application scenario can also limit what can be achieved with PdM. Understand how ML drives PdM systems, observe how various aspects of ML engineering influence PdM solution design, and finally – how to evaluate the efficacy of an ML-based PdM implementation.
How machine learning drives predictive maintenance
Machine learning can be said to function as the brain of a predictive maintenance system. When it is supplied with data about the current state of an asset, it is able to infer the ‘health’ of the asset, it can quickly identify when its performance starts to deteriorate and can predict when the asset is likely to fail.
To make this possible, an ML model is created, which is trained on historical data recorded from a machine. Training refers to the process where the model learns what the optimal state and oncoming failure of a machine looks like in terms of the data that is supplied to it. Based on these learnings, the ML model is able to provide forecasts on when a machine is likely to fail using the input data.
Machine learning techniques are typically able to ingest tens, or even hundreds of input variables about the current state of the machine to forecast its future state. That’s why, the insights obtained for predictive maintenance with machine learning cannot be uncovered, even by the most experienced plant operators.
Machine learning approaches for PdM solution design
A number of machine learning approaches are available for use in PdM. The selection of these approaches depends on the desired end goal, and the constraints posed by the application scenario.
Broadly, there are two key approaches to predictive maintenance with machine learning. These are:
- Supervised learning: In this case, the training dataset (historical data recorded from machine operation) contains instances of occurrence of failure.
- Unsupervised learning: In this case, the data streams from machine operation are available, but the dataset doesn’t contain occurrence of failure.
Supervised vs unsupervised learning: when to use which?
The selection of supervised over unsupervised learning depends on the availability of the training data. In plants where run-to-failure maintenance is conducted, the training dataset will contain occurrence of failure – and as a result, supervised techniques can be used. However, when a preventive maintenance strategy is being used, the training data is unlikely to have instances of machine failure, because a repair was carried out well in advance. This will necessitate the use of unsupervised techniques.
Apart from these factors, the objective of the PdM system also dictates the selection of the ML technique. For instance, a system which predicts when a piece of equipment is likely to fail, will use a different ML approach than one that predicts what will cause an oncoming failure. Here are three key methods that are used in the design of a PdM solution.
Examples for Supervised learning:
Binary classification
Binary classification is essentially used to distinguish between healthy and faulty machine behavior. It estimates the probability at which a machine may fail over a given time period. This time period is usually the minimum possible downtime of a machine. Some of the common algorithms used for binary classification include Naive Bayes, k-Nearest Neighbours, and Decision Trees.
Regression modelling
This approach is leveraged when a solution needs to predict the remaining useful life (RUL) of a given asset. In such applications the dataset contains machine state observations which correspond to a specific period of time. RUL is typically formulated as a continuous number, and is updated based on incoming input data over time. Some common regression modelling algorithms include linear regression, Lasso regression, ridge regression, etc.
Multi-class classification
Binary classification has limited impact on key maintenance decisions – for example, what is the likely cause of an upcoming failure, or which parts may need to be replaced in the upcoming maintenance cycle. To mitigate these shortcomings, multi-classification techniques are used. These techniques enable the prediction of specific failure mode associated with a machine, along with the likely time-to-failure. This enables plant operators to schedule targeted maintenance activities, while speeding technician workflows with contextual information.
Multi-label classification
Multi-label classification involves assigning various labels to plant operation and maintenance scenarios, such as labelling with “normal operation,” “routine maintenance,” and “critical issue.” In contrast, single-label classification assigns only one label, multi class says – temperature high or vibration high (only one of them) whereas multi label says – both temperature high and vibration high
Examples for unsupervised learning:
Clustering
Clustering might group similar vibration patterns, aiding in predicting machinery health. By identifying these natural clusters, plant managers can make informed decisions about maintenance strategies and resource allocation, enhancing overall operational efficiency and reducing downtime.
Auto encoder
It encodes sensor data from machinery states, like temperature, into a compressed format. Subsequently, it reconstructs the original data, aiming to minimize reconstruction error. This process captures essential features while filtering noise. Autoencoders aid in anomaly detection by highlighting deviations between reconstructed and actual data. For instance, if a motor’s encoded representation deviates, it might indicate a fault. In plant management, autoencoders enhance predictive maintenance by learning normal patterns, enabling early anomaly identification, and optimizing equipment performance.
Understanding model effectiveness in PdM implementations
Predictive maintenance with machine learning generally enables accurate predictions with 90%+ accuracy. However, the performance of the model is evaluated with the following four parameters, and it is important to know what they represent, to judge their effectiveness.
- Accuracy: This is the ratio of the sum of true positives and true negatives (TP and TN) and the sum of true and false positives and negatives (TP, TN, FP, FN).
- Precision: Precision measures the ratio of correct true positive predictions and true and false positive (TP and FP) observations.
- Recall: Recall is the ratio between true positives (TP) and false negatives (TP and FN).
- F1 score: Finally, F1 score is a weighted average of precision and recall. It is a more comprehensive metric for evaluating ML models.
Depending on the chosen metric, the performance of the model in a given scenario may vary. For example, in a PdM study, the accuracy of a prediction of a decision tree was 98%, whereas its precision was 71%. In such scenarios, F1 scores represent a better way of evaluating model performance.
Next steps
Machine learning is a crucial technology in PdM solutions. With numerous ML approaches available for use, the selection of the right technique depends both on the constraints posed by the application scenario, and by the objective of a PdM implementation. Finally, knowing how to evaluate a model’s performance is crucial to judge the efficacy of a PdM solution in real scenarios.
Power your predictive maintenance strategy with UptimeAI, an industry-leading PdM solution featuring advanced ML capabilities. Contact now to get started.