Author(s): Luis Ramirez Originally published on Towards AI. For industrial equipment the default approach is preventive maintenance, which involves servicing equipment on a fixed schedule, such as monthly or semi-annually. While better than reactive maintenance (fixing equipment after failure), this one-size-fits-all strategy has significant drawbacks. Equipment units experience different operational conditions, loads, and degradation rates, yet receive identical maintenance schedules. This leads to two suboptimal outcomes: either overspending on unnecessary maintenance (including operational downtime costs) or not maintaining equipment frequently enough, risking costly failures. A better solution is an intelligent maintenance strategy tailored to each piece of equipment based on its actual condition and predicted degradation; this is precisely where predictive maintenance comes into play. Starting simple Predictive maintenance often requires complex machine learning models that can be difficult to implement and interpret. In this post, we take a different approach, focusing on a visual analysis built on top of PCA projections. By focusing on visualization first, we can: Build intuitive understanding of degradation patterns. Establish a baseline for more advanced models. Create a common language for maintenance teams. Identify different failure patterns visually. Achieve quick wins while developing more sophisticated solutions. Source: Image by the author. The Dataset For this analysis, we’ll use the NASA turbofan engine dataset, which contains run-to-failure data of 100 engines. Each engine starts in good condition and develops a fault over time until failure. The dataset includes: Multiple sensor readings (temperatures, pressures, rotation speeds Operating condition indicators The cycle number Besides the initial features, I have added some aggregated data of mean tendency metrics, minimums, maximums and standard deviations that were generated using a length-5 rolling window. This preprocessing helps capture temporal patterns that might indicate degradation. Source: Image by the author. Step 1: Basic Health Visualization Understanding PCA Projections Our first goal is to create an intuitive visualization of equipment health status. We’ll use Principal Component Analysis (PCA) to reduce the many sensor dimensions into a two-dimensional plot that can be easily interpreted. PCA works by: 1. Finding the directions (components) of maximum variance in the data. 2. Projecting the data onto these components.3. Ordering components by how much variance they explain.Source: Image by the author. For this analysis we will only use the first two components, the result is a two-dimensional plot where similar operating conditions cluster together, besides the two main components we will use a gradient to represent the Remaining Useful Life (RUL). We can already see some patterns emerging in this visualization. As the data points move from left to right along PC1, they show increasing engine degradation. However, the linear scale makes it difficult to distinguish between different stages of low RUL values, which are the most critical for maintenance decisions. Source: Image by the author. Let’s apply a log scale to the RUL gradient to highlight these important end-of-life patterns. Now we’re seeing clearer separation between health states. In this enhanced visualization, we can distinguish: Healthy equipment clustered in the center left Moderately degraded equipment in the middle Severely degraded equipment spread in the right side Source: Image by the author. How To Use This Visualization? This basic visualization already provides valuable insights for maintenance teams: 1. Current Health Assessment: By plotting new data points from an operating machine, we can immediately see where it falls in the health spectrum 2. Early Warning System: As a machine’s position shifts toward the degraded regions, we receive an early warning before failure occurs 3. Fleet Comparison: Multiple machines can be plotted simultaneously, allowing us to prioritize maintenance efforts on the most degraded equipment We could even draw rough decision boundaries to separate the sections: Source: Image by the author. While this visual separation into low, medium, and high-risk zones roughly represents RUL > 100, 100 > RUL > 40, and RUL < 40, respectively. Now we have a better grasp of what this PCA represents, but to make decisions based on this, we need to polish our approach. Step 2: Defining In-Sample Regions with Gaussian Mixture Models Identify out of sample regions The previous classification has an important limitation: it extends the classification space infinitely in all directions. This means a point with extreme values, e.g., (-20, 25) or (-20, -15), would still be labeled as “low risk”, or a point like (5, -15) would be labeled as “medium risk”, simply because it falls on that side of the decision boundary. In reality, such extreme values likely represent sensor malfunctions or completely novel operating conditions outside our training data; they can be closer to a failure than to normal operation. Source: Image by the author. To address this limitation, one powerful enhancement is implementing Gaussian Mixture Models (GMMs) to define out-of-sample regions in the PCA space. GMMs model the actual distribution of “normal” data, allowing us to identify when new observations fall outside the regions of known behavior: Source: Image by the author. By using a GMM with a single component (which results in an ellipse), we establish a basic boundary. Although this offers an initial approximation, we can notice that the elliptical shape is not the best match for the distribution of data we have. Source: Image by the author. To improve the quality of the region definition, we can use a GMM with multiple components. This allows us to model the complex, non-elliptical distribution of machine states. This approach allows us to identify unusual degradation patterns that fall outside normal behavior. Source: Image by the author. Step 3: Feature Selection for Better Results Using Mutual Information The PCA algorithm is a deterministic process that maximizes the explained variance kept from the original data in the projected components, but not all sensor variance correlates with equipment health. For example, some sensors might fluctuate due to ambient conditions rather than degradation. To address this, we’ll use mutual information to identify which features most strongly correlate with Remaining Useful Life: Source: Image by the author. Source: Image by the author. This approach allows us to focus on the most relevant indicators of degradation, simplifying the analysis while keeping the number of variables manageable. Improved Visualizations Now, if we repeat the PCA plot using only the selected features, we can see that the shape of the projection has changed: Source: Image by the author. However, the core behavior remains, with healthy machines (datapoints) on the left and degraded ones on the right. Let’s apply the log scale again to better visualize the end-of-life period. Now we can see a clearer picture, there is a better distinction between the different RUL values and we have a delimiting curve/shape for the out of sample data. Source: Image by the author. Step 4: Risk Classification SVM in action After applying feature selection and creating an improved PCA projection, we can implement the classification of the risk regions, from our previous hand draw borders to a more robust method. For that we can use Support Vector Machines with a linear kernel, which will find the better line to separate the region. As the SVM are a supervised method, we need to define some labels to depict the level of risk. For the purposes of this analysis and based on the results we have got with the PCA, we will choose as thresholds: Source: Image by the author. With the definition we can now train our SVM classifier Source: Image by the author. This classification allows us to divide the PCA space into clear regions representing different risk levels. The color of the region represents the SVM classification while the coloring of the dots represents the actual category, based on the labels we have defined. Source: Image by the author. Why use linear SVM and not Neural Networks or boosting trees? I chose to keep the analysis simple yet well-founded, and linear SVMs offer a good trade-off by providing robust yet easily interpretable results. They essentially draw the optimal straight lines that separate the classes — similar to my hand-drawn approach earlier, but with mathematical precision. Now, let’s change the plot to focus on analyzing classification errors. This approach can further refine our understanding. The plot shows that 11.6% of the data points have been classified incorrectly. Among these errors, the most common misclassification occurs between low and medium risk, with errors between medium and high risk being the second most frequent. Fortunately, in these training results, there are no errors between low and high risk categories, which would represent the worst case scenario. Source: Image by the author. Actionable Maintenance Decision Points This classification approach provides clear, actionable information: 1. Low Risk: — Action: Continue normal operation — Monitoring: Routine checks during scheduled maintenance — Frequency: According to standard maintenance schedule2. Medium Risk: — Action: Plan intervention during next scheduled maintenance — Monitoring: Increase monitoring frequency — Preparation: Order potential replacement parts — Documentation: Begin documentation for upcoming maintenance3. High Risk: — Action: Schedule immediate maintenance intervention — Monitoring: Continuous or daily monitoring — Preparation: Ensure all replacement parts are available — Planning: Coordinate with production to minimize impact4. Outside In-Sample Region (Dashed Line): — Action: Investigate unusual behavior — Indication: Potential novel failure mode or sensor issue — Response: Engineering analysis requiredStep 5: Trajectory Analysis Individual Equipment Tracking Over Time Perhaps the most insightful visualization is the degradation trajectory plot, showing how individual machines move through the PCA space over time: This visualization reveals: Different paths to failure, suggesting distinct failure modes. Non-linear degradation patterns. Acceleration of the degradation after the crossing the medium range area. Source: Image by the author. Velocity Field Visualization Building on trajectory analysis, we can create a velocity field visualization showing the average direction and speed of degradation throughout the PCA space: The arrows in this visualization show both the direction and rate of degradation in different regions: – Brighter arrows indicate areas where degradation occurs more rapidly– The overall flow pattern reveals typical degradation paths– Areas with divergent arrows may indicate different failure modesSource: Image by the author. Predicting Future Degradation Paths For maintenance teams, trajectory and velocity field analysis provides great insights: 1. Degradation Rate Estimation: By tracking a machine’s movement through the PCA space over time, teams can estimate how quickly it’s degrading 2. Time-to-Failure Prediction: The velocity field helps predict how long before a machine enters the high-risk region 3. Maintenance Planning: Understanding typical degradation paths allows teams to plan interventions at optimal points 4. Failure Mode Identification: Different trajectory patterns often correspond to different failure modes, helping teams prepare the right replacement parts and procedures For example, if a turbofan engine follows a trajectory similar to Equipment 3 in the visualization, maintenance teams know to focus on specific components that typically fail in that pattern. If it follows a different path, they’d prepare for a different type of intervention. Conclusion In this blog, we have gone through the core steps to create a robust PCA analysis for Predictive Maintenance that can be replicated for other datasets to: Create intuitive representations of equipment health Identify different degradation patterns visually Establish a common language for discussing equipment state Build a foundation for more sophisticated modeling It’s important to keep in mind that PCA is a starting point, not a substitute for more complex models. A natural next step from here is to dive deeper into the dataset, understand feature behavior and interpretation, and build more complex models either for classification or as a forecast having RUL as the target. The second point to consider is that the data used here is exceptionally clean, with 100 run-to-failure assets. In a real-case scenario, if we are starting from scratch, the first big challenge will be to build a useful dataset. More often than not, data from industrial applications has missing values, mislabeled events, noisy sensor data, and few actual failures identified to train on. Understanding the nuances of that specific use case and industry will be crucial for success. Here are some options to continue exploring Aker BP’s Valhall oil platform and Awesome Industrial Datasets. Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor. Published via Towards AI