Interpretable predictive maintenance connects a model’s remaining useful life (RUL) estimate with information that helps engineers understand how the estimate was formed. For teams working with multivariate sensor time series, a single prediction may not show which measurements mattered or when they became important. VisioRed, a research visualisation tool, brings local explanations, what-if modifications and forecasting into one workflow to help users examine those questions.
The distinction matters as predictive maintenance develops toward prescriptive maintenance. Predictive models estimate when equipment may fail; prescriptive approaches also explore possible actions or measurement patterns associated with a different outcome. These tools can inform investigation, but their suggestions need to be interpreted in context rather than treated as guaranteed maintenance instructions.
Why explainability matters for predictive maintenance
Industrial predictive maintenance commonly uses readings from multiple sensors over consecutive time steps to estimate a component’s remaining useful life. A model may produce a useful RUL estimate while leaving operators uncertain about the evidence behind it. Understanding which sensor features and time periods influenced a result can help users investigate the prediction and compare it with equipment knowledge.
Time-series explanations pose a particular challenge. A model’s input contains both features, such as sensor channels, and time steps. An explanation can therefore assign importance to every feature at every point in the input window. That level of detail may be difficult to interpret at a glance, while simply averaging importance across time can conceal how a feature’s influence varies.
Interpretable predictive maintenance with VisioRed
VisioRed is designed to explore predictions for multivariate time-series data. Its input consists of several features measured across a selected sequence of time steps, and its predictive maintenance model produces an RUL estimate. The interface combines local interpretation methods with tools for modifying input values and examining forecasts.
From time-step scores to feature importance
For local explanations, VisioRed uses LIME and LioNets to create neighbouring examples around the instance being examined and fit a transparent linear model. The resulting explanation can assign an importance value to each feature at each time step. To make the results easier to summarise, one straightforward option is to average a feature’s importance across the input window.
VisioRed also introduces iPCA, or interpretations through principal component analysis. It applies PCA separately to each feature, using the selected instance and its neighbours, then represents each feature with its first principal component. This reduces the representation from one value per feature per time step to one value per feature for the explanation model. The resulting feature importance maps directly to features without first averaging time-step scores.
The method can also retain information about time-step contributions through the coefficients of each PCA transformation. These coefficients are not direct time-step importance weights; they indicate how measurements contribute to the feature’s reduced representation. The research reports that iPCA captured feature importance as well as or better than averaging in its evaluated metrics, while noting that this is a method for interpreting model behaviour, not a causal analysis.
What-if analysis and modification recommendations
To explore how input changes affect a prediction, users can select a feature, a subsection of the time window and a modification. Supported options include adding uniform noise, adding Gaussian noise, replacing values with the feature mean, or replacing them with zeros. The tool then shows the effect of the modified sequence on the predicted RUL.
Because users could test many combinations, VisioRed provides recommendations based on feature importance. It selects the two features with the strongest negative importance and the two with the strongest positive importance, then tests the available modification types for those features. The recommendations are intended to help focus exploration on changes expected to move the RUL estimate most. They do not establish that a corresponding physical intervention is feasible, safe or effective; those questions require engineering and maintenance validation.
Conditional forecasting for prescriptive exploration
VisioRed’s XYZ neural model explores a different question: given an input sequence and a preferred RUL target, what values might the features take in future time steps? During training, XYZ uses an initial part of an instance and the predictive model’s output, then learns to produce the instance’s later measurements. During use, a user supplies the initial values and a target RUL, and XYZ proposes future feature values associated with that target.
The tool also offers three forecasting options to compare with XYZ: a custom neural forecaster, N-BEATS and a static baseline. The static forecaster copies the final forecast-window values and reverses their order. Comparing forecast outputs with XYZ’s proposed values can help users examine which measurements differ in a scenario aimed at a preferred outcome. It does not mean that the model has found a verified operating recipe.
Reading the turbofan example carefully
In the paper’s turbofan demonstration, a model predicts an RUL of 25.05 time steps for an instance. The example sets a desired target of 312 and examines a five-step forecast window. The neural forecaster estimates an RUL of 16.63 after the forecast period, while the prediction associated with measurements proposed by XYZ is 22.53.
The proposed measurements do not achieve the requested target in this example. However, their associated prediction is higher than the forecaster’s estimate, illustrating how the interface can compare scenarios rather than simply label one as successful or unsuccessful. The result is specific to the model and instance used; it should not be read as evidence that changing real sensor readings would produce the same equipment outcome.
Practical implications for engineering teams
For system integrators, reliability engineers and operations leaders, the main value of this approach is a more inspectable path from prediction to investigation. A useful workflow can include:
- Review the RUL estimate alongside feature-level and time-step-level explanations.
- Use controlled what-if changes to examine how the model responds to input variations.
- Compare conditional suggestions with independent forecasts and equipment knowledge.
- Validate any proposed maintenance action against physical constraints and operational requirements.
VisioRed demonstrates how interpretability, modification tools and forecasting can be combined to support exploration of predictive maintenance models. Its outputs can help users ask more focused questions about sensor data and model behaviour, but they remain model-based estimates. The practical implication is clear: treat explanations and proposed scenarios as decision-support information, and pair them with engineering review before translating them into maintenance action.
Access Official Technical PDF