Predictive maintenance teams may collect extensive vibration, temperature, pressure and torque data without having reliable labels that show when a machine became faulty. This makes conventional supervised feature selection difficult: a model cannot rank indicators by their relationship to fault classes if those classes are unknown or inconsistently defined. Unsupervised feature selection offers another route, using the observed behavior of health indicators over machine life to assess their prognostic value.
An adaptation of minimum redundancy maximum relevance (mRMR), called prognostic mRMR, combines that relevance assessment with a check for redundancy among features. A rotating-machine case study illustrates how the method can help engineers narrow a large feature set without relying on pre-existing fault labels.
Why feature selection matters in predictive maintenance
Predictive maintenance commonly involves acquiring sensor data, constructing health indicators, identifying health states and estimating remaining useful life (RUL). Feature selection sits between indicator construction and downstream tasks such as state classification or RUL prediction. Its purpose is to identify indicators that carry useful information without retaining many that describe essentially the same behavior.
Two common selection approaches have limitations. Prognostic metrics can rank individual indicators but may ignore overlap between them. Supervised methods, including conventional mRMR, use known labels to assess relevance and redundancy, but trustworthy labels for machine malfunctions are not always available. Assigning time-to-failure as a substitute label can also be problematic: degradation may begin well after installation and may not progress linearly.
Unsupervised feature selection using prognostic behavior
Prognostic mRMR replaces the usual comparison between a feature and a class label with a relevance score derived from three characteristics of a health indicator: monotonicity, trendability and prognosability. Together, these measures ask whether an indicator changes consistently over time, follows a comparable pattern across machines, and shows separation between its starting and failure values.
Three measures of indicator relevance
- Monotonicity: Measures the strength of an indicatorâs association with time across run-to-failure records. The proposed method uses Spearman rank correlation, which can capture monotonic non-linear relationships and is less sensitive to outliers than the slope-counting approach it replaces.
- Trendability: Assesses whether an indicator has a similar degradation shape across machines. The method compares pairs of machine histories, which may have different durations.
- Prognosability: Evaluates how closely failure values cluster relative to the indicatorâs range between the start and failure of machine life. The failure value is calculated over a user-defined time window to reduce the influence of a potentially noisy final observation.
The three measures are scaled and combined into a fitness score. In the reported implementation, they receive equal weight. The weights and the time window used for start and failure values are design choices, so they should reflect the application and available run-to-failure data.
Accounting for redundancy between health indicators
A high-scoring indicator is not automatically a useful addition to a selected set. It may closely track an indicator already chosen, adding little distinct information. The mRMR principle balances maximum relevance against minimum redundancy: it favors indicators with strong prognostic scores while discouraging excessive similarity among selected features.
For the redundancy comparison, the study evaluates correlation, mutual information and dynamic time warping (DTW). DTW compares time-series patterns while allowing local stretching or compression of their timelines. That matters when similar degradation behavior begins at different points in different machinesâ histories. A low DTW distance indicates similar patterns and is converted into a redundancy score.
The selection is incremental rather than an exhaustive search across every possible feature combination. It first chooses the indicator with the highest relevance score, then adds features according to a combined relevance-and-redundancy objective until the desired set size is reached. This makes the requested number of features an explicit input to the process.
What the rotating-machine case study found
The application involved a high-speed rotating condenser operating at 7,500 RPM. Data from 11 run-to-failure histories were used to construct 317 health indicators, including vibration time- and frequency-domain features and aggregated measurements from other sensors. Vibration measurements included statistical measures such as RMS and kurtosis, as well as frequency-related indicators.
Because the evaluation required classification labels, the researchers created artificial healthy and faulty periods using engineering assumptions about the time before failure. Those labels were used to compare feature-selection approaches; they were not required by the proposed unsupervised selection step itself. The distinction is important when interpreting the results: the method selects features without labels, while the case-study evaluation uses assumed labels to measure classification performance.
In this dataset, DTW-based trendability performed better than the tested correlation-based alternatives for some feature-set sizes, particularly in the range of 20 to 40 features. Accounting for redundancy did not consistently improve results when selection began across all features. After first restricting the candidates to the 100 highest-relevance indicators, the DTW redundancy approach performed slightly better than relevance-only selection, reaching a reported maximum F1 score of 0.78 with 30 selected features. These findings apply to the evaluated machine and setup; they do not establish that the same settings will be best for every asset.
Compared with conventional mRMR across multiple assumed fault-label windows, prognostic mRMR performed better on several aggregate comparisons and was consistently more effective at selecting compact feature sets in the reported analysis. Results varied with the label definition, reinforcing the practical difficulty of using assumed labels as a benchmark.
Practical implications for engineering teams
For teams building condition-monitoring pipelines, the method provides a way to rank and reduce candidate health indicators when labeled fault events are scarce. It also makes feature overlap an explicit consideration rather than treating every highly ranked indicator as independently valuable.
- Build run-to-failure histories that capture observable degradation where possible.
- Define the start and failure windows used for prognosability based on the asset and application.
- Consider DTW when machines degrade at different rates or degradation begins at different times.
- Preselect indicators with adequate relevance before applying redundancy reduction, as the case study found this important to the value of the combined approach.
- Validate selected features on the target equipment and downstream task rather than assuming one feature count or metric will transfer unchanged.
Unsupervised prognostic mRMR does not eliminate the need for engineering judgment or validation. It offers a structured way to combine degradation-related relevance with feature diversity when fault labels are unavailable. For industrial automation teams, that can support a more focused set of health indicators for subsequent fault detection or RUL modeling, while keeping the assumptions and limits of the available data visible.
Access Official Technical PDF