📘 Resource Overview & Documentation Details

You are viewing the technical directory listing for Predictive Maintenance Without Vibration Data. This engineering resource is curated under our Predictive Maintenance division to support automation specialists, control systems engineers, and manufacturing technicians looking for verified reference materials.

🔍 Contextual Abstract & Technical Scope

Share Summary:LinkedInEmail

Predictive maintenance for rotating machinery often relies on vibration monitoring, particularly when the goal is to identify bearing degradation. But vibration data are not always available. A case study of a rotating condenser inside a synchrocyclotron shows how other operating signals can be used to build a failure-warning system—and why the alarm logic should reflect maintenance needs, not just model accuracy.

Why bearing-box failures need advance warning

The rotating condenser, or RotCo, is part of a proton therapy system. It has a stator and a rotor with eight blades, running at a constant 7,500 rpm. Rolling bearings in the bearing box support the shaft and help limit friction between moving components. A failure can disrupt system availability and delay patient treatments, making useful advance warning operationally important.

In this application, operators may see signs of weakness only hours before a failure. The desired warning window, however, is several days: an alert received early in the week could allow maintenance to be scheduled for the weekend. The challenge is that the system’s sensor set does not include vibration measurements, which are commonly used in bearing diagnostics.

Building a predictive maintenance model from available signals

Eight sensors recorded RotCo operating conditions at one sample per second. The measured signals included rotor speed, applied torque, current, voltage, bearing-box temperature, motor temperature, rotor temperature measured by pyrometer, and vacuum-chamber pressure. Vacuum is relevant because friction in the bearing box can generate gas release and affect the chamber reading.

The study gathered about a year of data, covering nine replacements: eight following failures and one preventive replacement. To make the time series more manageable, readings were averaged over 30-second intervals. The resulting data were cleaned to remove physically impossible values associated with sensor errors.

Feature extraction then summarized the signals over adjacent, non-overlapping two-hour windows. Time-domain features included the mean, standard deviation, skewness, kurtosis and number of peaks. A peak was defined relative to the window’s median and median absolute deviation. Frequency-domain features were derived using a Fourier transform: amplitudes were divided into five frequency bins, and the five largest amplitudes in each bin were averaged.

Two input designs were compared. The simpler design used features from one two-hour window. The history-enriched design also included the 12 most recent windows, representing the previous 24 hours. That added temporal context while increasing the initial feature set from 80 to 960 variables.

Classification, aggregation and a business-focused alarm

Rather than estimate a component’s remaining useful life, the approach treated the task as binary classification: would a failure occur within a defined time horizon? Training windows from the five days before a failure were labelled critical; earlier windows were labelled nominal. This framing directly addressed the operational need to identify risk within a useful planning window without assuming a specific physical degradation process.

A support vector machine (SVM) with a radial basis function kernel classified each window. Its decision value—the position relative to the classifier’s separating boundary—was passed to a second-stage aggregator. The aggregator calculated an eight-hour rolling average and raised an alarm when that average exceeded zero. Smoothing the output this way made the alarm decision depend on sustained evidence rather than a single classification.

The model was evaluated using a custom business score based on the interval between an alarm and a replacement or failure. Seven days of warning received the best score. Scores declined for predictions closer to the event and for warnings earlier than seven days, reaching zero beyond 15 days. This reflects a practical trade-off: a late alert may not leave enough time to plan maintenance, while an excessively early alert may not make effective use of the component’s remaining service life.

What the results suggest about historical context

Feature selection and SVM parameter tuning were performed across nine folds, with each run-to-replacement period held out in turn for testing. Backward feature selection removed signal or feature groups when doing so improved the cross-validated business score, helping reduce the risk of overfitting from an unnecessarily large input set.

For the history-enriched model, the selected features came from current, bearing temperature and pyrometer temperature. The retained feature types included peak counts, frequency bins and standard deviation. This reduced the input from 960 features to 144. Voltage, among other candidates, was removed during selection; in this dataset, including every available feature did not produce the best result.

The history-enriched approach achieved an average business score of 0.76, compared with 0.57 for the single-window design. The study also reported that alarms for most run-to-failure cases arrived several days before the event; one prediction was considerably earlier than the preferred window. These results are promising for the specific dataset, but the small number of replacement events means they should not be treated as a universal performance guarantee.

Practical implications for industrial teams

This case illustrates several useful principles for condition-monitoring projects:

  • Start with the signals available. When a preferred measurement such as vibration is missing, existing process and temperature signals may still contain useful indicators.
  • Represent operating history. Recent trends can add context that a snapshot of one time window misses.
  • Design alerts around maintenance decisions. A model score should be evaluated against the lead time and intervention options that matter to the plant.
  • Keep the alarm logic distinct from classification. Aggregating model outputs over time can translate changing risk estimates into a more stable decision signal.

For engineers and operations leaders, the broader lesson is that predictive maintenance is not only a modelling problem. It also requires careful data preparation, validation across distinct operating periods, and an evaluation metric tied to the realities of scheduling work. In this synchrocyclotron application, combining time-series features with a classifier and rolling aggregator offered a way to forecast bearing-box failures without vibration data. Further testing would be needed before applying the method to other machines or operating environments.

Access Official Technical PDF