The vibro-acoustic signatures of monitored machinery contain information that can be used to indicate the development of faults in rotating elements such as bearings, gear-wheels or shafts, as well as development of faults in the structure. R.K. Diagnostics' unique technology understands the relation between the healthy machinery internal processes and their manifestation in the vibro-acoustic signature. It performs high-resolution analysis of this data to detect anomalies that point to the development of a future breakdown, thus delivering the highest levels of reliability and accuracy of early detection.

In addition to the vibro-acoustic data, this breakthrough technology uses other types of parameters, e.g., temperature, pressure, currents, and others, to deliver reliable diagnostics and prognostics to a wide array of critical machinery.

The rich information analyzed/crystallized by the system is integrated in the health assessment phase using sophisticated artificial intelligence algorithms and expert knowledge. Physical model based algorithms are implemented for prediction of the remaining lifetime of a component.

The vibration based Diagnostics & Prognostics process is divided into four major stages:

  • Data acquisition and raw signals validity check
  • Signal processing
  • Feature extraction
  • Artificial intelligence methods for Classification, Pattern recognition and Fusion

Data Acquisition

Data acquisition is the first stage in any diagnostics system. Accurate data acquisition is critical to allow system sensitivity to small changes in the vibro-acoustic signatures and to prevent false alarms. The quality of the acquired data is determined by the following activities:

  • Optimal selection of sensors position, direction, characteristics.
  • Definition of the data acquisition system parameters, e.g. signal-to-noise considerations, signal conditioning, requirements of analog to digital (A/D) converters, requirements of memory management and processing power
  • Selection of machinery operating conditions for recording (operational regimes, repetitiveness, components loading parameters, effects on signals, etc.).
  • Selection of the most appropriate snapshots for recording. Because recording of wideband signals requires large memory space, only several snapshots are sampled. The embedded system applies this selection automatically.
  • Definition of raw signals validity checks. Validity checks are applied to determine problems of measurement and for signal quality scoring

Signal processing


The goal of the signal processing stage is to separate the effects of the possible failure modes of each component from among the numerous excitation sources of vibrations in a complex machine.

The signal processing stage maps the raw signal from the time domain into different processing domains ("views"), thus obtaining the vibro-acoustic signatures that characterize the specific machine in a specific operating mode.

R.K. Diagnostics' signal processing algorithms accurately distinguish between different excitation sources, allowing for the fact that the signals recorded during the Data Acquisition stage are polluted by various sources of noise. In most cases the signals are also non-stationary because of their dependence on rotation speeds and varying operating conditions. This implies increased complexity of the signal processing algorithms. The technology applies highly innovative multi-perspective signal processing algorithms that generate various "views" of the information carried in the raw signals. The software applies a unique combination of many signal processing algorithms, including:

  • Power spectrum estimation
  • Statistical moments
  • Spectrograms
  • Enveloping
  • Synchronous time averaging
  • Cepstral analysis
  • Synchronized resampling
  • Special time-frequency/RPS-orders representations
  • Orders tracking
  • Adaptive clutter separation

The signal processing algorithms enable high accuracy rotating speed analysis, which is particularly important for more sophisticated vibration analysis algorithms, i.e. cyclo-stationary analysis methods.

Feature extraction

In the Feature extraction stage, the system assigns values to a set of health indicators (features) that are correlated with various failure modes of the mechanical components. A selection of features that provide the best indication to the evolving damage is a key factor affecting the quality of the diagnostic conclusions. The numerous features generate a hyper-space (tens of thousands of features for complex machinery) that is managed by the algorithm.

Features trending

The feature extraction stage plays an important role as a sophisticated data reduction process extracting only the relevant information carried in the wideband signals. Consider the following example: The space required to store wideband signals sampled at 50kHz for 120 seconds from 5 sensors is approximately 60MB.
The amount of space required for storage of extracted features is ~100kB, i.e. data reduction of ~10-3.

The features extraction implements two types of algorithms: model-based and novelty detection.

Model-based algorithms

Model based algorithms are used for detection of patterns associated with known mechanical failure modes.

In most cases, the available data contains examples of normally operating systems. The model-based approach is a way to overcome the shortage of data examples that cover all possible failure modes. The model based feature extraction relies on a thorough understanding of the entire system mechanics, the failure mechanisms, and their manifestation in the vibro-acoustic signatures.

The model based algorithms cover a diverse set of mechanical failures. Examples are:

  • Damage in bearings (damages of the surfaces as well as deformations of the inner race, outer race, rolling elements, or cage)
  • Damage in gear wheels (distributed and localized damages, surface wear, partial or total tooth break, cracks)
  • Misalignment or imbalance
  • Oil whirl or whip
  • Pump malfunctions
  • Rotor blades partial or total loss
  • Rotating stalls
  • Installation problem

Novelty detection

Novelty detection algorithms are used to detect anomalies representing unknown or unanticipated failure modes.

Novelty detection algorithms for vibro-acoustic signatures are capable of differentiating between anomalies due to rotating parts and structural changes.

Artificial intelligence classification methods, pattern recognition and fusion

In this stage, all the features derived in the previous stage are analyzed and combined (fused) with all the other available parameters. The tasks are to recognize evolving failure patterns automatically, and to provide accurate and specific alerts related to the specific problematic component or failure mode.


Since examples of all possible mechanical failure modes are not available, the pattern recognition task becomes a one-class novelty detection classification problem. In order to achieve better performance, the system usually incorporates expert knowledge with self-learning capabilities that enable it to gradually improve performance over time.

R.K. Diagnostics' technology applies hybrid decision schemes capable of incorporating expert knowledge provided in different forms: physical models, rules defining the relationships between different parameters, patterns of parameters associated with system performance, etc.

Several pattern recognition methods are used for classification:

  • Clustering
  • Fuzzy logic
  • Self-organizing maps (SOM)
  • Neural networks
  • Bayesian networks
  • Support vector machines (SVM)
  • Variety of novelty detection schemes

The overall goals are to improve the detection rate and to diminish the rate of false alarms. It is possible to improve both rates over time, by improvements made in all four processing stages. Improvements in the first three stages will mainly affect the detection rate. Improvements made in the fourth stage (decision and fusion) will mainly affect the false alarm rate.

Copyright © R.K. Diagnostics 2009. All Rights Reserved. CFM Engine photo by Eric Drouin courtesy of Snecma