An Integrative Approach for Machinery Diagnostics

Fault identification and estimation is an important and necessary step in Condition Based Maintenance. In general, there are two prevalent methods that are used for this purpose, data driven techniques and model based techniques. Data driven techniques use data collected from experiments, learn about the system and then use this knowledge to infer the system’s current state of health. Model based techniques use models to derive the knowledge that can be used to determine the machine’s condition. Each of these methods has its own limitations and strengths. This research proposes a novel approach that integrates the information from both data and models in an optimal fashion to provide accurate diagnostic information about the health condition of machinery. Specifically, rotor-bearing systems are used as test cases to develop these methods.

Overview of Diagnostics and Prognostics
Diagnostic and Prognostic Methodology

Figure 1 shows the basic steps involved in any diagnostic methodology. Sensors are used to record various signals from the machine but these signals cannot be directly used as such signals are often very large in size and noisy. Signal processing techniques are hence used to derive useful and compact information about the system from these measurements. These information packets are called features and the process is called feature extraction.

Diagnostic and Prognostic Methodology

Typically the tasks of a diagnostic and prognostic methodology are as follows.

  • Detect any change in performance.
  • Determine if the change is a degradation in the system.
  • Determine the cause of degradation.
  • Identify the amount of degradation.
  • Estimate the remaining useful life.

These duties need to be performed in a sequence. Figure 2 shows the typical sequence of operations to be performed as the degradation in system increases.

Objectives of the research
The objectives of the research are as follows
  • Develop detailed nonlinear dynamic models of machines with and without faults.
  • Combine models and data and use appropriate signal processing techniques to develop features that are rich with information on the state of the machine.
  • Compare the performance and interaction of these novel features with the existing data based features.
  • Develop an optimal set of features and diagnostic methodology for machine fault detection.

It should be noted that model and data integration is achieved in two ways. First, features are developed by combining data and models (model based features). Second, integration is also achieved by smart fusion of model and data based features in a hybrid feature set.

Classification performance

Referring to Figure 3, two machines belonging to a same class have certain similarities; these similarities are captured in the physics-based models. Further, each machine based on its environment has a certain individuality; this individuality is captured in the data. Thus, using both models and data could increase the efficiency and accuracy in performing diagnostics and prognostics. The current research aims at integrating the models and data to better predict the performance of the machine. Also models generalize the system so that efficient algorithms based on models can be applied to various systems with minor modifications.

This new idea of integration of information has been developed and validated to detect, and to identify type and severity of localized defects in rolling element bearings. Rolling element bearings have been chosen for this study as they are the load carrying elements in many high speed machinery, one of the primary sources of nonlinearity and often the cause of failure.

From figure it can be seen that
  • Data driven features outperform the model based features when considered independently. Single data driven or model based feature performs well by itself.
  • In both cases, the performance increases steeply after the first couple of features.
  • It is interesting to note the decline in the performance of the model based features
  • After the first three features. This is probably because the rest of the model based features are redundant.
Publications List
  1. Kappaganthu K., Nataraj C., 2011, Nonlinear modeling and analysis of rolling element bearing with clearance, Communications in Nonlinear Science and Numerical Simulation, 16(10), 4134–4145
  2. Kappaganthu Karthik, Nataraj C., 2011, Modeling and analysis of outer race defects in rolling element bearings, Advances in Vibration Engineering, 11(4),4, 371-384
  3. Kappaganthu Karthik, Nataraj C., Samanta B., 2009, Model Based Bearing Fault Detection Using Support Vector Machines, Annual Conference of the Prognostics and Health Management Society