Moving forward with vibration-based condition monitoring
30 May 2017
Big data pattern recognition analysis of the measured vibration data bank from a number of identical machines from one plant or many plants can form the basis for carrying out centralised vibration-based condition monitoring (VCM). The concept 'Internet of Things (IoT)' or Industry 4.0 can be used to visualise the approach, as Jyoti K. Sinha from the University of Manchester, explains
Vibration-based condition monitoring (VCM) of rotating machines is well-known and adopted in industries. The objective of the VCM is to predict the presence of any fault(s) at early stage so that maintenance or remedial action can be performed before any catastrophic failure. These practices generally avoid machine downtime and maintenance overheads, and maintain plant safety.
The data collection and signal processing required in the VCM is now easier due to a number of technology advancements in instrumentation and signal processing capabilities over the past two to three decades. However even with advancements in the technologies, the historical vibration data from a machine on a foundation are found to be difficult to apply directly for fault identification in identical machines installed on different foundations. In general this is due to the different dynamic behaviour of the machines that results from their different foundation layouts. It is very common for a plant to use many identical machines to meet production requirements.
The application the VCM-based fault detection process individually for each identical machine at one plant or many plants is a time- consuming process. Hence a generic approach is required to make the fault detection process easy. This paper presents such an approach and proposes use of the measured vibration data from many identical machines at one plant or more different sites, that may or may not have identical foundation structures, including all auxiliary structural components in a common data bank together with machine health status. This data bank can then be used for training purposes so that the new measured vibration data from any identical machine can be used to identify the fault or health of that machine.
Current VCM practice
Figure 1 typically demonstrates the current VCM approach for the detection of any fault(s). The vibration measurements are generally carried out at each bearing pedestal in the machine in 3 mutually perpendicular directions (e.g., vertical, lateral and axial directions for a horizontal machine as per the ISO code. Overall vibration value (RMS vibration velocity) is then compared with the ISO severity limits. The presence of harmonic peaks (1x, 2x, 3x,…) and/or sub-harmonic peaks related to the machine RPM in the vibration spectrum/spectra and their amplitudes trending is then used to identify the presence of different faults.
Figure 1: Simple VCM approach
A little more involved signal processing in the VCM approach used in the industry for the rotating machines is shown in Figure 2 where the measured vibration data during the machine normal or/and transient operation are analysed in a number of different formats to identify many well-recoganised defects.
Figure 2: Different data processing used in the VCM approach
Figure 3: Complex process to identify any fault in a machine
Most machines have many bearings and hence more measured vibration data. A typical machine with 4-ball bearings is shown in Figure 3 with the measured vibration spectra and orbit plots. This makes the vibration analysis and the fault detection processes data intensive and complex. Therefore the current practice poses the following limitations:
• Greater number of instruments that generate more data to analyse.
• Fault detection process is subjective.
• Requires experience & engineering judgements to carry out fault detection.
VCM is a practice which was developed over several decades. The use of a greater number of sensors at each bearing pedestal and simple signal processing in different formats seems reasonable. However, the following significant advancements in technologies have been observed over two to three decades:
• Computational power
• Facility of integration of all instruments/sensors to a system for data collection and storage
• Number of different signal processing techniques
Proposed approach: Big data analysis
Considering the advancements in computation power, signal processing techniques and instrumentation, the following improvements are proposed and adopted.
- Sensor reduction: Just one sensor per bearing pedestal at 45° from the horizontal and vertical directions is proposed. It is typically shown in Figure 3. The reduction in the number of sensors and related instrumentation will to reduce the cost of initial setup and maintenance cost thereafter.
Figure 4: Only one vibration sensor at 45° per bearing pedestal
- Data Bank: Generation of measured vibration Data bank from all identical machines and their fusion together with the machine health status is then suggested. It is good to use the parameters like – Overall RMS vibration value, amplitudes at 1x, 2x, 3x, …, Crest factor or Kurtosis, Spectrum Energy (SE), etc at each bearing from all identical machines. These parameters are very likely to be available for the machines in old plants, together with the machine’s heath status. The vibration data from a bearing of a machine can be stored as shown in Figure 5. Similarly the data from all bearings from many identical machines can then be populated in the data bank.
Figure 5: Big data storage format per bearing in a machine
- Mathematical Model and Computational Power: The mathematical model can then be constructed for the data bank by an appropriate pattern recognition technique by utilising the computational power for data training and the fault classifications (machine health status). Typical results of a typical rotating rig where the rotor was supported through four ball bearing but different foundation structures and flexibilities are shown in Figure 6. This figure clearly indicates the vibration data and machine health status (heathy or faulty) can be separated clearly by properly training the data bank from many identical machines from a plant site or many different plant sites worldwide. Here four different machine health conditions (healthy, shaft misalignment, crack, rub) were experimentally simulated in the rig.
Figure 6: Data classification with different machine faults
Fault Detection Approach
Once the mathematical model with the training algorithm of data bank and their classifications related to machine status (healthy or faulty) is available then the model can be used to access the status of any identical machines – pictorially demonstrated in Figure 7. This typically shows that the vibration data from a machine with the rotor crack is fed onto the training data bank algorithm. The mathematical model then estimates the pattern of the new data (Blue Dot) which clearly classified within the crack box. Hence the machine can be diagnosed as the rotor crack defect without involving engineering judgments and the experience of the maintenance engineer doing the routine VCM.
Figure 7: Fault detection process
Figure 8: Typical rotor crack detection (new data 'Blue Dot' in crack box)
The proposed method uses the concept of big data analysis to unify the identification of faults in the identical machines at a plant site or many plant sites across the globe. This approach is expected to lift the limitations of experience of any maintenance professionals, engineers and practitioners involved in the VCM and fault detection process. This approach is likely to be the future for the VCM. It can be managed centrally within an organisation, irrespective of machines installation locations. This is a realistic and feasible approach in the VCM application using the concept of 'Internet of Things (IoT)' of Industry 4.0.
• Dr Sinha is programme director, reliability engineering and asset management (REAM) MSc course head, Dynamics Laboratory,
School of Mechanical, Aerospace and Civil Engineering (MACE), The University of Manchester. Dr Sinha would like acknowledge a number of his PhD students who were involved in this research.