Developments In Process Parameter Monitoring

Published Date: 02 Nov 2017

Introduction

Preventive/Predictive programs are usually focused on vibration technology, oil analysis, thermography and similar technologies, yet it is important that any implemented predictive maintenance program should take into account process parameter monitoring on some level.

Predictive maintenance is seen as a general philosophy whereby various maintenance tools that monitor the actual condition of a plant are utilised to achieve optimal plant operation. This mean ensuring not only that equipment is not operating within, say an acceptable vibration range, but more importantly, is operating in the most efficient way, in turn, the most cost effective way. (Mobley 2002)

Process parameter monitoring is one of the most important condition monitoring techniques, as in addition to indicating potential failures, it also alerts operations of deviations from optimal operating conditions. This thus results in savings not only through downtime avoidance, costly overhauls, etc, but through the corrections to the inefficiencies pointed out through a proper Process Parameter Monitoring program; these have the most impact on plant production and profitability.

Parameters monitored on a plant could be exhaustive, yet usually include:

Temperature - Process fluid, intake air, exhaust, rotor bar temperature etc

Pressure – suction, discharge, etc.

Current - motor over current conditions

Voltages - e.g for voltage imbalances

Flow rate

There may also be custom parameters such as:

Sub-cooling/superheat (Refrigeration)

Process fluid density

Brake horsepower

Equipment efficiency

Differential pressures

In order to benefit in any significant way from process parameter monitoring, the selected parameters are required to be properly, and accurately monitored/recorded over a reasonable time frame.

This could be carried out either manually or automatically via microprocessor based systems, although most systems will incorporate a hybrid of the two. An extensive manual process parameter program on a medium to large plant would be very resource intensive, thus automated data collection, monitoring and analysis is required.

In recent years, process parameter monitoring and analysis has evolved into more than just trending data, but using advance pattern recognition algorithms to detect deviations from the norm. the integration of artificial intelligence into the analysis have been used to train the programs to not only detect, but diagnose faults based on either historical data, or fuzzy rules. This paper will expand on these advances in the implementation of Process Parameter Monitoring, and also outline associated limitations.

Developments in Process Parameter Monitoring

In this section we consider a few advancements in process parameter monitoring, laying out the advancements in a rough order of less to more advance developments, i.e. more traditional smaller systems, to larger complex programs.

Advancements in Data collection

Process parameter monitoring developments have traditionally been focused on data collection, its efficiency and accuracy, and the ability to present such data in ways that could be better analysed by the engineer.

Most plants are built with a range of instruments such as thermometers, and pressure gauges to allow operators to spot any irregularities, amongst other things. The problem has been that these gauges have short life spans, and require calibrating often. It is also a resource intensive approach to have a technician log measurements manually.

Technological advancements have led to process parameters being measure easier, faster, more accurately, and remotely. This has led to more efficient plant operation and better utilisation of maintenance staff. Sensors, via a standard 4-20mA, or 0 – 5 DCV output are capable of tying into a number of standard or custom process management programs, thus accurately relaying process temperature, flow etc directly to a control room on the plant.

Equipment self-monitoring

Various smaller equipment, e.g. chillers, compressors, pumps, have in recent years, developed the feature of process parameters being directly displayed on equipment via on-board controllers. These systems provide basic alert notifications through use of statistical control charts, allowing the operation to further investigate the alert. Some systems will also stop the process operation as a safeguard to prevent damage to the equipment.

At the same time these controllers have evolved via use of various integration protocols to allow integration into overall plant monitoring systems, and building management systems.

The integration of Process parameter monitoring with other CBM data – Sensor fusion

Sensor fusion is essentially the processing of data derived from various sensors into a result/output that leads that output to be more useful than possible when sensors are used individually. (Elmenreich 2002)

Sensor fusion may simply be the integrated signal of various temperature transducers around a process tank to give an average temperature of the process fluid.

With the complexity of machines, and hence seemingly exhaustive list of failure modes, process parameter monitoring has being evolving to ‘system’ parameter monitoring. Multi-sensor based CMB gathers the various data and uses a data fusion model, similar to that in figure 1 to bring everything together and perform fault detection and diagnostics.

Figure - Data fusion model

An example of where this development in parameter modelling has been visible is in the use of Motor Current Spectrum Analysis. Unlike vibration analysis, the more recent technique of MCSA does not benefit from the range of specifications indicating limits for various faults, and diagnostics has been difficult given that various faults generate very similar signatures. By monitoring motor speed, torque, acoustics, vibrations, temperature, etc, the incorporation of all parameters leads to better diagnostics and prognostics. (Xue, Sundararajan and Gonzalez-Argueta 2007)

Some systems have even began to incorporate process parameter monitoring into the datalogging of vibration monitoring. This leads to better analysis of vibration spectra as the analyser is made aware of the conditions under which the equipment is operating at the time of elevated vibration levels.

Development of statistical data driven fault detection

Statistical modelling for process parameter fault detection analysis has been around for some time, yet over the years, more robust statistical methods have been introduced. Fault detection was initially based on control charts (something still widely used today). These set control limits for real time monitoring of the machine parameters. Developments in this area included the Multivariate statistical process control, whereby advances in data/signal acquisition led to the ability to process many parameters simultaneously.

Yet the real development in came about in the 90’s when Artificial Intelligence was factored into the analysis of process parameters for condition monitoring.

Integration of Artificial Intelligence in to Process Parameter Fault Detection and Isolation Systems

The implementation of Artificial Intelligence into process parameter monitoring is fairly recent, with work being carried out primarily in the past decade (Masory 1991) (Das, Maiti and Banerjee 2012). AI techniques commonly used include Artificial Neural Networks, and Fuzzy Logic, with the former being utilised more often. These are expanded on in the subsequent paragraphs. Genetic Algorithms is another technique sometimes used in hybrid systems, yet this will not be delved into.

Artificial Neural Networks

Artificial Neural Networks (ANN) are modelled on the operation of the human brain, whereby an input is placed into a the neural net of ‘neurons’ and undergoes weightings, that are based on differences between the inputs and ‘ideal’ outputs, and then gives an output. While the output may not initially be correct, the beauty of ANN is that using all the input data, the system teaches itself what outputs are expected, given a host of inputs. (Blais and Mertz 2001 )

This ‘learning process’ is the selling point for ANN, yet in supervised learning methods large validated data has to be inputted into the system to reduce the learning time.

There are two approaches to ANN. The first is a case where clear differentiation between the fault condition and normal condition, as well as between various fault conditions, is established. The second is where residual values (the difference between the measured parameter and the expected value) are generated and inputted into a second network; high residual values would indicate a fault. (Das, Maiti and Banerjee 2012)

http://www.fhwa.dot.gov/publications/research/safety/98133/images/ch02/p017-1.gif

(http://www.fhwa.dot.gov/publications/research/safety/98133/images/ch02/p017-1.gif)

Figure - Simplified schematic of Neural Network

ANN is suited to processes that are very difficult to model, where set rules are not evident, and non-linearity is an issue.

An example of such an application is depicted in the simplified schematic of a neural network for determining flank wear on a cutting machine, whereby ANN was used to predict at what point the cutting tool would fail based on the cutting speed, cutting force, acoustic signals, as well as vibration signals.

Fussy Logic

This is a formal logic concept, that unlike Boolean logic that is based on true and false outputs, fuzzy logic produces an output on a degree of uncertainty, e.g. a degree of tallness, not tall or short, much how human reasoning behaves. The inputs and outputs of the fuzzy system are set in fuzzy representations (fuzzy sets) , while they are linked via well defined fuzzy relationships (fuzzy rules).

http://www.scielo.br/img/revistas/ca/v14n4/a05fig01.gif

Many complex machinery failure modes and mechanism are understood by humans, yet the knowledge on these failure modes/mechanisms is vague so as to make it very difficult or almost impossible to model for a computer to process (Wang 2003). This is where fuzzy logic comes into play; this approach is very useful in monitoring very complex systems that are not fully understood, so as approximating the outputs would suffice. It is also noted that with approximation comes faster processing times.

Limitations of Process Parameter Monitoring Techniques

While limitations of many systems used today will be addressed, it is noted that advancements in artificial intelligence are slowly breaking down these barriers, and meeting new ones.

Limitations based on traditional analysis of process parameters

As noted, traditional analysis of process parameter is still widely used on plants, smaller equipment, and in less sophisticated predictive maintenance programs. Traditional methods are based on analysis using simple statistical methods based on single data streams, more complex multivariate statistical methods, threshold limits (control charts), or system identification methods. Some of the drawbacks include;

Inability to provide real-time diagnostic in cases where all parameters are not available, or incomplete, or where conflicting results may appear.

These methods are not usually able to deal with non-linearity, where clear relationships cannot be established between a parameter, and a fault.

Accuracy in Fault Isolation/Diagnostics

Root cause analysis based on traditional process parameters is usually difficult as discrepancies in a particular parameter may be a symptom of numerous faults in a system (Mobley 2002).

For example, in a centrifugal pump, a low discharge pressure may be due to;

Pump cavitations

Entrained air in the system

Blockage of the inlet strainer

Internal wear

Leakage of the pump or pipes

Wrong rotation of the pump, etc.

Traditional systems would be great at detecting a fault, as control charts would set various limits for alerts. This would result in an operator using other methods to identify the root cause of the alert.

Priori based model approach has had some success in fault diagnostics, yet this assumes the system could be accurately fitted to mathematical model.

Complexity of modelling

Another area where non-AI based techniques have reached their limit is in model complexity. Modelling of process parameter monitoring for complex machinery could be very difficult. The use of model-based, or Priori type fault detection and diagnostics has proven to very accurate, yet some processes are too complex to be modelled, hence the need for fuzzy logic hybrid systems.

Monitoring during Transient operation

Most traditional process monitoring programs are not able to accurately flag a system during transient operation.

In the case of an air conditioning screw compressor, during start up, the system has to ignore process parameter limits and instead sets a buffer time for the system to achieve a steady-state. During the transient operation of a system, system pressures and temperatures are higher than would normally be deemed acceptable, hence if they are taken into account, would be classified as faults.

With a traditional system, at part-load operation, process parameter monitoring may mask the faults by not considering the various parameters relative to the load of the system. Such a case was seen with a York screw compressor and its associated 12 year old controller, whereby the discharge pressure and motor current load were both just below the alert levels in the monitoring system, yet the load was only at 60%. If the system was operating at full load there would be an obvious alert condition, yet due to the non-linear relationship between the slide valve loading and discharge pressure, the interrelation was not present in the system.

Even with more advance analysis techniques, transient and non-linearity in operation has been a challenge. The use of Priori based models for fault detection is one analysis technique that while very accurate in fault detection and isolation, has only proven satisfactory when it comes to doing so in transient situations. (Das, Maiti and Banerjee 2012).

Limitations of more Advance Process Parameter Monitoring Analysis

To counteract the limitations of traditional trending, advance pattern recognition algorithm have been used in the analysis of process parameter monitoring.

Expansion of the term ‘Process’

In order for an advance process parameter monitoring system to be truly robust, it is required that they incorporate other condition monitoring techniques into its analysis.

This means that in addition to the various process parameters, vibration, acoustic, or even thermographic signals are incorporated to lead to better predictability and fault isolation abilities. This is where the concept of data fusion came into play.

Limitations arising from Training of Neural Networks

While Neural networks could be very powerful, their limitation usually lies with the ability to train the network, as well as the difficulty and time involved in doing so. The models require large databases of information and carefully designed system models to be effective in the training process. The complexity of modelling some systems, as well as the layers of hidden nodes and weightings make development of these systems very complicated (Aljoumaa and Soffker 2010) (Talebi, et al. 2010).

One approach to reducing the training time for such systems may be by incorporating some hybrid of fuzzy-logics into such models, thus giving the system some level of accuracy in its early life and allowing more accurate predictions to be done as more data is acquired.

It is noted that some studies (Li 2001) have sort to implement a hybrid Fuzzy Neural Network approach. The logic is that fuzzy systems are easily put together for complex systems, yet are not good at the machine-learning process. ANN, on the other hand are good at learning, yet do not work will with expressing rules. The hybrid focuses on the strength of each method.