Industry

Power plant maintenance

Power plant

Problem

Maintenance services based on information system implementation concept

Maintenance of valuable power plant assets is one of the key processes at a running power plant. Maintenance tasks are planned, controlled and reported by a maintenance system. The goal of effective maintenance process is to achieve short overhaul breaks, low costs and good performance. A well planned maintenance system supports maintenance personnel in these tasks.

The goal of the project is to create a standard and well structured procedure to implement a maintenance system.

Web based architecture for process information and condition management

Desktop client applications take time and effort to install, administer and maintain. Web applications offer improvements to these problems but there are concerns on how well web applications work in environments where massive amounts of real time data and graphical displays are needed.

Service concept development for automatic condition monitoring

Automated condition monitoring technology is an enabler to reduce unavailability costs caused by unplanned outages as well as certain non-optimal operating states. The goal is to develop and create a systematic approach how to utilise effectively multi-discipline learning methods in incipient fault/deviation detection and to pilot and create a service model where the technical findings are effectively and naturally communicated between the organisations.

Enhancing Power Plant simulator to support design and operation optimization

Maintaining of power plant simulator up-to-date is very important. Although principals of thermal processes haven’t changed much recently, variety of the new standards are published. These include components like boilers, steam turbines, cooling towers etc. Neglecting continuous development of optimization tool decreases value of the tool and accuracy of the decisions and optimization based on the tool results.

Results

The significant results has been gained in those four earlier mentioned areas:

Benefits

The industrial impacts for D2I came from successful implementation of the results in business models and the development of new business concepts. Some pilots provided results for future development projects and some more direct business impacts.

The asset lifecycle management service concept offers savings in cost and time during system implementation. The user also benefits from a well-planned system as it offers the best support for their maintenance related works.

The intelligent applications for operation support testing through cloud enable new usage areas and longer life cycle for operation support services. The improved services provide better availability and maintainability of user plants and components reducing production costs and improving competitiveness.

Development on the simulation tools resulted in increased flexibility for user to use any of the main standards in their calculations. The developed calculation algorithms and simulation components improve the design and optimization of O&M, fuel consumption, boiler processes and steam turbine calculations.

Future

Succesful piloted concepts and methods will be applied in future products and services. Some of the results need further development work. D2I results will be used as a input for Fimecc Step program and EU H2020 program "Cyber Physical System based Proactive Collaborative Maintenance" MANTIS program.

Screenshots for intelligent O&M concept:

Mobile UI for power plant maintenance
Lassie in web browser

Monitoring and control of industrial processes

Modeling

Problems

The problem area is found in modern process industry where the production processes are run by machinery and equipment that is monitored, guided and supported by sensor-based, intelligent systems. The systems collect data at intervals that can be decided and calibrated; these intervals are quite often short because this is made feasible by the technology and it is assumed that by collecting larger amounts of data there is a better probability of collecting data on anomalies, errors and malfunctions. This creates so-called big data problems.

One of the tasks addressed in the D2I program is to work on data representing the operational state of valves to identify so-called weak signals that are indications of anomalies, errors and malfunctions of the valves and to build support tools to guide decisions on predictive maintenance, i.e. to (i) carry out maintenance immediately, (ii) wait until the time for scheduled maintenance and carry out the necessary operations, or (iii) decide to do nothing, wait for a possible (or probable) failure and then carry out necessary operations (either immediately or at a scheduled maintenance time). The valves are in pipelines, the pipelines are in the process industry, there may be hundreds of valves in operation and the predictive maintenance problem was formulated by Metso Automation.

Results

There are three main states of a production process: (i) normal, the process is fully operational, there is no need for any type of intervention; (ii) failure, the process has to be shut down and the cause of the failure identified; (iii) gradual transition from normal operation to failure, the behaviour is still normal, or almost normal, but there are signs indicating a possible failure.

The data that is collected from the valve controllers consists of time series of n process diagnostic variables (if n > [100, 1000] we will easily get a big data problem). We decided on a two-stage process for failure prediction.

First choose a subset of the recorded diagnostic variables that contain the information that is essential for predicting a failure; the identified features are used to classify a state as critical or non-critical (i.e. to predict a failure).

Second, failure prediction can be done in real time or with a small delay.

The two-stage process needs to identify a subset of variables which has the most information on the state of the system; for this we found out that we should work on feature selection (extraction), not component extraction as this will add new constructs to the big data problem. This advised against using, for instance, Principal Component Analysis as this method decomposes variables into new constructs.

In the specific case we used feature selection to identify variables that represent a data set.

We used a test case as the basis and test environment for the development work. We had data on 32 valves and simulated performance of each valve for 120 days; in the data there were three simulated faults describe: (i) air leakage, (ii) faulty H-clip and (iii) friction; with 16 diagnostic variables that had been identified there were 61440 data points that were analysed.

The process worked out in the following way: we first identified the optimal feature subset for the three fault types separately; second, we evaluated the selected feature subsets by comparing their prediction performance to the model using all the 16 diagnostic variables, using a multinomial logistic regression analysis; third, we selected, implemented (using the R statistical software) and tested three different fuzzy entropy measures.

We found out that the predictive performance in general shows high accuracy with different varying performance depending on the fault types and that there is no single (fuzzy entropy) approach that performs better than the others but there is at least one method for every fault types with high accuracy.

The use of entropy was suggested by Dr Mats Friman of Metso Automation who had tested classical Shannon entropy on valve faults with very good results; the algorithms produced good accuracy and were shown to be very fast. We compared Dr Friman's results with the results we got with fuzzy entropy-based feature selection algorithms and found them to be about the same quality.

Fuzzy entropy has two benefits compared to classical entropy: (i) it is a measure which is not necessarily related to random experiments; (ii) it has been shown to discriminate the actual distribution of patterns better using membership functions instead of distributions. The standard definition of fuzzy entropy is:

where μA is the membership function, xj is the observed factor. The Metso Automation case and test data were used to work out the features of classical and fuzzy entropy and to evaluate the usefulness of entropy measures as predictors of faults in valves based on time series data. This proved that the approach will work and that the methods are fast, cost-effective and accurate as predictors.

C. Carlsson, M. Heikkilä & J. Mezei. Fuzzy Entropy Used for Predictive Analytics. Fuzzy Logic in Its 50th Year: New Developments, Directions and Challenges, Springer, pp. 187-209.

C. Carlsson, M. Heikkilä & J. Mezei. Fuzzy Entropy Used for Predictive Analytics. Proceedings of the 2015 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1-6, Istanbul, August 2015.

C. Carlsson & M. Heikkilä. Soft Computing in Handling Tacit Knowledge. Proceedings of the 2015 International Conference on Fuzzy Theory and Its Applications (iFUZZY), DVD 038-1158, pp. 124-129, Taiwan, November 2015.

M. Friman & H. Happonen. Taming the Big Data - Entropy Driven Statistical Models. Finnish Society of Automation, 2013.

M. Friman. A Method for Automatic Generation of Plant-Wide Inference Engines. ETFA'2014 - IEEE International Conference on Emerging Technology & Factory Automation, 16-19 Sep 2014, Barcelona, Spain.

The Metso Automation case was, secondly, used to develop methods for valve risk profiling and monitoring. This is to work out assessments of how long a device can be run at acceptable risk levels after data has been compiled and analysed that predicts that a fault may occur within an identified time interval. There were possibilistic Bayes models developed to support the judgment of risk levels and decisions on when to initiate predictive maintenance. The possibilistic Bayes models are part of what we now call predictive analytics and proved to be useful as decision support tools when the data for risk profiling is incomplete and when the monitoring produces data of not sufficiently good quality for classical Bayes modelling.

C. Carlsson, M. Heikkilä & J. Mezei. Possibilistic Bayes Modelling for Predictive Analytics. Proceedings of 15th IEEE International Symposium on Computational Intelligence and Informatics, Budapest, Nov 2014, pp. 15-20.

J. Mezer, R. Wikström & C. Carlsson. Aggregating Linguistic Expert Knowledge in Type-2 Fuzzy Ontologies. Applied Soft Computing, Vol 35, October 2015, pp. 911-920.

The Metso Automation case was, thirdly, used to develop methods for decision support that will support the process operators in the monitoring and control room. The approach is to find and identify similar cases for situations when something goes wrong in the production process; in terms of the Metso Automation case when a valve package breaks down and the affected part of the process needs to be closed down. The approach developed builds on the use of fuzzy ontology to describe valve (and pump) problems and to support problem solving by finding and identifying similar solved cases in large databases. We developed a working prototype in Protégé and worked out ways to update the ontology entities with type-2 fuzzy sets, which allows for imprecise data to be used without expanding the ontology (which is impossible to accomplish with classical ontology). The results show that fuzzy ontology is an effective way to build decision support for the professionals in monitoring and control rooms and could be developed as part of the maintenance program that is part of the delivery of advanced systems for the process industry.

R. Wikström. Fuzzy Ontology for Knowledge Mobilisation and Decision Support. Doctoral Thesis, IAMSR/Åbo Akademi University, 2014.

J. A. Morente-Molinera. Sistemas de Ayuda a la Toma de Decisiones Basados en Informacion Linguistica Di-fusa. Doctoral Thesis, University of Granada, 2015.

J. A. Morente-Molinera, R. Wikström, E. Viedma-Herrera & C. Carlsson. A linguistic mobile decision support system based on fuzzy ontology to facilitate knowledge mobilization. Decision Support Systems, 1/2016, pp. 66-75.

J. A. Morente-Molinera, J. Mezei, C. Carlsson and E. Viedma-Herrera. Improving supervised learning classification methods using multi-granular linguistic modelling and fuzzy entropy. To appear in IEEE Transactions on Fuzzy Systems.

Benefits

Predictive maintenance theory, models and tools will show the critical targets for maintenance work and the optimal time for carrying it out. This can be shown to save (i) costs for process breakdowns, (ii) costs for restarting processes and (iii) maintenance costs. These elements can be included in a maintenance package or program to be sold to customers; predictions on beginning process failures can initiate service actions or programs in optimal just-in-time fashion. In this way a maintenance program can add competitive edge to process technology deliveries. Fuzzy entropy based algorithms are effective tools for sufficiently good predictions.

The professionals in the monitoring and control rooms – or in the control rooms that support maintenance programs installed for customers – will benefit from the risk profiling and monitoring solutions that can give them estimates of the time remaining before predictive maintenance operations must be initiated. The possibilistic Bayes models allow them to decide on this even when the data is imprecise, has missing elements or outliers, or is distorted. This will speed up decision processes and safeguard against fatal mistakes.

The third approach is again targeting the professionals in the monitoring and control rooms, both in-house and monitoring maintenance programs for customers. The situation now is one when there has been some serious process problem, the causes need to be identified and problems need to be solved. The professionals taught us that the fastest way is to find out if there have been similar problems – quite often there has been – and how they were successfully solved; this as the first step to get the process back and up and running, after which there will be time to work out better and more sustainable solutions. The fuzzy ontology turned out to be a good approach to find similar cases, to select them from available databases and to present them to the professionals. Experiments show that the identification and selection process is fast, the precision is sufficient for the professionals who are good at quickly identifying valid similarities and the problem solving is fast and cost-effective; an added benefit with using a fuzzy ontology is that the quality of the ontology improves with repeated and frequent use.

Future

The methods, tools and technology – fuzzy entropy, possiblistic Bayes modelling and fuzzy ontology – that were developed with the help of the Metso Automation case and simulated database will be further developed and implemented for the other partners of the D2I Industry ecosystem in cooperation with Tieto.

References

C. Carlsson, M. Heikkilä & J. Mezei. Possibilistic Bayes Modelling for Predictive Analytics. Proceedings of 15th IEEE International Symposium on Computational Intelligence and Informatics, Budapest, Nov 2014, pp. 15-20.

J. Mezer, R. Wikström & C. Carlsson. Aggregating Linguistic Expert Knowledge in Type-2 Fuzzy Ontologies. Applied Soft Computing, Vol 35, October 2015, pp. 911-920.

C. Carlsson, M. Heikkilä & J. Mezei. Fuzzy Entropy Used for Predictive Analytics. Fuzzy Logic in Its 50th Year: New Developments, Directions and Challenges, Springer, pp. 187-209.

C. Carlsson, M. Heikkilä & J. Mezei. Fuzzy Entropy Used for Predictive Analytics. Proceedings of the 2015 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), pp. 1-6, Istanbul, August 2015.

C. Carlsson & M. Heikkilä. Soft Computing in Handling Tacit Knowledge. Proceedings of the 2015 International Conference on Fuzzy Theory and Its Applications (iFUZZY), DVD 038-1158, pp. 124-129, Taiwan, November 2015.

J. A. Morente-Molinera. Sistemas de Ayuda a la Toma de Decisiones Basados en Informacion Linguistica Di-fusa. Doctoral Thesis, University of Granada, 2015.

J. A. Morente-Molinera, R. Wikström, E. Viedma-Herrera & C. Carlsson. A linguistic mobile decision support system based on fuzzy ontology to facilitate knowledge mobilization. Decision Support Systems, 1/2016, pp. 66-75.

J. A. Morente-Molinera, J. Mezei, C. Carlsson and E. Viedma-Herrera. Improving supervised learning classification methods using multi-granular linguistic modelling and fuzzy entropy. To appear in IEEE Transactions on Fuzzy Systems.

M. Friman & H. Happonen. Taming the Big Data - Entropy Driven Statistical Models. Finnish Society of Automation, 2013.

M. Friman. A Method for Automatic Generation of Plant-Wide Inference Engines. ETFA'2014 - IEEE International Conference on Emerging Technology & Factory Automation, 16-19 Sep 2014, Barcelona, Spain.

R. Wikström. Fuzzy Ontology for Knowledge Mobilisation and Decision Support. Doctoral Thesis, IAMSR/Åbo Akademi University, 2014.

Screenshots and illustrations:

Mobile UI for power plant maintenance
Lassie in web browser

Remote predictive monitoring of lifting equipment

Lassie

Problem

Konecranes maintains and services a large global fleet of lifting equipment. A growing part of the fleet is connected to Konecranes remote services that support predictive condition monitoring and diagnostics, and the proper timing of field maintenance and service work.

The challenge is to develop methods for versatile and intelligent remote data analysis. There are needs to focus both on fleet level analysis as well as on individual equipment. To be efficient, the remote operations need to process and combine information from separate sources, and to select the best data combinations.

Solution

The methodologies for crane usage, performance and condition monitoring were developed further to support the equipment lifetime management and optimization of maintenance and service operations. The development was conducted in co-operation with Konecranes, VTT and Datactica.

Results and benefits

The work focused on five cases:

Failure identification of hoist gears

In general, gear faults are rare in machineries with slow rotation (e.g. cranes). In machines with high speed rotation gear faults are a significant cause of failures. The faults are mostly related to tooth faults: pitting, chipping, wearing and tooth cracking. Under normal conditions, contact fatigue is one of the most common failure modes for gear teeth. In lifting process gears are critical components: the required hoisting motion and power is transmitted through the gears. Gear measurement analysis and failure identification during the hoist lifetime is important from the reliability and maintenance point of view. Especially in larger units the down-time, spare part and repair costs can be substantial.

The aim of the studies was to develop and test analysis tools to produce features which can be used to identify mechanical gear failures at early stages leaving time to plan remediation and maintenance actions.

There are different methods available for gear condition monitoring and failure analysis. Typically the analysis is based on vibration acceleration measurements and the analysis is carried out in time or frequency domain, or in both. In gear vibration acceleration analysis, vibration amplitudes at the gear mesh frequency and its harmonics typically dominates the signal regardless of the actual gear condition. To reduce the effect of gear mesh originating from normal operational variations and to emphasize features indicating impending gear failures, special signal processing technique were applied for the analysis of spur and bevel gears. The focus in detection was on tooth pitting damages, initial faults and features which continues to increase as the damage spreads. The methods were tested during Konecranes gear testing where under different loading and driving conditions.

The applied method can be used for gear anomaly detection and for optimizing maintenance plans when incipient faults in operating gear teeth are detected. The methods also reduce sensitiveness arising from normal operational variances, and thus can be used to reduce the risk for false alarms. Since gear the vibration acceleration is in practise measured from the bearings supporting the gear shafts, it can be applied also on the bearing analysis.

Practical way to develop the analysis method principles further is to perform lifetime tests for the gears and their key components. The methods can then be tested and verified in field applications and iterated based on the experiences.

Poster: Gear Box Fault Detection

Future

The remote service concepts are further expanded to a growing fleet of equipment. The work is conducted in parallel with the build of new Industrial Internet services. The field data analysis is supported by new algorithms and methods that are able to focus the equipment on fleet level and that will develop to be intelligent and capable to detect abnormal behaviour of a single machine. This development will enhance the performance and productivity of the remote operations and services.