8 Conclusion

This chapter will provide a conclusion to the research and summarize the extend to which the research objectives were achieved. In addition, the results from the statistical analysis of cement mills and fans will be summarized, detailing the implications that the results have on equipment maintenance. Finally, a post-analysis discussion will provide identify suggestions for improvement of the methodology and recommendations for future work.

8.1 Research objectives

In summary, the main objective of this research was to develop an integrated predictive model, incorporating failure event records, production output records, and vibration observations, that can be used to predict the reliability and behavior of the mechanical equipment of the plant. In order to achieve this objective, several specific objectives were first completed.

First, each data source was aggregated, pre-processed, and structured in order to make them suitable for statistical analysis. The second specific objective was to perform a descriptive analysis on the prepared data sources in order to identify important characteristics regarding the scope of the data collected, and to extract preliminary insights. Additionally, the data was used to perform a criticality analysis which identified the cement mills and fans as critical equipment within the plant. The third objective was to build and demonstrate the use of reliability models for the critical equipment. The finally objective involved an integration of the failure event records, the production output records, and the condition monitoring(vibration measurements) records for the purpose of building an integrated predictive model.

As the research successfully accomplished all specific objectives, the primary objective was also accomplished, and several integrated predictive models were generated. The stoppage event records and the monthly production records were integrated and used to build extended Cox and accelerated failure time reliability models for 4 cement mills in plant section 6. A complete integration of all three data sources was performed and used to estimate an extended Cox model for 10 fans and an accelerated failure time model for 4 fans. Furthermore, the integrated fan data was used to build artificial neural network and logistic regression models for the purpose of predicting future maintenance actions from the observed event, production, and vibration measurements.

8.2 Analysis results

As the integrated predictive models estimated the effect of equipment-specific variables on the reliability of the equipment, the results of these models, and respective inference, can be used to derive future maintenance decision support. Based on the integrated data for cement mills, the extended Cox model(semi-parametric) and accelerated failure time model(fully parametric) both indicate that an increase in production rate corresponds to an increase in the risk of failure, or decrease in reliability. Additionally, both models indicate that replacing a component of the cement mill in response to a failure significantly increases reliability immediately after, as compared to simply repairing the cement mill. Finally, the accelerated failure time model also indicated that performing additional maintenance interventions, before failure occurs, significantly increases the reliability of the cement mill.

Based on the integrated data sources for fans throughout the plant, the extended Cox model provided the same conclusion regarding production rate(an increase in production yields a decrease in reliability) and equipment replacement(replacing a component increases reliability more than repair). However, this model also indicated that each time a fan must be stopped, for reasons other than failure or maintenance, the reliability significantly decreases once it resumes operation. As the accelerated failure time model was estimated using only 4 of the fans from the same section, it identifies several different significant effects. This model suggests that as the duration of the repair time increases, the reliability of the fan immediately following the repair is decreased. Furthermore, it indicates that as the observed vibrations(ANDE and HNDE.ENV) increase, the reliability of the fan decreases.

Lastly, the ANN and logistic regression classification models managed to provide more accurate predictions of future action as the complexity of the models decreased. When predicting the future maintenance action from covariates, the model behaved poorly. However, when reducing the output to a classification replace or non-replace the model prediction greatly improved.

In conclusion, these predictive models provide a clear indication of the value of using data integration to identify changes in equipment reliability. In addition to tracking and accounting for historical behavior and performance, monitoring the condition of an equipment through vibrations can provide a key indicator of the health of equipment, which can be used to motivate maintenance interventions.

8.3 Post-analysis discussion

This section will serve to briefly summarize some points of discussion regarding the research, including suggestions for future work, improved methodology, and alternative analyses.

As the source data for this research is extensive, yet largely unstructured, this research demonstrates the need for structure and standardization when making and recording observations. Additionally, it demonstrates the benefit of having compatibility between data sources based upon a standardized reference. Although text mining was used to classify a failure mechanism and maintenance action for all failure events, there is tremendous value in documenting this information in real-time according to engineering standards. Additionally, tracking and classifying the degree of repair(e.g., complete, partial, etc.) or specific component replaced will allow for improved estimation of reliability, and more specific maintenance actions.

The use of vibration measurements to predict equipment reliability demonstrates the necessity for more frequent, or continuous, monitoring of equipment. Given the irregular observation intervals and need for imputation, it is likely that the model estimation would greatly improve with more frequent observations. Continuous vibration monitoring would allow maintenance engineering to precisely identify the health status of an equipment and allow for instantaneous maintenance intervention if deemed necessary.

Additionally, when integrating the data sources and accounting for both the occurrence of non-failure stoppages(Numstop) in addition to maintenance interventions(Maint.), the use of a cumulative sum makes an assumption that the effect of both types of events has an additive relationship. In terms of non-failure stoppages, this assumption may be correct, as repeated stoppages may cause an accumulation of wear, which might be well represented by a cumulative sum. In contrast, the effect of planned maintenance, in theory, should promote an immediate increase in reliability, which would eventually diminish with time. In this regard, the effect of repeated planned maintenance interventions may be better represented by a different function. An alternative to a cumulative sum could be a measure of the time since last maintenance, which could receive a slight reduction following each maintenance intervention, while still accounting for the effect of time. This reduction in time could also vary depending upon the type or severity of the maintenance intervention.

Although the plant has experienced a large number of mechanical and electrical failures, the majority of the stoppage events records have not been thoroughly studied. Although the cumulative number of non-failure stoppages was accounted for in the data integration, the number of different stoppage categories could be particularly useful in distinguishing between intermittent and extended failures, as discussed in the literature review. It could be useful to account for these different non-failure stoppages in reliability models, or even to study the occurrence of these events themselves using intensity or count models.