Predictive and proactive maintenance for critical industrial equipment and processes can help to reduce unplanned downtimes as well as to improve process availability and stability. Artificial Intelligence (AI) methods for process data analysis can provide plant operators with much needed insights for decision support and predictive plant maintenance strategies. Let me bring you more insight with a concrete project.

The project aimed to create an AI-based solution for predictive maintenance of a rock grinding mill using SiePA – our application for equipment predictive analytics. SiePA (Siemens Predictive Analytics) represents a software solution for model training, based on the integration of machine learning with domain expertise. It is designed for support of prealerting/predicting critical equipment conditions and identification of issues based on natural language processing. SiePA also provides an integrated dashboard, focused on user interaction support and data visualization, enabling users to monitor the equipment sensors grouped by subsystems.

Anomalies as a basis for process optimization

By using SiePA, the project intends to assist the plant operator with anticipating failures at several points within the process, based on correlations among sensors. Many failures of process-critical equipment are preceded by anomalies in process data. SiePA is specifically designed to identify such anomalies. Such diagnostics is accordingly enabled by stemming predictive models from historical data, accomplished through advanced machine learning algorithms. Assessment of potentially large number of anomalies, based on numerous variables in common processes, calls for the involvement of domain experts. We have therefore reinforced the SiePA data analysis tool with domain expertise and consulting services to facilitate the definition of anomalies in a given process. Moreover, focusing anomaly detection on those process variables that are specifically relevant to equipment or process performance reduces the complexity whilst increasing usability in many applications.

Calculating correlations and creating models

Based on historical data, the SiePA system automatically calculates how sensors correlate with each other to define a starting point for the modelling procedure. Once the correlations have been established, results are evaluated and enhanced with the support of a domain expert. This is done with an aim to integrate domain knowledge into the modelling procedure, thus creating high-quality models that will be able to detect potential risks better than merely relying on the correlations alone.

Training models

Since SiePA’s purpose is to inform the user whenever a target sensor is behaving abnormally, we must identify what the sensor’s normal behavior is in the first place. The system therefore needs to be “educated” with historic operational data from “healthy” periods of equipment operation, so to learn the likes of normal behaviour.

Dashboard and alerts

Once the system has been set up, the operators are able to monitor its subsystems on the dashboard. Should an anomaly be detected, an alert is issued by the system. The alerts are categorized as low- or high-risk. In general, the system provides an overview of all monitored equipment along with a modern dashboard, showing a detailed status of relevant information.

Implementation

The methods and tools were applied in the specific case of application to the process for grinding phosphorous rocks. Our data scientists and in-house domain experts performed the analysis of processes and assets in the mill—mined rock is ground before sorting and further processing. The product is then transported from the mill to the sorter in a gas stream that is recirculated into the milling process. By using six months of acquired data, our data scientists and in-house domain experts analyzed the mill processes and assets thanks to SiePA. Considering that the plant was a typical brownfield installation, this study also required matching tags to the positions of probes in the processes, as well as identifying missing data and ensuring consistency in data naming and values. Accordingly, the SiePA analysis was then applied with the goal of detecting and predicting several failures within the milling process.

Enabling preventive maintenance for mill rollers

The analysis demonstrated that SiePA can, for example, successfully be used as a tool to detect gradual equipment wear through modelling; in this case to identify deterioration of the rollers and thus distinguish it from low quality of raw materials. We demonstrated that SiePA can detect slow changes in the behavior of mill rollers and differentiate normal/learned variations from anomalous changes on a multivariate basis. By detecting the increasing deviation of the target sensor, this analysis helps operators keep track of changes in the equipment and supports them in adjusting settings and planning maintenance.

Summary and future work

Following the training on Process data, the SiePA solution showed capacity to generate early warnings of sensor failures or equipment deterioration, as well as detect slow changes to the processes in a system of multiple correlated sensors. These early warnings and the app’s ability to connect with different maintenance, repair, and overhaul (MRO) systems, can be effectively used as a trigger for specific preventive maintenance workflows.

With the upcoming V3.0, SiePA can make predictions of failures based also on analyzing raw vibration data. The integration of modelling templates for different equipment types brings our offering to the next level, guiding the customer through a domain-expert driven workflow.

High-quality demands, cost pressures, continuous specification changes, and final usage requirements for chemical production was the trigger for the execution of SiePA for batch processes. The first global release focuses on batch production transparency, optimal (golden) batch identification, and real-time detection of anomalies powered by AI.

As technology evolves, new opportunities arise to improve plant strategy, production, and overall system profitability for our customers.

Find out more about Siemens Predictive Analytics (SiePA):

And get in touch with us! At: https://new.siemens.com/global/en/markets/chemical-industry.html