What is Automated Anomaly Detection?
Automated Anomaly Detection (AAD) is broadly monitoring assets in a unit or process using process variables such as temperature, pressure, flow, and vibration in computer models that continuously evaluate real-time performance against past performance and/or design expectations. When the difference in current performance versus predicted performance is statistically significant an anomaly is detected, and an alert is generated. This leverages the comparative advantage of computers to perform continuous repetitive analysis on large amounts of data to detect and report anomalies freeing up humans to use their comparative advantage in prioritization and complex problem solving to profitably address the reported anomalies.
AAD is part of an Advanced Monitoring capability that continuously evaluates and implements the most profitable monitoring strategies.
Why Implement Automated Anomaly Detection?
If a picture is worth a thousand words, the P-F curve provides the best picture of AAD benefits. AAD pushes the point of detecting failures as far back on the P-F curve as presently possible. As a result of finding potential failures earlier, fixes are smaller (e.g. process adjustment to improve NPSH vs pump seal replacement), larger repairs can be planned and executed at the most opportune time, and asset/unit performance is maintained at or near capacity a greater percentage of the time. The end result is a more reliable and profitable operation.
How is Automated Anomaly Detection Implemented?
Implementing AAD is a four-step process that takes approximately one month to complete. The first step is configuration. During this step, process data is backfilled into the AAD software to make it available. In parallel, the physical assets are placed into a facility/system/asset hierarchy within the AAD software. During the second step, models are constructed. Standard templates are used to quickly deploy a large quantity of models. Additionally, custom, and proprietary performance models are deployed. The third step is a baseline review to maintain the initial models. Although the models will be continuously trained over time, this initial maintenance step is a more concentrated effort to minimize false positives and maximize finds. Finally, stakeholders are trained in best practices that allow them to leverage the information in the AAD software. Not to be mistaken with the Digital Twin approach, the MFG Analytic Model Based approach quickly delivers bottom line results to most of the equipment and processes.
How to Realize the Value of Automated Anomaly Detection
The value of AAD can only be realized when the detected anomaly creates an action that either prevents a failure or helps prepare for an oncoming failure. The difficult part of implementing an Advanced Monitoring capability for organizations is twofold; trusting the model and anomaly are valid and acting on the anomaly once detected.
For the past several decades failure mitigation has been focused on detecting an active failure mode. One example of this is periodic manual vibration monitoring of rotating equipment. This is a predictive action that looks for vibration levels usually monthly that are approaching or above a set alarm point indicating the machine has entered a failure mode. After the failure mode has been detected the machine is scheduled for an outage to correct the failing component before a total failure occurs. Comparing this to Advanced Monitoring, a sensor (wired or wireless) is connected to the machine and vibration feeds the computer model which may look at several variables. When the real time variables produce an “unexpected” response from the model an anomaly is created. If the anomaly is validated, it is sent to be investigated and potentially corrected before the defect grows into a failure.
The process seems straight forward however there are several obstacles that need to be overcome to get value out of AAD. The first obstacle is the notion that each anomaly must be academically challenged to assure it is valid. The second obstacle is the thought that the anomaly has not yet reached a limit or threshold and there is time before this happens. The third obstacle is it takes new skills to effectively diagnose anomalies which must be learned and practiced. Let’s take a deeper look at each of these obstacles.
The nature of building a model that will detect a valid anomaly involves both science and a bit of form fitting. Since there is variability in the early developed model outputs, each output is challenged and retrained a few iterations until the model is stable. Once the model is validated through feedback and retraining, it is production ready; however, the organization may still view it as in training. If this viewpoint is not overcome, anomalies will become an academic exercise to prove or disprove validity by the organization and value will be left behind.
When vibration predictive technology entered in the early 80’s, there was skepticism. Until this point people used their senses to estimate the condition of the machine, however most machines ran to failure. When vibration technicians recommended shutting down a machine due to vibration readings there was skepticism as well. Over time the technology earned credibility and now it is a proven technology. As in the early days of vibration monitoring, we are in a similar place with AAD. The technology will earn credibility over time, however early adapters can get early benefits. Partnering with an experienced AAD company and treating the validated model as an accurate and not debating it can shortcut the proving phase and lead to earlier value. There will be times when the model reports a false condition however this can be overcome, and learnings captured with the benefit of acting on anomalies outweighing the cost of waiting on perfection.
As in the vibration monitoring example, technicians had to learn what a failing bearing in the early stages looked like. There were instances where a machine was disassembled, and the report back was “the bearings were perfect.” Over time knowledge and skills grew, bearings were cut open and observed under high power microscopes showing defects that were early in manifestation. The newly observed defects were then analyzed to gain a deeper knowledge of how bearings fail. These learnings led to better practices in bearing handling, installation, lubrication, temperature limits, etc. Now we are at a similar place as vibration technology in the 80’s, we need to invest in diagnosing and correcting Advanced Monitoring anomalies. When an anomaly is validated, it should flow to the work execution capability for diagnosis and mitigation. The work execution capability should understand the Advanced Monitoring capability and vision so they can respond accordingly. Craftsmen typically do not diagnose issues in the failure mechanism territory of the P-F curve so the skills required to diagnose and mitigate failure mechanisms will develop over time.
The MFG Analytic / Atonix Advantage – A Holistic Approach
Successful implementation of Advanced Anomaly Detection and continuous monitoring/improvement can be thought of in two pieces; VisiblePlant software powered by AtonixOI continuously running data driven predictive models identifying anomalies and the site resolving the anomalies. Not to be mistaken with a Digital Twin approach, the MFG Analytic/ AtonixOI Model Based approach quickly delivers bottom line results to most equipment and processes (80/20 rule) at a reasonable cost.
Customers can expect models running and validated anomalies communicated within one month of data acquisition.
Atonix Digital brings 30 years of modeling software experience along with over 1500 model templates for rapid deployment. MFG Analytic brings greater than 100 years of combined industry experience in plant operations, maintenance, processes, and AAD implementation. The combination of these companies brings years of experience and efficiency in implementing Advanced Monitoring. This unique combination of skills and experience allows our customers to expect models running and validated anomalies communicated within one month of data acquisition. The MFG Analytic and client partnership continues past AAD implementation to the execution phase. With AAD implementation experience, MFG Analytic provides recommendations and answers to the obstacles and challenges presented during implementation. The vision is to have AAD deliver value as soon as possible requiring both the anomaly detection and a profitable response.
MFG Analytic monitoring and diagnostic service will monitor, diagnose, and communicate anomalies short-term or long-term depending upon customer needs. MFG Analytic has a library of over 50 Knowledge Maps used as model input and to help diagnose anomalies. With feedback from anomaly resolution, these Knowledge Maps are continuously updated with new learnings and knowledge for future reference.
Once AAD is implemented, MFG Analytic monitors and processes anomalies from VisiblePlant software. If the anomaly presents sufficient risk to the plant, it will immediately be proposed as a Top Business Risk via the Top Business Risk app. Anomalies are reviewed at weekly Client/MFG Analytic anomaly communication meetings where new anomalies are presented and an initial RCA originated with next steps defined (referring to Knowledge Maps for assistance). Unresolved anomalies are also reviewed in each meeting leaving no anomaly behind. If diagnostic capability is required past the client’s ability, MFG Analytic can provide diagnostic support to assist in anomaly resolution. Once resolved, the anomaly is closed, the RCA is finalized, and the appropriate Knowledge Map is updated to assist with future issues.
Bob Gleichman & Dale Whittenberg
Both Bob & Dale are Senior Partners at MFG Analytic.