Analytic behind Asset Advisor - ED?

[Sireethorn_M]

Hello All,

I just got questions from customer while we've done POC with them on our EAA-ED where they would like to understand more detail about AI/ML that we use for notifying before failure happens or suggest condition-based maintenance. Below is the list of questions so would need all of your suggestion here.

1.  Predictive vs. Condition-Based Maintenance: Could you clarify how your software supports predictive maintenance compared to condition-based maintenance?
2.  Data Models and Analytics: What data models and algorithms are used for predictive maintenance? How is historical data processed for predictions?
3.  Factors and Parameters: What factors and parameters does the software use for analyzing asset health and predicting failures?
4.  Customization and Accuracy: Are the predictive models customizable? What is the accuracy rate of the predictions, and how is it validated?
5.  Threshold Adjustments: Can the thresholds for predictive alerts and notifications be adjusted to fit specific asset requirements or operational conditions?

Note: For offer level that we propose is ESP-Prime (ACB + MV SWG + Dry TR)

Thank you in advance for your comments and support.

[Victor_Marina]

About Question #1: Could you clarify how your software supports predictive maintenance compared to condition-based maintenance?

EAA-ED application provides Maintenance indexes to keep under control over the next maintenance date, which had been initially set with an extension period agreed in the service contract.

So, it is relevant to know last maintennace date (or commissioning date if no maintenance executed yet) and next planned maintenance date. In the meanwhile of those 2 dates, application runs timely assemment of the aging of any revant compnent of the connected equipment, and that assessment is converted into an  estimeated remaining useful lifetime (RUL) of the component. 

In case a single component's RUL is shorter than time to Next planned maintenance date, EAA application update index value, which in turn, is scored depending on the RUL left; and from that update some actions are lead by our CSH experts to reschedule the intervention.
In case RUL goes beyond Next planned maintenance date, index stay in low/safe values, and no additional action is required, because Next mainenace date is properly scheduled.

In summary, our service (including application, CSH support on remote and FSR onsite interventions) does not deliver pure predictive maintenance, leaving assets without planned date till symptoms of failure come up; but secures an extended maintenance schedule, while keeping online assessment of asset condition to anticipate upcoming failures at component level.

[Victor_Marina]

About Question #2 and #3: Data Models and Analytics, Factors and Parameters

2. What data models and algorithms are used for predictive maintenance? How is historical data processed for predictions?

Refer to Reply to Question #1 to understand properly our position around pure predictive maintenance, as our proposed approach for maintenance strategy.
a) We run propietary models, to breakdown the assets on their main components that are at the initiating causes of failure, based on manufacturer experience

b) Then, our laws and rules to compute the aging are asset specific (at the level of range and model in most of the cases) and based on design parameters, quality issues on the product or similar products, and services guidelines from past experience

c) Finally, we run our analysis daily, taking into account up to 11 months of data measurememts, never less than 2 months, to run analytics behind each asset.

3. What factors and parameters does the software use for analyzing asset health and predicting failures?

Each asset type has their own data model, including

  - some parameters or properties for the range/family/model (i.e type of mechanism, type of grease...) and

  - required measurements that should flow from the device to our application in the cloud, including electrical data (load, power factors, power quality... ), mechanical (number of operations, number of magnetizations...),

  - measurements from specializzed devices or sensors: thermal points (absolute temperatures from specific locations where sensors get fitted), insulation measurements (either for bushing or cables), measuments from oil in transformers (water ccontent, H2 content, temperature...),

  - environmental data
We cannot disclose the full list in our documentation, however, during set up phase, customer may have access to libraries deployed in the gateway to retrieve data, through the gateway interface. 

[Victor_Marina]

About Questions #4 and #5: Customization and Accuracy, Threshold Adjustments

4. Are the predictive models customizable? What is the accuracy rate of the predictions, and how is it validated?

Our models are specific for asset types, and reusable. They are not customizable, but there is always room to evolve the standard model, indeed we have internal releases for advanced analytics.

About the data available at asset level, unfortunately we cannot guarantee all parameters would be measured, as they depend on the electronic devices fitted into equipement, their connectivity and internal features. As a consequence, in some assets we do not break it down on compnents that does not have required data inputs, reducing model accuracy.

In any case, we validate our models with product line of business, who run initial test (non destructive) to validate model in the lab, and along the history of our service, with more than 10.000 connected assets, a bunch of them for years.

5. Can the thresholds for predictive alerts and notifications be adjusted to fit specific asset requirements or operational conditions?

Yes, Thresholds are customizable, while they are not part of the advanced analytics described above, they are a mechanism to trigger a ticket on any indicator that varies along time surpasing the threshold value.

There are 6 (potential) thresholds to define for an indicator: upper and lower thresholds, and in both cases 3 levels: information, warning and alarm. Each one would trigger a ticket with corresponding severity level (low, medium, high). Once they are set, they get displayed on the widget whre the were set.

In addition, it can be defined thresholds for the trend of each indicator.
Currently, many thresholds are set by default during initial configuration, while they are adjustable, and histrory of updates is also recorded.