Machine Learning due diligence

The Real Requirements for Machine Learning PdM

With the proliferation of vendor solutions in the Industrial Analytics marketplace, it may be tempting to bypass the Business, Functional and Technical Requirements phase of the purchase process. The danger of diverging from standardized requirements and due diligence practices is that you become over-dependent on the information provided by the solution vendors.

The selection process for an IIoT Industrial Analytics solution is based on industry, facility and solution-specific requirements. Although no two requirements documents are identical, there are several functional and non-functional requirements common for most analytics and maintenance scenarios.

Below are key variables to consider when defining the requirements for an Industrial Analytics solution for Predictive Maintenance. Use this checklist as a starting point to develop your Requirements Document. For each variable, we provided a Basic or Baseline Functionality and an Optimal or Best Practice Functionality.

Please note that our evaluations are high-level, and individual vendor solutions will vary. We strongly recommend developing your own scorecard based on the list of Functional and Non-Functional requirements identified by the internal stakeholders of your industrial facility.

Practitioners’ Guide: What can go wrong?

Based on our experience at SKF Enlight AI, the following are the most common mistakes made during defining requirements for Industrial Analytics.

– Excluding key stakeholders in the process

– Failing to agree on a prioritized list of requirements

– Over-reliance on vendor pitches

– Divergence between views of production facility and headquarters

– Conflicting agendas of IT and OT

– Limited understanding of underlying Machine Learning data science

High Priority Functional Requirements for an IIoT Predictive Maintenance Solution

The following Functional Requirements need to be defined by stakeholders within your organization:

  • Interoperability / Open Architecture
  • Asset and Sensor Neutrality
  • Alert Generation
  • Machine Learning Methodology
  • Asset Visualization

#1 Interoperability / Open Architecture:  There is no standard or uniform IIoT infrastructure platform. The key consideration is whether the analytics solution works with multiple platforms or is a closed add-on to one platform.

#2 Asset and Sensor Neutrality: The key consideration is whether the solution functions in heterogeneous plant environments with data from all production assets. In some cases, the solution is tied to one class of sensors or processes.

#3 Alert Generation:  When a machine degradation or potential asset failure is detected, this is communicated to the relevant facility stakeholders.

#4 Machine Learning Methodology: Each Predictive Asset Maintenance solution is based on a Big Data methodology. Is this a manual process or is Artificial Intelligence used to automatically select the optimal algorithm for the specific scenario?

Practitioners’ Tip: The Importance of Machine Learning Methodology Selection

Why is the learning model important? Each specific Machine Learning Methodology approach (manual, Supervised, Unsupervised etc.,) requires differences in levels of internal staff time commitments and infrastructure investments. At the most extreme, manual statistical modelling is an offsite activity that is almost completely non-disruptive. At the other end of the spectrum, to deploy Supervised Machine Learning, an accurate virtual clone of the underlying asset is necessary. The cost differential between solutions based on different models is significant.

#5 Asset Visualization: At a facility level, technicians accessing the user-interface will not be trained in Artificial Intelligence and Big Data. The key considerations when defining this requirement are the visualization of machine behavior and the ability to depict the health of machinery or the entire facility, and take specific action as a result.

Practitioner’s Tip: Don’t Overlook Visualization

Visualization of data generated by Predictive Maintenance solutions is often an overlooked requirement. In terms of business impact, Visualization is one of the most critical components of Industrial Analytics for Predictive Maintenance solutions. There is a shortage of highly skilled Big Data Scientists and Big Data Engineers that is expected to worsen over the next decade. If analytics generated from a solution are actionable for end users, then the solution needs to provide an intuitive user interface and to summarize the data for plant and maintenance staff that lack formal training in the discipline of Big Data. Furthermore, the tool should be accessible at both a machine, plant and multi-facility level by stakeholders on the business, IT and operational areas of the organization.

High Priority Non-Functional Requirements for an IIoT Predictive Maintenance Solution

The following Non-Functional Requirement needs to be defined by stakeholders within your organization.


#6 Scalability: Analytics platforms must be applicable to a machine or facility of any size. The solution must be able to add assets without a need for any incremental investment in hardware, software or dedicated labor hours.

#7 Performance: The objective for an industrial analytics platform is to provide the production facility with accurate and timely data. Targeted performance measurements of the following will need to be defined:

  • Correct Alerts (True Positives)
  • False Alerts (False Positives)
  • Missed Failures (False Negatives)
  • Recall
  • Precision
  • F-score

Practitioner’s Tip: Which stakeholders should define requirements?

Predictive Asset Maintenance solutions are a critical component of a plant’s Industry 4.0 strategy. Based on our experience with some of the largest industrial companies, we recommend that you form a cross-functional team that includes the following:

– Plant Asset Maintenance

– Plant Engineering

– Plant Management

– Headquarters (IT, Finance)

With the Convergence of IT and OT, it is important to include stakeholders from IT. Even if they are not responsible for the business or functional requirements, access to accurate and updated data is a core element of any Industrial Analytics solution. By including representatives of IT in the Planning and Requirements phase, data related issues (such as availability, security and privacy) can be identified upfront and mitigated.  

Other Non-Functional Requirements

The following is a list of non-functional requirements. The specific details will need to be defined by internal stakeholders.

  • Response Time
  • Availability
  • Stability
  • Maintainability
  • Usability
SKF Enlight AI

Industrial plants generate terabytes of process data. SKF Enlight AI is a SaaS Predictive Maintenance solution that uses Automated Machine Learning to identify emerging asset failure patterns within this data. It provides early warnings and sensor-level intelligence to help avert unplanned downtime and meet production goals. For more information on how SKF Enlight AI can improve performance and reliability, click here.

Leave a Reply

Your email address will not be published.