Artificial Intelligence in Product Lifecycle Management

Introduction

Artificial Intelligence (AI) seems to be all around us and seeping into every area of life and work, and, so it is no surprise that it is finding its way into Product Lifecycle Management (PLM) Tools. What I will cover here is a very brief introduction of AI before detailing some of the current capabilities of PLM systems. I’ll finish off with describing some of the more challenging engineering problems that AI could help engineers solve (thanks to Darren Cleal MCMI ChMC for his input to this section).

A Brief Introduction to AI

Yes, AI is a big topic. It seems to do everything from playing games, to reading and interpreting writing, to recognizing faces and aiding medical diagnostics. Each one of these is a field that a Google Search will provide tutorials, introduction and research fields. However, a starting point is to take a higher level view where AI can be described as below:

This show that AI encompasses Machine Learning (ML) and Deep Learning. It is ML that is most relevant to PLM systems and can be split into three learning methods, Supervised, Unsupervised and Reinforcement.

Briefly, Supervised is training the ML model using data with known output values. When an unknown value is entered the output is based on the learning from the training data. Unsupervised takes away the known output values and the training ‘looks’ for clusters of values. The model can either be told how many clusters to look for or determine for itself. The output for an unknown value is then which cluster does the trained model determines it fits best. Reinforcement, I know least about beyond the basics that it is based upon trial and error and rewarding the model when it does what you want.

Taking a closer look at traditional supervised classification models, the basic approach is to select a particular model algorithm and decide upon a metric i.e. what is ‘good’. There are many potential metrics with the obvious one, how often you predict the correct result i.e. accuracy. However, this has its limitations e.g. cancer diagnoses where 90% may look good but when you look at the numbers of False negative results it is not acceptable. A place to start on understanding metrics is https://scikit-learn.org/stable/modules/model_evaluation.html. Having decided upon an algorithm and the metrics another part of the process is to ‘tune’ the algorithm. Each model has a set of parameters that can be varied to improve the performance and there are available methods that can do the heavy lifting when given input ranges for parameters.

Just like all PLM systems high quality data is required when training an algorithm. For ML the amount of data, the balance of each output classification, amount of missing data and outliers are all considerations that require strategies to minimize, mitigate or remove.

Neural Networks (NN) are deep learning methods which are also used for classification problems. Although they can have 10’s or 100’s of thousands of nodes, each node can simply be thought of as summing the inputs (weighted) to give an output. This output is passed to the next layer and node(s) and is covered in the diagram below:

NN can also minimize error (improve performance) by back propagating the error from the output backwards through the entire network. Greater performance can be achieved, but their complexity leads to issues with explainability on why and how they are performing so well. There is some debate and research on making these models more explainable and hence more easily scrutinized for hidden bias.

PLM Incorporating AI

A look at two of the larger PLM Tools, namely, Teamcenter and 3D Experience showed where PLM is embracing AI.

Siemens themselves have developed two AI driven capabilities in Teamcenter:

Teamcenter Assistant

On the User Interface, it analyses the context of what you are doing and shows the commands and data it determines is most relevant. How it does this is through using Unsupervised Machine Learning Techniques. Simply put it ‘remembers’ (learns) what you do in certain situations and presents this when it determines you are following a similar sequence. It can also learn from your colleagues interactions with the system, and if you doing a task for first time it will present options based on their previous actions. This is especially useful for a new user, who now has available the combined knowledge of the team’s history of usage.

Separate from Siemens, Nordic PLM has also created an AI powered Teamcenter Chatbot:

•  https://www.plmnordic.com/plm-smartbot/

AI Classifier for objects

There are multiple drivers that lead to developing an automated, accurate classification system. It includes time saved in manually classifying an object but, also, in searching for existing objects. This leads to more re-use and less duplication. Can also be run as a batch job, in background or during off-peak hours

The classifier is Supervised Deep Learning Neural Network that is trained on a few objects, then it will automatically suggest classifications for the rest. Blind reliance on the classifier will cause prediction accuracy issues (e.g. as data changes over time or ‘bad’ data is used for retraining) which will further cause issues, rework and negate time savings. Therefore, a human is still in the loop as the Subject Matter Expert (SME) to review the suggested classifications and accept/reject or cleanse i.e. ensure ongoing system accuracy with human validation.

The use of AI and Machine Learning in their PLM products is not as obvious for Dassault Systemes. 3DExperience has capabilities through:

• EXALEAD has AI incorporated and it’s search could be used for PLM
• NETVIBES

However, I could not find any obvious capabilities embedded into ENOVIA

Wider view of where PLM could benefit from AI

Where could AI impact PLM to bring benefits? There does not appear to be a ‘must have’ concept emerging from either user base or the PLM companies.

The major strength of a PLM System is in the data it stores, and this provides opportunities if the compelling use case can be created. However, organisations need to stop focusing on controlling data (storing, versions, workflows) and extract ‘smartness’ from it and its interconnections.

Possible concepts or extensions to PLM in using AI

Searching

• Image search (e.g. Centric AI Image Search): instantly search PLM image archives to find the closest match. If a similar item has been approved or used before, a designer can find out immediately. If there isn’t an existing match in archives, can help find a supplier that provides a close match.
• Template search: input a number of data features and find exact or closest match. Then provide ‘where used’ so can see how it was used to provide re-use of design thinking.

User Interface/User Experience (UI/UX)

• This journey has started e.g. Teamcenter, but there is still more that could be done in e.g. following the hyperpersonnel recommender AI/ML models of Amazon.
• Incorporate the Business Process Models (BPMs) into the PLM system (not simply the workflows that support them). ML could then look at the sequence of individual user operations performed, show the applicable business processes, the current step and next steps. Ensure that central templates (common templates) are filled in correctly and supply links to relevant training.
• Identify a reduced set of common parts and present these to the user as options, as the user either searches for a part or inputs the attributes for a new part.

Supply chain

• Gather data on vendors based on quality of products received, on-time delivery and raw materials. Use automated AI models to learn (and track) each suppliers performance and predict potential issues and suggest alternate suppliers.

Sustainability

• Create an algorithm to calculate all of the possible permutations of component cost, logistics, and eventual part disposal to develop a weighted sustainability score.

Wider Value chain

• PLM systems often provide input to a data lake. Extend the concept of PLM as a collaboration hub to include incorporating a data lake with AI/ML algorithms to provide the insights. This would require that the additional data is assessed against metrics and corrected as required to ensure its quality and usefulness.
• Add data sources to PLM to include Production Monitoring and incorporate IoT devices.

Adoption and Change

• While it can be argued that engineers drive the future, they can be conservative in their habits leading to an adoption challenge. As part of a business and process transformation incorporate AI/ML algorithms that allow some learning from the end users based upon actual interactions e.g. abandoned processes, duration of steps and process workarounds.

Tackling Engineering Challenges with AI?

Below are some of the most common engineering challenges that face an organization as it moves to embrace new technologies and the benefits they can give.

• Managing Data over time
• An application and IT landscape that develops over time
• Processes – ensuring their relevance
• Commonality of parts across time and different programmes

Data Fragmentation over time

Over time (years), data is re-used and copied. It can appear in multiple databases used by different tools with no control or single master. It can have a lifetime measured in decades with defined traceability to individual data items.

Example – The same Aircraft Engine mounting bolt is certified once but used for different 1000’s engines by proxy i.e. engine 6 referred to Engine 1 bolt then Engine 66 (10 years later) refers to the bolt used in Engine 33.

If data is moved and then lose the tenuous links to/from certified version, then will need to recertify many other versions which can be a prohibitive cost.

Opportunity

Use experienced engineers to identify the original certified data. Train on this certified data. With ML search other databases to find matches:

• Create the links and cement them in digital thread, and
• Going forward for new designs / implementations supply the certified data to engineers at point of use (link to UI/UX)

Advantages

• Quickly find data for regulatory compliance e.g. crash investigation.
• As now have links, can move data and keep traceability.
• Improve Maintenance Repair Operations through (search & retrieve) only the necessary / relevant information.

Legacy application landscape

An organically grown mesh of procured and developed IT, point solutions and black box technology cost huge amounts to maintain and creates People centric and Application centric cultures.

Opportunity

A mix of datalake, Business Intelligence (BI) and ML.

1. Identify data source(s) to be used for training with outputs required and the associated inputs
2. Export all the databases into a data lake
3. Use trained ML to search in datalake for exact or near matches across different sources
4. Use BI to join dBs based on attributes in different sources e.g. procurement Id in PLM to find & extract all details from the Supply dB.
5. Combine all data columns from all searched dBs giving a comprehensive report of the search results (digital thread)

e.g. train on TC db of parts to classify certain parameters to object (bolt5, nut1, nut3, wash2). Then in datalake run search and cannot find exact match to search parameters but find things that ‘look’ similar (high probability – e.g. 90%) in CAD db but with different classification it’s a ‘nut1’ in TC but a ‘nuta’ in CAD. Note both are ‘nuts’ not one a bolt or washer.

Business Process

Many organisations work with MS Word or MS Excel to describe business process. These should be converted to Process Models. This requires Experienced Process modellers to interpret and challenge the existing content (which can be vague, duplicated, missing and/or conflicting).

Opportunity

• Use AI NLP to extract the relevant sections of documents / excel.
• Group processes into functional areas (predefined)
• Create visual processes that link the sections as per the written ‘joins’
• Indicate health of each process based on e.g. clear purpose, duplicate sections, sections with no inputs/outputs, missing conditions

Extra

Have examples of ‘best practise’ processes that can be compared to the converted processes

Identify process gaps i.e. functional areas with none (or few) processes, areas that may require business processes (compared with ‘best practise’)

Common Parts across programmes

The same requirements are often fulfilled by different parts in different programmes. Also, same need may be met by different solutions.

Example from car manufacturer: for one vehicle model, they had: 4 Diesel engine options and 4 Petrol.

Across the 8 engines they used 3 different types of bolts to link the Cylinder Head to the Engine block. Each Bolt type required different torque settings.

Resulting in additional tooling to be able to handle the different bolt types, additional programming to manage the different torque types, with an increase in testing needs.

Opportunity

AI can be used in different ways, first:

• AI NLP to compare requirements on different programmes to find exact or near matches. Then find the parts linked to these requirements (in both programmes)
• User select same functional part of BOM in two different programmes

Then use ML trained on one programme (parts) to search for each part in both programme (or near match) to show the commonality between and differences

Engineer can then decide if there can be re-use / commonality or there are other factors that require the differences (make an informed choice).

Possible Future of PLM is it PLI

Kalypso advocates that Product Lifecycle Intelligence i.e. using the PLM data to provide AI powered change as the evolution of PLM. PLM is not the end point of the journey to improve efficiencies, drive innovations and improve time to market.

In joining more parts of the organization digitally there may not be a single source digital truth but many data masters across the organization. In this case, an enhanced ability to integrate and share data with more and different systems and data models combined with PLI could be the future of PLM.

Conclusions

AI is moving into PLM and it’s a trend that will only continue. There are some obvious places where it can expand, for example, improving UI and Searching and helping in Adoption and Change. There are also some interesting real world engineering challenges that it could be targeted towards in Data, Process and Commonality where it could play a key role in simplifying or solving.

In summary, AI is beginning to realise its potential in PLM but it has not yet eliminated the need for ‘Real’ Intelligence and so our roles are not going to be replaced, but it will augment our capabilities, turning engineers into a form of digital cyborgs.

Resources

1. Machine Learning for Dummies IBM Edition (2018)

https://www.ibm.com/downloads/cas/GB8ZMQZ3

2. Teamcenter Assistant, an Artificial Intelligence Application (Oct 28, 2020)

https://blogs.sw.siemens.com/teamcenter/teamcenter-assistant-an-artificial-intelligence-application/

3. PLM after 20 years – What is the position of Siemens Digital Industries Software today? (Nov 12, 2021)

https://www.industrie-digitalisierung.com/en/plm-after-20-years-what-is-the-position-of-siemens-digital-industries-software-today/

4. You’re in Luck – AI Classification and PLM Helps You Work Smarter! (Jan 16, 2021)

https://blogs.sw.siemens.com/teamcenter/youre-in-luck-ai-classification-and-plm-helps-you-work-smarter/

5. Classification AI ‘makes the system work for you’ (Nov 11, 2021)

https://blogs.sw.siemens.com/teamcenter/classification-ai-makes-the-system-work-for-you/

6. PLM SmartBot – A Teamcenter-Specific Chatbot

https://www.plmnordic.com/plm-smartbot/

7. Artificial Intelligence and Machine Learning at Dassault Systèmes – Part 7/7 (Jun 4, 2019)

https://blogs.3ds.com/netvibes/2019/06/04/artificial-intelligence-and-machine-learning-at-dassault-systemes-part-7-7/

8. Are Artificial Intelligence and Machine Learning catching eye words for PLM Projects? (Jan 15, 2022)

https://beyondplm.com/2022/01/15/are-artificial-intelligence-and-machine-learning-catching-eye-words-for-plm-projects/

9. Centric AI Image Search

https://www.centricsoftware.com/artificial-intelligence/

10. How PLM Can Learn from Machine Learning (Aug 8 2017)

https://www.whichplm.com/plm-can-learn-machine-learning/

11. Do you have data you can trust? Preparing for the AI world (Jun 2, 2020)

https://www.whichplm.com/do-you-have-data-you-can-trust-preparing-for-the-ai-world/

12. When PLM Met Machine Learning: The Beginning of a Great Relationship (Mar 8, 2019)

https://www.aberdeen.com/blog-posts/blog-when-plm-machine-learning-meet/

13. Actionable Insights for the Pharmaceutical & Biotech Industry (2019)

https://vertassets.blob.core.windows.net/download/e5a082f5/e5a082f5-c507-41a0-af9f-4760a58ca102/plm_analytics_for_ls.pdf

14. Exclusive Interview with Drew Cekada, Senior Manager- Digital Transformation Consulting, Kalypso (Jun 17, 2022)

https://www.analyticsinsight.net/exclusive-interview-with-drew-cekada-senior-manager-digital-transformation-consulting-kalypso/

15. PLM's Analytics Imperative: Drive Innovation Improvements with Actionable Insights

https://kalypso.com/pliinfographic