Empowering a citizen data scientist for hardware design & manufacturing

Empowering a citizen data scientist for hardware design & manufacturingPartha DekaBlockedUnblockFollowFollowingJan 25Improving productivity of a hardware design and manufacturing professional with an advanced AI toolAuthors: Partha Deka and Rohit MittalWhat is a citizen data scientist?Expert data scientists rely on custom coding to make sense out of data.

The use case could be data cleansing, data imputation, creating segments, finding patterns in the data, building a supervised model to predict a target, feature importance analysis for a classification/ regression problem, understanding the possible cause of an outcome, finding tunable parameters that can affect an outcome, even image preprocessing for a deep learning problem etc.

Citizen data scientist are power users of advanced analytics as discussed above.

But the trait of a citizen data scientist is different than an expert data scientist.

Their primary responsibility is outside the realm of statistics & analytics, they complement expert data scientists.

A citizen data scientist is more contextual, has unique perspective of a business area, focus more on a business problem with the application of an analytic technique, has more appetite for matters relative to business priorities & has the ability to justify a business value.

In a manufacturing organization, a hardware design engineer or a factory floor manufacturing engineer can also wear the hat of a citizen data scientist.

Optimizing a design process or improving throughput/yield in the factory floor could be their goals to bring more business value out of their processes.

Automate advanced analytics with an in-house built tool for hardware :We built an advanced analytics tool for hardware professionals.

By hardware we imply both semiconductors as well as board level systems.

with drag and drop features .

Our goal is to democratize analytics including AI & ML so that a citizen data scientist can focus more on the business value without spending time to program/code to perform advanced analytics.

Our tool is constantly evolving, some of the capabilities we have already built-in our tool:· Auto data preprocessing & data cleansing· Auto feature analysis including but not limited to:Ø Filter methods e.

g.

Pearson’s correlation, Anova, LDA, Chi-squareØ Recursive feature elimination· Principal component analysis (PCA) & understanding the impact of features → getting a sense of the segments in the data with respective feature over loadings· Dimensionality reduction with PCA· Various Clustering techniques for an unsupervised learning problem· Deep learning capabilities:Ø Image recognition: creating pixel features using popular CNN models with weights pre-trained on ImageNet images ( http://www.

image-net.

org/)Ø Automatically finding the best model for a classification problem with features from the previous step· Automatically finding the best supervised model (Classification / Regression) by training various models on the data.

This is a key feature enabling non-trained professionals to utilize ML for their problems.

· Additional features can be added for specific manufacturing needs.

For instance, fine tuning a model for false positives vs false negatives through hyper parameter tuning or ROC threshold.

Use case specific modelling which are not so common.

A modular flow below on ML capabilities:A modular flow below on deep learning capabilities:Example Searching pattern within the data set:Our “Analyze pattern/segment” module (in-built in our tool) provides automatic insights on the existing segments in the data relative to feature weights.

Our tool automatically perform principal component analysis on the data and auto render a plot with information variance explained by feature weights.

Additionally, it also provides smart auto recommendations on existing patterns / clusters in the data.

We found this capability extremely useful in finding patterns of product test failures by feature importance.

Below is a flow on pattern analysis with our tool:Conclusion:We have built an advanced analytic tool and constantly enhancing its capability.

Our goal is to democratize the usage of AI & ML within the hardware engineering / manufacturing community of our organization.

With our tool we have realized good ROI on many of our design & manufacturing use cases such as yield improvement, throughput optimization, hardware material defect detection, identifying tunable parameters for optimizing hardware design etc.

We are seeing increase adoption of our tool and also envision to bring in specific but generic reusable analytics that would suffice critical business use cases for various sub organizations within our organization.

References:https://scikit-learn.

org/stable/https://www.

tensorflow.

org/http://www.

image-net.

org/.

. More details

Leave a Reply