Case Study: Welding Lines

Overview

The goal of this case study is to showcase PraxisAI’s work with an automotive supplier. We will walk through how PraxisAI was implemented and utilized by various groups of engineers responsible for plant operations.

Disclaimer: All screenshots are from demonstration and do not have any customer data.

Customer

A high-volume Tier-1 welding plant, operating 3 shifts to produce ~5,000 welded assemblies daily across 5 robotic lines. With 45 robots and modular fixtures, it supports over 50 part variants. The facility employs ~200 staff.

Problem

The supplier had large amounts of under utilized plant data. The team’s ability to drive efficiencies through AI was significantly hindered by the challenge of connecting siloed data systems, reducing the effectiveness of their analysis and slowing down potential cost-saving initiatives.

This resulted in three main operational issues

  1. Welding spot defect rate of ~0.6% leading to a FPY of ~92%. The teams goal was to achieve a 95% FPY.
  2. Highly manual production planning resulting in suboptimal plans.
  3. Delays in communication between procurement and planning teams leading to lines bottlenecked due to part shortages.

Data

Sensors

With approximately 100 sensors per machine, the fastest sensors recorded data at intervals as short as 0.2 seconds. Data was not being collected in a historian, however it was exposed through an OPC UA endpoint.

MES

An on-premise MES collected process data and synced with the cloud data lake approximately every 6 hours. This delayed data flow led to issues for cloud-based solutions, which struggled to adapt to changing conditions on the plant floor.

ERP

Cloud based ERP (S4/HANA) for supply chain data.

Solution

There were three main phases of the solution

  1. Contextualize siloed plant data (ERP, MES, sensors) stored in various on-premise and cloud systems.
  2. Applying AI securely with a hybrid deployment for defect prediction and material allocation.
  3. Working with process engineers to configure and deploy custom agents to respond to issues in real time.

To implement this on PraxisAI, we utilized various co-pilots that help with the complex engineering and change management required to deploy AI solutions in factories. These are built based on our founding team’s experience in deploying production grade applications, as well as using these applications on the factory floor.

  • Data: Build applications without having to rip and replace existing systems
    • Builds connections to integrate data
    • Understands source systems
    • Transforms and maps data correctly to the inputs of the required solution
  • Dashboards: Reduce complexities in collaboration between software and manufacturing teams
    • Understands requirements in the context of manufacturing
    • Outputs code in React
  • Tools: Reduce time to value to build custom solutions for common complex manufacturing problems
    • Builds machine learning models and complex algorithms
    • Auto ML capabilities with the ability to find the best performing model / algorithm
    • Explains model and algorithm outputs
  • Analysis: Enable your process engineers to run ad-hoc analyses on demand
    • Uses appropriate tool (custom models and algorithms) based on task
    • Performs advanced statistical analysis

Contextualize Data

Before any solutions were implemented, we first understood the source systems that we were working with. Praxis provides a FactoryOS, a pre built knowledge graph that helps our agents find relationships in the data and gather the correct data for a solution.

Below we can see users working with our co-pilot to build their FactoryOS.

We can analyze and visualize our data through natural language.

Tools

Engineers used various tools to optimize quality and production scheduling. Tools are complex algorithms and machine learning models provided by the Praxis team and based on extensive research. These tools are pre built for manufacturing use cases and have the flexibility to fit the different and changing requirements of plants across an enterprise.

Defect Prediction

With sensor data and process data contextualized, process engineers identified various defect types they want to predict. Data teams monitored auto ML training pipelines and made custom tweaks to the models when needed. The defect prediction tool outputs a time till defect and helped improve FPY to 95% by enabling smarter predictive maintenance.

Material Allocation

Procurement teams need to make intelligent allocations of material across plants based on requirements of upcoming orders. These decisions have a large direct impact on the revenue realized from production. Users created optimization engines based committed inventory from raw material suppliers and available inventory at other plants, as well as other constraints such as order priority, cost and due dates.

A custom agent was built to react to changes in inventory levels and committed supply and recommend optimal allocation changes.

Production Scheduling

The goal was to optimize changeover time between different product families scheduled on the line. Production analysts identified various custom requirements and built several engines for different scenarios.

A custom agent was built to react to live changes such as machine downtimes and recommended an optimal schedule based on the various optimization engines created.

Custom Agents

Through PraxisAI’s automations product, engineers defined custom agents for workflows they would like to automate. Engineers can upload various documents such as SOPs and maintenance manuals. Each agent takes in custom prompts and instructions from the user.

As mentioned in the last section, agents were built to react to live issues on lines to build new production schedules, or manage the tasks required to fulfill a part shortage. For example, to aid in defect prediction, users a predictive work order or alerted maintenance teams accordingly. Below, we can see the drag and drop UI/UX which enables process engineers to define these workflows.

Results

Over the course of 6 months, engineers observed the following

  1. Changeover time reduction of 14%
  2. Welding spot defect rate of ~0.4%, down from ~0.6%.
  3. Increase in FPY to 95%
  4. 12% decrease in late orders

For further details or more information on how these solutions can be used at your plant, please contact atewari@praxis-tech.ai.

Join Us Now!

Ready to Reduce Unplanned Downtime by 30%?

Leverage the latest AI tools to optimize your processes from procurement to assembly. Reduce bottlenecks and improve effeciency.