AI in Manufacturing - Improving Process using Prescriptive Analytics

With the rise of Industry 4.0, computation power, data warehousing and automation, factories have been increasingly becoming intelligent. Preventive maintenance of Machines and predicting the failures have become an increasingly common sight. AI has also empowered in planning and logistics, where the quantity of item to be manufactured and the timing of it, have been decided through the outputs of ML models. Now the manufacturers are increasingly focused on improving the quality of the process and the throughput through sustainable methods as rising global warming is a concern. To improve the efficiency and to make the process sustainable, Machine Learning models coupled with optimization are used for Prescriptive Analytics. Data of the industrial process is often huge data with many process and control variables involved. Understanding the variables requires domain knowledge expertise coupled with feature engineering techniques. A search-based optimization can be used for finding the Pareto optimal solution with objectives to maximize the KPI and finding the support in historical data. Identifying the interaction effects is done by learning the data through a prediction model. The performance after the process is predicted using modelling for the KPI. Sensitivity analysis was conducted to understand the effect of variables on the uncertainty of model output and the KPI. The process, then optimized for maximizing throughput provides prescriptive analytics thereby improving the performance and reducing energy consumption.

 
 

Outline/Structure of the Talk

  • Problems in Industrial Analytics (2 mins)
  • Approaches (5 mins)
    • Feature Engineering
    • Bayesian Inference
  • Modelling Approaches ( 3 mins)
  • Sensitivity Analysis ( 3 mins)
  • Optimization - Prescriptive Analytics ( 5mins)

Learning Outcome

  • Methods to deal with high-frequency Industrial data
    • For Feature Engineering
    • For modelling
    • Bayesian Inference
  • Using Model-based Optimization for Prescriptive Analytics

Target Audience

Data Engineers, Data Scientists, Machine Learning Developers,Industrial Engineers

Prerequisites for Attendees

  • Basic Industrial experience
  • Basic Analytics experience

schedule Submitted 5 months ago

Public Feedback

comment Suggest improvements to the Author
  • Ravi Balasubramanian
    By Ravi Balasubramanian  ~  5 months ago
    reply Reply

    Hello Aravind,

    Thanks for your proposal. Can you please explain which specific industrial use case you would be presenting to illustrate the topics mentioned in your talk? Walking through one use case would be helpful for the broader audience.

     

    Thanks 

    • Aravind Kondamudi
      By Aravind Kondamudi  ~  5 months ago
      reply Reply

      Hi Ravi,
      Thanks for your comment. We would be presenting real industry optimization scenarios used for quality improvement, throughput improvement and energy optimization however, we cannot provide the exact case as it is confidential. We will be explaining the end-to-end scenario, explain the differences in predictive and prescriptive analytics specifically for Industrial analytics.

      • Ravi Balasubramanian
        By Ravi Balasubramanian  ~  5 months ago
        reply Reply

        Hi Aravind,

        Thanks for the response. Can you please provide details on below questions:

        1. Are the KPIs (say a quality metric) online measurements at specific frequencies or they are offline? Can you comment about the uncertainty of measurements in the KPIs and how that uncertainty is accounted for in your modeling?

        2. Because you cannot describe the exact use case,  let's consider for example an industrial use case at a water treatment plant - one of the KPIs could be a quality metric on contaminant ppm in water, that must stay below 100ppm, is measured maybe every 12 hours; the high frequency sensor data is sampled every 1sec. If my understanding of this talk is correct, you are building a ML regression model with specific set of variables (process parameters, sensor measurements, set-points. etc. - some could be actionable & controllable) for predicting the KPIs.

        (i) Can this model be treated like a proxy model for the online or offline KPI measurement?

        (ii) Is this model subsequently used in the optimization problem to arrive at the Pareto of operational conditions?

        (iii) At what time-range will be these Pareto recommendations for prescribing operational settings?

        (iv) Will you be covering the time challenges between synchronising KPI measurement and sensor-data measurements? How is model accuracy measured and what is it's impact on the recommendations?

        (v) Can you please include details on (i) to (iv) in your talk as they could be very valuable to the audience.

        Thanks,

        Ravi

        • Aravind Kondamudi
          By Aravind Kondamudi  ~  4 months ago
          reply Reply

          Thank you Ravi, for the excellent and important comment. Although we did not explicitly mention, our focus of the talk would be answering these questions. Below are the answers :

          Are the KPIs (say a quality metric) online measurements at specific frequencies or they are offline? Can you comment about the uncertainty of measurements in the KPIs and how that uncertainty is accounted for in your modelling.

          Ans : It can be both online measured KPI or an offline KPI like throughput recovery.

          As per uncertainty of KPI value, there can be two types - process inherent uncertainty or noise in the variable. The former is explainable while the latter not. We are trying to explain the first part by adding the significant process variables, disturbance variables which can be measured. There would still be a part which is not explainable like operator error, data capturing error or random variation which cannot be considered for modelling as it is not the true behaviour. We would try to minimize that part by adding more information. If you are asking about any other type of uncertainty, it will be extremely helpful if you can give some example for my understanding


          Because you cannot describe the exact use case,  let's consider for example an industrial use case at a water treatment plant - one of the KPIs could be a quality metric on contaminant ppm in water, that must stay below 100ppm, is measured maybe every 12 hours; the high frequency sensor data is sampled every 1sec. If my understanding of this talk is correct, you are building a ML regression model with specific set of variables (process parameters, sensor measurements, set-points. etc. - some could be actionable & controllable) for predicting the KPIs;

          (i) Can this model be treated like a proxy model for the online or offline KPI measurement?

          (ii) Is this model subsequently used in the optimization problem to arrive at the Pareto of operational conditions?

          (iii) At what time-range will be these Pareto recommendations for prescribing operational settings?

          (iv) Will you be covering the time challenges between synchronising KPI measurement and sensor-data measurements? How is model accuracy measured and what is it's impact on the recommendations?

          1. The first part of the model is predicting the KPI as accurately as possible, in our use case, we have built it to identify the optimum process variables and not to measure the KPI. A prediction model with all the response and control variables can be used as a process simulator where, for a given condition the model can predict the KPI value.
          2. It will be good if you can explain what exactly the Pareto of operational condition. In our use case, the model gives the optimum operation condition for optimum KPI.like how to get maximum throughput etc, which is the pareto of the objective function
          3. We have built the model for a batch process scenario which will give the optimum operational setting for a given batch for a given the product type or for a given set of disturbance variables.
          4. The main challenge is the synchronisation between KPI and the sensor parameters, which will be covered in detail, in the talk.
          5. We measure model accuracy in two ways :
            1. Historical error when predicting the KPI value and in error in a validation set with plant team.
            2. Creating a doe with the optimum parameter recommendations from the model and compare the performance of those parameters in the real KPI output.
          6. (v) We will include these details in the talk.
  • Natasha Rodrigues
    By Natasha Rodrigues  ~  5 months ago
    reply Reply

    Hi Aravind,

    To help the program committee understand your presentation style, can you provide a link to your past recording or record a small 1-2 mins trailer of your talk and share the link to the same?

    Thanks,
    Natasha

    • Aravind Kondamudi
      By Aravind Kondamudi  ~  5 months ago
      reply Reply

      Hi Natasha,

      Thank you for the comment. I will be uploading the video shortly and update you.

      Thanks,

      Aravind.

       

      • Aravind Kondamudi
        By Aravind Kondamudi  ~  5 months ago
        reply Reply

        Hi Natasha,

        I have updated my proposal with links to videos from both speakers.

        Thank you,

        Aravind.

        • Natasha Rodrigues
          By Natasha Rodrigues  ~  5 months ago
          reply Reply

          Hi Aravind,

          Thanks a ton for the same. 

          Regards,

          Natasha