Resource and Operations Management for Industry 4.0

Foundation Models for Time-series Forecasting

In process and manufacturing industries, we typically have 100s of sensors capturing different statistics of resources and assets in a periodic manner. Accurate forecasting of these sensor streams greatly helps in improved business decisions and efficient downstream tasks. To solve this, our team at IBM Research focuses on building a reusable & scalable platform for multi-variate long-term time-series forecasting incorporating a wide spectrum of various novel state-of-art algorithms and scalable ML pipelines. One of the key focus area in this work stream is to learn generalized representation for time-series in different industries and domains. The primary goal is to capture the domain knowledge of the industry as a time series embedding in a semi-supervised way, that can be used to bootstrap and speed-up various downstream tasks (like forecasting, anomaly detection, classification, etc.). These techniques greatly help in improving the accuracy of the forecasting models especially when there is data scarcity and less availability of labelled data. These assets are packaged as part of our Foundation Model ecosystem and can be customized based on various client requirements.

Process Insights and Operation Control

Fast-changing demand and market competition force every business and the process industry to continuously adjust to cater to these ever-changing needs, making it difficult to track and analyze the impact of every adjustment on the overall system stability. Adopting a data-driven model for process control and decision-making provides a scalable solution, which comes with many additional requirements,

1. Create a trustworthy model-driven solution.

2. Enabling a model-driven process control.

3. Monitoring model health.

Most validation approaches focus on accuracy metric, which often does not guarantee the model's correctness. This project focuses on model-driven process insights generation to gain trust, and use user feedback to correct model performance.

Active forecasting models the effect of external factors on the forecast. While essential for business strategy evaluation and process control, it demands process knowledge (the cause and effect chain), which is often unknown in many scenarios. As part of this work, we explore various modeling strategies, viz. data-driven simulated experiment-based resolution of causal impact, a differential equation-based model for approximating the data generating process. The integrated framework enables the direct incorporation of such approaches into the forecasting model.

Further most industries experience significant drift in data distribution over time, which is a big challenge to data-driven models, as their performance deteriorates with data drift. Continuous monitoring of the model health, identifying data drift, and updating the model when needed to avoid anomalies in the process control is an important component of this project.