In recent years, the trend toward using larger, more complex AI models has become a popular approach. With advancements in large language models (LLMs) and deep learning architectures, the Data Science field is increasingly characterized by high-powered, high-resource algorithms.
At APERIO, however, we have achieved excellent results using a different approach. For industrial data management, our R&D team has turned towards small, targeted models – but at a very large scale. This approach has allowed us to deliver fast and reliable, real-time detection of data quality issues across massive time-series data streams without the overhead and performance limitations of traditional, large-scale models.
Why Go Small?
Each industrial data stream that our clients monitor—whether it’s pressure readings from a factory sensor or temperature fluctuations in a pipeline—presents unique, evolving patterns. Rather than relying on monolithic models that demand extensive tuning and retraining, our approach uses “micro-models” that are lean but highly adaptable. For every data channel, we maintain around eight active models at any given time, each with only a handful of parameters, allowing us to detect anomalies and disruptions with remarkable precision and efficiency.
Our micro-modeling framework also lends itself to high automation. Models are automatically built, validated, updated, and deprecated within a streamlined pipeline. This setup ensures that our algorithms are constantly learning and evolving in tandem with the data they analyze, without requiring manual intervention or even preliminary configuration.
Scaling Responsively
Apart from overcoming the limitations of large-scale models, the models we employ don’t need to be deeply complex to deliver value. With consistent updates, typically weekly, each model remains relevant and fine-tuned to the latest data patterns. This frequent refresh ensures that our clients receive the highest level of accuracy in anomaly detection without any lag that could impact operations. More importantly, because our models are updated automatically, our sophisticated streaming algorithms can keep pace with the data in real time, even when the environment is volatile or rapidly evolving.
Future-Proofing Industrial Data Management
Our unique model structure also allows APERIO to be highly scalable. As the volume of data grows, our system can add new models and adapt to additional channels without re-engineering. This scalability, combined with application-agnostic adaptability, makes our approach ideal for the demands of modern industrial data streams.
In a world where bigger models often come with trade-offs in cost, complexity, interpretability, and flexibility, we’re proud to embrace a smarter, more responsive method. With small, nimble models working in harmony with each other, APERIO is redefining what scalable, automated Data Science can achieve in an industrial context.
By Pele Schramm, Head of Data Science at APERIO.