Patented AI for large-scale time series data
Viaduct pioneered the industry's only AI specifically trained to detect patterns and make predictions on complex connected asset data – at massive scale.
GET STARTED.avif)
.avif)
Dozens of algorithms for time series search, detection, and prediction
.png)
High-performance, domain-specific query language for event-based data
.avif)
Customized language models for interactive, time-series data exploration.
Backed by industry-pioneering research
Spun out of Stanford AI Laboratory and based on some of the most foundational and frequently-cited research in time series AI:
- Toeplitz Inverse Covariance-Based Clustering of Multivariate Time Series Data
- Drive2Vec: Multiscale State-Space Embedding of Vehicular Sensor Data
- Greedy Gaussian Segmentation of Multivariate Time Series
- Network Inference via the Time-Varying Graphical Lasso
Foundation model for time series data
Discovering correlations across events and assets in large-scale multivariate time series data is computationally infeasible with traditional approaches. Our foundation model addresses this challenge by embedding asset states in a high-dimensional vector space, capturing relationships in time-series data to identify assets with shared conditions. This model integrates over 20,000 domain-specific features and incorporates patterns over short- and long-term time windows.
Unsupervised AI for early issue detection
Viaduct's Early Issue Detection (EID) algorithm employs a hierarchical framework for unsupervised anomaly detection in time-series data. Through a multi-stage process, the system integrates contextual signals to identify and rank emerging issues based on future projected impact. Benchmarked against standard approaches including Isolation Forest, Markov Random Fields, and autoencoder models, EID demonstrates significant improvement in detection latency and ranking accuracy.
LLM integrations for time series data
Off-the-shelf large language models struggle with time-series data analysis, where temporal relationships and multivariate patterns are critical. Viaduct's LLM integration combines our time-series foundation model with specialized language models optimized for industrial equipment data. This enables natural language interaction with complex datasets while maintaining the mathematical rigor required for industrial applications. The result is a query interface that understands both equipment behavior and time-series relationships—capabilities beyond generic LLMs.
