Architecture

Falkonry Time Series Intelligence (TSI) is architected as an intelligent decision support system designed to securely, reliably, rapidly, and cheaply translate large quantities of raw telemetry, sensor, or time series data into decisions. It specializes in streaming time series analytics. Its AI-native approach is fundamental to its design, addressing the complexities and volume of industrial and IT time series data that conventional analysis methods struggle with.

TSI has a cloud-native architecture, ensuring global accessibility, but can also be deployed in a private cloud or on compatible local hardware to meet specific security, legal, and integration requirements. Its interface can be used from any Web browser and is optimized for desktop and tablet displays. TSI is built on Falkonry’s proprietary technology, called PatternIQ™, into a commercially validated tool for industrial time series visualization, anomaly detection, classification, and collaboration.

By isolating remarkable events from time series data, Falkonry makes it possible to perform digital actioning and knowledge management, which subsequently enables the use of language models to bring together case histories, maintenance documentation, best practices, and other public materials to accelerate root cause analysis as well as for developing field service engineer procedures.

Here's how data moves through the TSI architecture:

1. Data Ingestion

In this stage, TSI ingests raw data and converts it into logical values. Additionally, it also enriches these values by attaching them to metadata that is organized in hierarchies. This metadata enables a dynamic and contextualized view of events and data, enhancing the system's ability to interpret and analyze time series information effectively.

Connect

Data enters the Falkonry Time Series Intelligence (TSI) system through Inbound Connections. These connections act as secure links between TSI and various external data systems. Data, which includes a time component and values corresponding to a named signal, are pre-processed using programmed transformations. This allows creation of synthetic signals and reduces the need for external data pre-processing for cleaning up bad values, renaming signals or performing simple formatting corrections. Once data starts flowing, TSI auto-detects new signals, which once approved, start receiving persistent time series data as and when data arrives over a connection.

TSI supports connecting to a wide range of industrial and IT data acquisition systems, including:

• Open Telemetry
• IoT Gateways like Litmus Edge, Ignition Edge, and Kepware
• MQTT brokers
• iba Systems (e.g., ibaDatCoordinator v3 or higher, ibaAnalyzer)
• File upload
• AWS IoT Core.
• Azure IoT Hub and Event Hub.
• S3 objects containing data files
• Falkonry observation API

Organize

Signals are organized into Signal Trees to provide a multi-headed organization of signals based on their metadata. This helps in visualizing data, setting context, and performing analysis. Multiple signal trees can be created for the same set of signals, serving different use cases (e.g., by system hierarchy, physical structure, signal type). Mutliple trees can be intersected more rapidly to locate the signals of interest for any use case.

This organization provides critical context for AI analysis, enabling visualization, setting context, and efficiently analyzing data, including AI-generated insights and alerts, across signal hierarchies. TSI leverages this organization to automatically learn about systems and their physical behavior without manual setup.

2. Time Series Processing (Understand, Find What Matters, Live Data Analysis)

An AI-native approach to processing time series data enables complex analytics on streaming data, which is achieved through a combination of infrastructure and a specialized time series processing core.

Foundational Infrastructure:

• Tiling Infrastructure: TSI uses a tiling infrastructure that stores raw data and univariate aggregates, supporting sample-hold lookback and a built-in time hierarchy. This allows for flexible visualizations and efficient queries, as basic statistical information is pre-aggregated and readily available, even when zooming out to large time ranges
• Clocks and Events: TSI manages data flow using a Signal clock, which advances with the latest received signal data and an Agent clock, which advances when the agent produces a new value based on its definition using all its input signals.
• Stream Processing: TSI is explicitly designed as a "data processing engine that is designed with infinite data sets in mind". This means it can handle massive, unbounded data sets, spreading workloads evenly over time for consistent resource consumption. This is a core reason why it's well-suited for AI, as AI models thrive on continuous, high-volume data streams.

PatternIQ™ (Time Series Processing Core)

AI-Native Processing Core: PatternIQ™: At the heart of TSI's analytical capabilities is PatternIQ™, Falkonry’s proprietary and patented technology. It is purpose-built to discover and explain patterns and anomalies from high-resolution time series data.

Rules (Logic-Based Alerting)

Rules convert any number of analog quantities (raw signals, anomaly scores, condition labels) into discrete events based on thresholds. Rules can be simple or compound/nested, using the output of other rules or AI inferences (from Insights or Patterns) as input. This allows for smart alerts based on AI inference, providing contextual control to prevent nuisance alerts and consolidate alerts.

Insights (Automated Anomaly Detection)

Insights automatically produces anomaly scores by systematically identifying deviations from normal operating behavior without requiring any setup or supervision, or labeled examples. It learns "normal" behavior from historical data. In the process, it takes into account noise, shapes, waveforms, value distributions and time dilation to understand the normal behavior. This is a key AI-native feature, as it tackles the "unknown unknowns" – problems that have never been seen before and for which no predefined alarms exist. It uses a Deep Neural Network (DNN) architecture to learn embeddings and to producing normalized reconstruction errors as anomaly scores.

Patterns (Multivariate Pattern Recognition)

Patterns discovers meaningful temporal correlations, such as early warning signs or deterioration stages, from multivariate data streams. It supports unsupervised, semi-supervised, and supervised learning, allowing pattern discovery with few or no labeled examples. It also doesn't require any feature engineering effort on the part of users. This capability can be used to automate root cause analysis for recurring issues.

Underlying AI Architecture Optimizations

• Resource Efficiency: The AI models are designed for efficiency, with training that converges quickly, very small model sizes compared to most DNNs, and efficient training/inference that does not require expensive GPU resources. This enables real-time inference on small-capacity edge devices.
• Incremental Evolution: Individual signal learning and iterative training allow models to evolve economically as system behavior changes.
• Data Pathology Readiness: TSI is designed to handle "data pathologies" like missing values, outliers, irregular sampling, and heterogeneous data sources without special handling or data engineering. This is a crucial AI-native design choice as real-world industrial data is often imperfect.

3. Data Output & Integration (Presentation)

Once data is processed and analyzed, TSI facilitates the presentation of results and integration with external systems in order to drive further action.

• Visualization and Reports: TSI provides intuitive and dynamic visualization tools for time series data, allowing zooming and panning. Reports enable analysis and documentation, including charts, rich-text summaries, and sharing of findings, which supports data-driven root cause analysis. A corpus of such reports further enables automatic recall of human expertise based on recurrent or similar event.
• Alerts: Rules generate alerts based on met criteria, which can be threshold-based or logic-based.
• Outbound Connectivity (Integration of Workflows): TSI can send live output (e.g., Rules output, alerts) to external systems through various outbound connections. This includes Webhooks (triggering JSON HTTP payloads), MQTT (publishing JSON messages to external brokers), and direct integration with CMMS/EAM systems (e.g., eMaint, IBM Maximo, SAP Ariba, Ultimo, MaintainX) to automatically create work orders or requests from alerts. REST APIs allow other systems to pull data from TSI, supporting custom integrations and bulk data operations

In essence, TSI is a comprehensive platform that takes raw time series data from various industrial sources, processes it through advanced AI models (Insights, Patterns, Rules) and internal infrastructure components, and then presents findings through visualizations and reports or pushes actionable alerts to external systems to drive decision-making and automation. This entire process is designed for scalability and efficiency, minimizing the need for manual coding or infrastructure management.