Peer-reviewed papers and presentations

Modern naval power systems face significant operational challenges as increasing power demands transform power systems engineers into data pipeline architects. Organizations currently struggle with disconnected or air-gapped environments, legacy data formats, and the inherent difficulty of visualizing high-frequency signals at scale. These technical bottlenecks, including missing sensor metadata and complex stream limitations, frequently cause tool chain failures and prevent timely analytics, ultimately hindering mission readiness and infrastructure resilience.

The Falkonry Time Series Platform addresses these challenges by serving as a standardized bridge between design and operations, facilitating seamless edge-to-cloud deployment. By automating the contextualization, normalization, and analysis of multi-sensor data, Falkonry enables advanced AI capabilities—such as physics-based modeling and unsupervised pattern clustering—to turn raw sensor data into actionable intelligence. This integrated approach supports interoperability with enterprise tools like CMMS, Power BI, and LLM-based workflows, ensuring that critical data is available, consistent, and ready for decision-making across the entire system lifecycle.

This presentation details the evaluation of the data of a 720-hour continuous reliability demonstration for the MTU 20V 4000 M93L Main Propulsion Diesel Engine. The project served as a critical testbed for evaluating engine performance without human intervention, while simultaneously experimenting with advanced computational infrastructure for data acquisition, contextualization, and anomaly detection.

We examine the implementation of both cloud and edge computing, the use of secondary sensing technology like vibration analysis, and the challenges of data integration from disparate sources. Key findings underscore that while the mechanical equipment demonstrated high reliability, success in unmanned operations requires significant improvements in testing capabilities, data strategy, and infrastructure robustness. The session concludes with actionable lessons learned regarding data acquisition, human factors, and the deployment of AI-driven analytics to support future autonomous maritime operations.

Effective real-time monitoring is crucial for modern iron and steel operations to prevent issues like cold furnaces or unscheduled shutdowns, which leave signatures in high-volume real-time data. Conventional monitoring often misses these signs.

Artificial intelligence enhances our ability to detect these signatures, helping to avoid lost production, improve quality, and increase campaign duration. Effective AI must offer significant avoidance opportunities, require minimal attention from operations and maintenance, and cover most operational issues to gain trust and advance digital manufacturing. Therefore, measuring and communicating this effectiveness is vital for digitalization in the steel industry.

This paper introduces a new metric for judging the effectiveness of real-time monitoring in digital transformation projects, regardless of the technology. We illustrate this metric with multiple real-world case studies of Time Series AI platforms applied by steelmakers globally.

Smart steelmaking is impeded by the need for long, manual cycles of data collection, data preparation, domain analysis, model building, operationalization, maintenance, and root cause analysis. It is simply inconceivable that supervised approaches can be applied at any scale, given data quality issues that are prevalent in the industry. Furthermore, supervised approaches simply cannot surface behaviors that are not previously recorded and understood, which further limits their adoption. We have therefore looked for new solutions to achieve line scale Smart steelmaking primarily in the form of anomaly detection from SCADA and PLC data. In this paper, we demonstrate the automatability and scalability of two approaches - semi-supervised and selfsupervised (hands-free) - with examples from continuous casting and demonstrate that the performance is comparable. A linescale anomaly detection approach enables Smart real-time decisions over the whole steel mill. The hands-free approach also does not require any upfront investment in data quality or labeling as well as the need for ongoing manual model maintenance.

This paper presents ways to leverage various data types acquired during steel manufacturing to enable fully data-driven decisions on the production floor. By combining a robust data acquisition system (iba) with an automated AI system (Falkonry), manufacturers can instantly gain operational visibility and insights without any data science/data engineering/IT/MLOPS resources, that enable proactive decisions. This enables operational users to proactively diagnose the root causes of operational issues and take timely corrective actions. Both secure and scalable, this integrated approach can expand to cover the entire plant. We showcase the benefits achieved by leading steel manufacturers in improving productivity and quality.

Due to the unavailability of highly curated data, manufacturers are increasingly adopting anomaly detection to analyze large volumes of operational data in real-time. It is very difficult to review every anomaly manually because of limits on available manpower. This paper, therefore, presents a novel approach to automatically prioritize anomalies to direct human attention to where it is most needed. Using anomaly severity, anomaly persistence, signal importance, anomaly spread, and contextual information, the automated anomaly detection AI enables prioritization of attention to review the critical anomalies for timely diagnosis and corrective actions. We present practical cases of deploying this approach in line-scale steelmaking operations.

In this presentation, we showcase a novel methodology for automated strip break classification that uses Falkonry’s time series AI platform to classify the complex waveforms of tandem mill parameters associated with strip breaks. Strip breaks are unwanted yet common occurrences during cold rolling of steel. Strip breaks have various causes, such as weld breaks, material defects, or steering issues, and result in lost productivity, material scrap, and equipment damage. Today, analysis of strip breaks requires manual extraction of multiple time series parameters from SCADA systems or historians and extremely resource and time-intensive human interpretation of these parameters just before the strip break event. The resources and time required to analyze strip breaks are expensive, but more expensive is the lost time between events and diagnosis. Breaks are temporally indiscriminate. Human analysts are neither on-call nor capable of near-instant analysis — which is required for production teams to determine causes and actions in response to strip breaks. Automated classification allows low-latency automated classification of strip breaks for use by operations teams to understand underlying causal factors and implement corrective actions.

This paper presents a self-supervised autonomous “plant scale” AI — capable of monitoring every PLC and IIoT parameter of a steel plant — automatically detecting and accelerating diagnosis of anomalies. Automatic anomaly detection proactively informs plant operations of conditions that otherwise would go undetected — leading to informed production and maintenance decision-making. Self-supervised AI overcomes the challenges of constant equipment, environment, and product changes that thwart classical machine learning approaches. Normalized severity scoring of the AI results further enable prioritization of the anomalies for investigation and action. We describe several use cases of this new AI in commercial operation along with the corresponding user workflow.

Artificial Intelligence (AI) and Machine Learning (ML) techniques have been used to solve complex manufacturing operations problems. Applying ML and AI to anomaly detection and diagnosis at scale, however, has been a significant challenge. This paper discusses how Falkonry’s Time Series AI platform leverages ML/AI for automated detection and diagnosis of equipment in steelmaking, detecting precursor conditions to critical equipment failures 3-10 weeks in advance of breakdowns. This methodology is scalable across use cases without the need for data scientists. Precedent detection of novel equipment conditions provides insight into maintenance operations that would otherwise have been missed. Such insights lead to proactive maintenance interventions, thus avoiding loss of production due to unexpected downtime events.

Recently, Artificial Intelligence (AI) and Machine Learning (ML) techniques have been used to solve complex operations problems. However, scaling ML/AI across a multitude of equipment types and use cases, a variety of signals, and over time with changing operations remains a significant challenge. This paper discusses how Falkonry’s Operational AI platform learns, detects, and predicts conditions in continuous casting. This methodology is scalable across use cases and time without the need for data scientists. Precedent detection of impending equipment failures allows operations to schedule necessary maintenance interventions, thus avoiding loss of production due to unexpected downtime events.

Ciner Resources, a global leader in soda ash production, faced a situation common in the industrial world. Investments in instrumentation and data collection were producing large volumes of operational data, but efforts to turn this data into meaningful improvements in operational efficiency were falling short.

Attempts to use statistical analytics or approximate engineering models proved to be time consuming and limited in applicability. While machine learning techniques held promise, any approach that required an external team of data science and software experts was a non-starter. Ciner needed an approach that empowered their process engineers to use their data and their deep understanding of the plant to gain better operational insights. They found that approach with Falkonry.

In industrial settings—from semiconductor fabrication to oil, gas, and metal production—modern equipment generates trillions of measurements daily, far exceeding the capacity of human analysis and conventional heuristics. To address this, we present “Time Series AI,” a novel, deep-learning-based approach for scalable, unsupervised anomaly detection in complex industrial time series data.

Unlike traditional methods that rely on labor-intensive manual rule creation and univariate statistics, Time Series AI leverages Convolutional Variational Autoencoders (CVAE) to automatically characterize sensor traces and identify anomalies. By defining an anomaly as the “residual difficulty” the model faces in reconstructing the input data, our approach eliminates the need for predefined anomaly modeling or extensive labeled training sets.

This presentation outlines how how to utilize time series AI to drive operational excellence in semiconductor manufacturing. By focusing on equipment performance visibility, the proposed solution enables manufacturers to reduce unplanned downtime, optimize resource usage, and improve product quality without the necessity of extensive data science teams.

The core of Falkonry’s approach involves mining multivariate, temporal patterns in equipment data to automatically discover and distinguish operational conditions. This allows subject matter experts (SMEs) to directly participate in the AI curation process—identifying, tagging, and explaining events—which facilitates the transition from raw data to actionable insights. By integrating with existing fab systems and supporting diverse deployment environments (cloud, on-premise, air-gapped), the technology serves as an end-to-end workflow solution for predictive maintenance, defectivity analysis, and equipment fingerprinting. Ultimately, the presentation argues that domain knowledge is critical to effective AI, and that by empowering engineers to use AI, organizations can achieve greater precision, higher yields, and improved operational efficiency.

This presentation details the process and results of applying machine learning to improve production yield and optimize semiconductor manufacturing processes.