Why NUBISON AX

Technology Innovation

Accepted at NeurIPS 2025

Oracle AD

A principle-based industrial AI technology setting new standards for complex multivariate time-series anomaly detection in industrial environments

Key Features
  • Anomaly detection based on multivariate time-series data
  • Precise identification of anomaly occurrence time and intervals
  • Root cause tracing at variable (sensor) level
  • Predictive maintenance and downtime minimization
  • Detection of complex long-term anomaly patterns
Key Applications
  • Semiconductor and display ultra-precision manufacturing
  • Robotic automation and smart factory equipment
  • Large-scale sensing-based industrial facilities
  • Energy, environmental, and plant operations
Technical Differentiators & Advantages

Innovative First-Principle Based Approach

01
World-Class Performance Verification

Proven technical excellence on major benchmarks including PSM, SMD, and SWaT, simultaneously improving detection performance and anomaly localization accuracy

02
Principle-Based Anomaly Detection

Breaking away from model complexity competition, achieving high reliability and stability through design based on fundamental principles of anomaly signal generation

03
Stability and Reproducibility

Structurally defining inter-variable relationships in latent space under normal conditions, enabling precise detection of long-term and cumulative anomalies

04
Dual Scoring Mechanism

Combining prediction error and structural deviation to minimize false positives while simultaneously enhancing detection sensitivity and stability

05
High Interpretability

Clearly identifying root cause variables from latent structures when anomalies occur, supporting engineers' immediate root cause analysis and predictive maintenance

06
Industrial Field Optimization

Optimized for complex sensor data in high-precision manufacturing environments such as semiconductors, displays, and robotic automation, ready for immediate deployment in real operational environments

AI Paradigm Innovation

Neuro-symbolic AI

An AI paradigm that integrates deep learning-based learning capabilities with rule, logic, and knowledge graph-based reasoning into a unified system for simultaneous prediction, explanation, and verification

Key Features
  • Neural network-based perception and pattern recognition (images, video, time-series, text)
  • Ontology, knowledge graph, and logic rule-based reasoning and consistency verification
  • Constraint checking on neural model outputs with automatic rule violation detection and correction
  • Background knowledge-based reasoning and generalization in low-data environments
  • LLM-integrated natural language interface for Q&A and verification pipelines
Key Applications
  • Smart factory and manufacturing process analysis (equipment, sensors, process history)
  • Quality inspection and vision inspection automation (video analysis + quality specification rule checking)
  • Process and equipment compliance and regulatory automatic inspection
  • Knowledge graph-based manufacturing analysis and root cause tracing
Technical Differentiators & Advantages

Structural Integration of Neural Learning and Symbolic Reasoning

01
Neural Learning + Symbolic Reasoning Integration

Combining logic, rules, and knowledge graphs on top of deep learning's high-performance prediction capabilities to implement logically verifiable decision-making beyond simple probability outputs

02
Explainability and Transparency

Explicitly traceable reasoning paths based on rules, ontologies, and logical expressions, mitigating the black-box problem and suitable for industrial and regulatory environments

03
Embedded Domain Rules and Constraints

Reflecting medical guidelines, manufacturing specifications, and safety/compliance rules in the symbolic layer, enabling automatic verification, blocking, and correction when AI outputs violate rules

04
Advanced Reasoning with Limited Data

Performing reasoning, analogy, and abstraction using domain knowledge and logic even in environments lacking large-scale data

05
LLM Hallucination Mitigation

Achieving enterprise-level reliability through pipelines that verify and provide feedback on LLM-generated results using symbolic reasoners

06
Industry and Manufacturing Scalability

A general-purpose AI architecture immediately applicable to industrial sites with clear domain structures such as processes, equipment, quality, safety, and knowledge graphs

Next-Generation Multimodal AI

INSPIRE

A core technology that fundamentally resolves modality imbalance in asynchronous multimodal environments and overcomes the limitations of real industrial data integration

Key Features
  • Fully asynchronous multimodal data integration
  • Mitigation of information imbalance between modalities
  • Enhanced expressiveness through indirect synthesis
  • Stable performance maintenance in noisy and missing data environments
  • Advanced multimodal classification and prediction performance
Key Applications
  • Industrial equipment and process monitoring (sensors + unstructured data)
  • Manufacturing and robotics-based multi-sensor systems
  • Sports analysis (images + trajectory and time-series data)
  • Healthcare multimodal diagnostic data
  • Complex AI analysis systems in asynchronous collection environments
Technical Differentiators & Advantages

Innovative Multimodal Integration via Indirect Synthesis

01
Pre-emptive Modality Imbalance Resolution

Unlike conventional direct fusion approaches, each modality absorbs meaningful information from other modalities before fusion, fundamentally solving the problem of weaker modalities being overshadowed

02
Cycle Consistency-Based Domain Preservation

Enabling information exchange while maintaining original domain characteristics through bidirectional transformation between modalities, achieving stable representation learning even in asynchronous and heterogeneous environments

03
Dual Alignment Mechanism

Simultaneously applying sample-level proximity alignment and distribution-level statistical alignment to achieve both local and global representation consistency

04
Optimized for Asynchronous Environments

Not requiring temporal alignment or common timestamps, enabling immediate application to irregular and asynchronous data collection environments in real industrial sites

05
Strong Performance and Robustness

Achieving state-of-the-art accuracy, F1, and AUC performance on the RallyPose benchmark compared to existing single-modal, ensemble, direct fusion, and cross-attention approaches

06
Scalability and Versatility

A general-purpose framework extensible to various modality combinations beyond image-time-series, applicable across complex data environments including industrial, healthcare, and sports

Unified Data Collection

Data Blind Spot Removal

A data integration technology that eliminates data blind spots in industrial sites through a collection system encompassing both standard and non-standard equipment

Key Features
  • Unified collection of standard and non-standard equipment and system data
  • Support for multiple communication methods including TCP/IP, PLC, Bluetooth, MQTT/S, REST API
  • Edge Gateway-based data collection and on-site integration
  • Collection system configuration based on Link Server / Driver Server
  • Transformation of non-standardized data into usable formats
Key Applications
  • Manufacturing sites with mixed standard communication interfaces
  • Environments operating both legacy and new equipment
  • Non-computerized equipment environments with manual input or display-based systems
  • Facilities establishing equipment data-based analysis and management systems
Technical Differentiators & Advantages

Innovation in Standard and Non-Standard Data Integration

01
Unified Standard/Non-Standard Data Collection

Enabling integration of industrial standards (Modbus, OPC-UA, etc.) and non-standard equipment data into a single collection system

02
Diverse Link-Based Data Integration

Providing flexible integration structures including Gateway, Mobile, Direct, LPWA, System, and Digital Convert Links according to field environment requirements

03
Data Blind Spot Minimization

Expanding data collection scope to include equipment areas where data collection was previously difficult

04
RPA and OCR-Based Non-Computerized Equipment Data Collection

For single-output port or display-based equipment: RPA-based automation, predefined region-of-interest data extraction, and OCR recognition for standard text and non-standard display values

05
Centralized Data Collection Architecture

Enabling consistent data management by transmitting extracted data to central data collection servers

06
Scalable Collection Architecture

An extensible architecture that flexibly accommodates new equipment additions and communication protocol changes

Time-Based Data Reconstruction

Legacy Data Integration

A data integration technology that restructures distributed legacy data on a time basis to enhance process data usability and AI applicability

Key Features
  • Integration of legacy system data from MES, PLM, and other systems
  • Mapping data without time information to time-axis references
  • Transformation of individual variables into time-series data format Xn(t)
  • Thing Driver code-based data transformation and time mapping
  • Web-based editing tools for immediate code modification and application
  • Real-time integration through dynamic Thing Driver application
Key Applications
  • Manufacturing sites operating MES-based production management systems
  • Industrial environments with legacy system-centered process management
  • Sites with AI application constraints due to insufficient time information
  • Industrial data lake framework implementation environments
Technical Differentiators & Advantages

Thing Driver-Based Time Information Reconstruction

01
Time-Centric Data Reconstruction

Reorganizing existing production data without time information along time-axis references to enable process-level analysis

02
Thing Driver-Based Integration Architecture

Applying Thing Driver codes to individual data for time information mapping and unified management through a single Driver

03
Flexible Code Management and Extensibility

Mapped codes can be immediately modified and applied through web-based editing tools, ensuring field responsiveness

04
Large-Scale Information Retrieval Support

Enabling retrieval of thousands of diverse information types using encapsulated code collections

05
Real-Time Integration Design

Completing data integration in real-time by dynamically applying generated Thing Drivers to the system

06
Maximized AI Applicability

Expanding the scope of machine learning and AI analysis model applications through data transformed into time-series format