Transformer MTS Manufacturing Patents 2026 | PatSnap Eureka
Transformer Models for Manufacturing Time Series
Transformer-based architectures for multivariate time series analysis are emerging as a foundational technology for manufacturing intelligence. This dataset spans 16 patent records from 2013 to 2026, with a clear inflection point around 2021.
Self-Attention Architectures Reshaping Manufacturing MTS Analysis
Within this dataset, transformer model applications to manufacturing multivariate time series span four core technical mechanisms: self-attention-based encoding of inter-variable and temporal dependencies, hybrid architectures combining transformers with LSTM and autoregressive components, 3D tensor-based convolutional deep learning for MTS representation, and ensemble and foundation model approaches over decomposed univariate channel streams.
Manufacturing processes generate concurrent, heterogeneous time series — temperature, pressure, humidity, yield, tool vibration, energy consumption, and logistics signals — that are structurally correlated but originate from disparate sources with varying sampling rates and noise profiles. Classical statistical models such as Vector Autoregression provide interpretable baselines but are constrained in capturing nonlinear cross-variable dependencies at scale.
The transformer architecture’s self-attention mechanism enables direct modeling of long-range temporal and inter-variate dependencies without the sequential bottlenecks of RNNs. IBM’s Time Series Forecasting patent explicitly claims decomposition of multivariate input into univariate subseries processed by a transformer model, connecting smart sensor inputs from manufacturing systems to planning system outputs.
In this dataset, IBM is the most prolific single assignee with 4–5 records across multiple patent families in retrieved records. Chinese university assignees — Shanghai Jiao Tong, Ocean University of China, Guilin University, and Qiqihar University — collectively represent the largest national filing bloc in this dataset, with China accounting for 10 of 16 patent records.
Architecture Clusters and Filing Activity Across the 2013–2026 Period
Four distinct technology clusters are identifiable in this dataset: transformer encoder with self-attention, hybrid transformer-LSTM-autoregressive architectures, 3D tensor and volumetric CNN approaches, and statistical VAR-based baselines. Filing activity in this dataset shows a clear inflection point around 2021.
Patent Records by Technology Cluster (Dataset Snapshot)
In this dataset, hybrid transformer-LSTM architectures and pure transformer encoder approaches each account for the most records, with 3D tensor/CNN and statistical baseline clusters contributing smaller shares.
↗ Click bars to explorePatent Filings by Era — Pre-Transformer vs. Transformer Period (Dataset Snapshot)
In this dataset, filings from 2021 onward represent the large majority of transformer-specific MTS patent records, confirming the 2021 inflection point when transformer architectures displaced or augmented statistical and LSTM-only approaches.
↗ Click bars to exploreKey Manufacturing Deployment Domains for Transformer MTS Models
Across the retrieved records, transformer and hybrid MTS architectures have been applied to at least six distinct manufacturing domains, ranging from predictive maintenance and yield forecasting to federated value chain analytics and pharmaceutical batch monitoring.
Production Tool Component Maintenance
National Cheng Kung University filed a US patent in 2022 (reissued 2025) using multiple-variable time series prediction with VAR models and information criterion algorithms to forecast accidental shutdown trends in production tool components. A corresponding CN-jurisdiction filing was published in 2023. This is the most patent-dense application domain in this dataset.
Predictive MaintenanceSemiconductor & Precision Manufacturing Yield
Shanghai Jiao Tong University’s 2024 CN patent discloses a dual-embedding hybrid model that predicts yield (产率) and quality rate (良率) from process environment variables including temperature, humidity, and pressure. The architecture combines LSTM encoding, cross-attention fusion, and parallel linear autoregression, targeting semiconductor and precision manufacturing production lines.
Quality & Yield ForecastingEquipment Manufacturing Material Forecasting
Qiqihar University’s December 2024 CN patent applies the iTransformer model variant combined with an improved Whale Optimization Algorithm (iWOA) for automated hyperparameter optimization, targeting material consumption time series in equipment manufacturing enterprises. IBM’s 2024 US patent separately links smart sensor MTS from manufacturing systems directly to downstream planning system outputs for supply chain integration.
Material Requirements PlanningSUSTech Manufacturing Value Chain Model
Southern University of Science and Technology filed two CN patents in July 2024 and August 2025 on a Manufacturing Value Chain Joint Large Model applying federated learning across supply chain partners to train shared large models for multi-dimensional manufacturing data analysis without centralized data pooling. Sartorius Stedim Data Analytics’ 2013 US patent provides the statistical baseline for batch process multivariate monitoring of dependent and manipulated process variables in pharmaceutical manufacturing.
Federated & Batch AnalyticsLeading Patent Assignees in Transformer MTS Manufacturing — Dataset Snapshot
In this dataset, IBM (International Business Machines Corporation) is the most prolific single assignee with 4–5 patent records spanning 3D tensor MTS prediction, channel-hybrid foundation models, and univariate-decomposition transformer forecasting in retrieved records. Chinese university assignees collectively represent the largest national filing group in this dataset, with Ocean University of China and Shanghai Jiao Tong University among the most active individual institutions.
Top Assignees by Filing Count in Retrieved Records (Dataset Snapshot)
↗ Click bars to exploreInternational Business Machines Corporation
IBM holds 4–5 patent records in this dataset spanning filings from 2021 to 2024 in the US jurisdiction. Key patents include 3D tensor-based MTS prediction (2021, 2023), the Hybrid Channel Modeling for Time Series Foundation Models built on PatchTST architecture (November 2024), and a univariate-decomposition transformer forecasting patent linking smart sensor inputs to planning systems (2024). IBM’s portfolio represents the broadest deep learning MTS architecture IP coverage in this dataset.
United StatesShanghai Jiao Tong University
Shanghai Jiao Tong University filed two CN-jurisdiction patents in this dataset (2022 and 2024 versions) covering a Manufacturing Multi-Data Value Collaborative Prediction Method using dual-embedding of target and environmental variables, LSTM encoding, cross-attention fusion, and parallel linear autoregression. The 2024 filing targets yield and quality rate prediction from temperature, humidity, and pressure process variables in semiconductor and precision manufacturing contexts.
China — CNFour Emerging Frontiers in Manufacturing MTS Transformer Technology
The most recent filings in this dataset (2024–2026) identify five emergent directions, with foundation model adaptation, AutoML-driven transformer deployment, and federated multi-enterprise training representing the most strategically significant near-term shifts.
Time Series Foundation Models with Hybrid Channel Architectures
IBM’s Hybrid Channel Modeling for Time Series Foundation Models (US, November 2024) introduces mixed channel-independent and channel-dependent backbones built on PatchTST, with pre-training via patch-based masking and normalization workflows. This signals the transition from task-specific transformer models to general-purpose foundation models that are fine-tunable for specific manufacturing deployments. Organizations investing in narrow, manufacturing-specific transformer models risk architectural obsolescence as pre-trainable foundation models mature.
Hyperparameter-Optimized iTransformer via Metaheuristic AutoML
Qiqihar University’s iWOA-iTransformer Method (CN, December 2024) demonstrates that metaheuristic optimization — specifically the improved Whale Optimization Algorithm — is being applied to automate iTransformer hyperparameter selection for manufacturing-specific material consumption time series in equipment manufacturing enterprises. This points toward AutoML-driven transformer deployment, reducing the expert knowledge required to configure MTS models for specific manufacturing contexts. GE Infrastructure Technology’s WO filing (April 2025) further signals low-code, automated MTS model pipeline deployment using TensorFlow/PyTorch-compatible deployable objects.
Transformer-Encoder vs. Hybrid Transformer-LSTM for Manufacturing MTS
Click any row to explore further.
| Dimension | Transformer Encoder (Self-Attention) | Hybrid Transformer-LSTM-Autoregressive |
|---|---|---|
| Representative Patent | IBM Time Series Forecasting (2024, US); Guilin Univ. MTS Forecasting (2026, CN) | Shanghai Jiao Tong Univ. Manufacturing Multi-Data Value Prediction (2024, CN) |
| Core Mechanism | Self-attention for global temporal and inter-variable correlation; positional encoding; frequency-domain (Fourier) filters | Dual-embedding of target and environmental variables; LSTM encoding; cross-attention fusion; parallel linear autoregression |
| Target Variables | Manufacturing sensor MTS decomposed to univariate channels; industrial MTS with spatial-temporal patterns | Yield (产率), quality rate (良率) predicted from temperature, humidity, pressure process variables |
| Key Strength | Long-range temporal dependencies; global structural pattern extraction; adaptable to foundation model pre-training | Captures both nonlinear long-range dependencies (attention) and local autoregressive linear structure simultaneously |
| Primary Application Domain | Manufacturing-to-planning system pipelines; material demand forecasting; industrial MTS broadly | Semiconductor and precision manufacturing quality and yield forecasting; production line process control |
| Filing Jurisdiction | US (IBM); CN (Guilin Univ., Ocean Univ. of China) | CN (Shanghai Jiao Tong University, 2022 and 2024 versions) |
| Maturity Signal | Active patent generation through March 2026; IBM transitioning to foundation model backbone designs | Two patent family versions filed (2022 and 2024), indicating iterative refinement of the same core invention |
Frequently Asked Questions: Transformer MTS in Manufacturing Patents
The earliest directly relevant filing in this dataset is IBM’s Performing Multivariate Time Series Prediction with Three-Dimensional Transformations (US, filed January 2021), which marked early deep learning architecture innovation for MTS in manufacturing contexts by moving from 2D to 3D tensor representations.
In this dataset, IBM (International Business Machines Corporation) is the most prolific single assignee with 4–5 patent records, spanning 3D tensor MTS prediction (2021, 2023), channel-hybrid time series foundation models (2024), and univariate-decomposition transformer forecasting (2024) across US-jurisdiction filings.
10 of the 16 patent records in this dataset originate from CN-jurisdiction filings. Assignees include Ocean University of China, Shanghai Jiao Tong University, Guilin University of Electronic Technology, Qiqihar University, and Southern University of Science and Technology (two filings).
According to the content, predictive maintenance and yield/quality forecasting are the most patent-dense manufacturing application domains in this dataset. These are described as the most commercially validated use cases, with differentiation occurring at the architecture level (hybrid attention-autoregressive models) or the deployment level (federated/privacy-preserving training).
The iWOA-iTransformer method was filed by Qiqihar University (CN, December 2024). It combines the iTransformer model variant with an improved Whale Optimization Algorithm (iWOA) for automated hyperparameter optimization, applied specifically to material consumption time series in equipment manufacturing enterprises.
IBM’s Hybrid Channel Modeling for Time Series Foundation Models (US, November 2024) introduces a hybrid channel-independent / channel-dependent backbone anchored on the PatchTST architecture for time series foundation models, with pre-training workflows including normalization, patching, masking, and permutation of univariate decompositions, enabling fine-tuning for specific manufacturing deployments.
Data and insights on this page are based on a limited patent and literature dataset and are for reference only. Figures may not represent the complete technology landscape.