Zhengyang

Research profile

Research summary

Zhengyang is an early-stage industrial AI researcher whose visible record — resolved from public bibliographic data — consists of three first-authored arXiv preprints, all dated 2026. Output is evenly spread across three subfields, each accounting for about one-third of the record: industrial machine learning and MLOps for cyber-physical manufacturing systems, machine learning for industrial AI and predictive maintenance, and reinforcement learning for industrial control. The most applied work describes autonomous model lifecycle management for digital twin-based manufacturing control in automotive production, reporting process capability (Cpk) improvements of 28–45% versus uncontrolled operation — though evaluation was internal-only, with no code or data released. A second preprint proposes a hybrid PID controller augmented with TD3-discovered Gain/Shift actions, matching the best deep RL agent on the Industrial Benchmark (-203.46 vs -204.92) in a single-episode evaluation. The third benchmarks hybrid LSTM-Transformer architectures for Industry 4.0 predictive maintenance over more than 700 runs, claiming noise-robustness advantages that remain untested against linear baselines. Evolve's reproducibility review averages 32/100 across the three papers (two rated low, one medium readiness), reflecting closed evaluation, missing significance testing, and synthetic targets. The collaboration footprint is compact: seven distinct coauthors on the three preprints, all first-authored, with no corresponding-author role yet. Impact, output, and activity metrics are not yet available.

Research keywords

Highlights

• Deployed digital-twin MLOps with +28–45% Cpk gains

  First-authored 2026 arXiv preprint on autonomous model lifecycle management for digital twin-based manufacturing control in automotive production. Evolve's verdict: a deployed system reporting Cpk improvements of 28–45% versus uncontrolled operation, but internal-only evaluation with no code or data released.

• Hybrid PID–TD3 controller rivals deep RL on Industrial Benchmark

  First-authored 2026 preprint augmenting PID control with TD3-discovered Gain/Shift actions, matching the best deep RL agent (-203.46 vs -204.92). Evolve flags single-episode evaluation, no significance tests, and an unproven MIMO premise.

• 700+ run benchmark of hybrid architectures for predictive maintenance

  First-authored 2026 preprint comparing hybrid deep architectures for Industry 4.0 predictive maintenance; the LSTM-Transformer is claimed to win on noise robustness across 700+ runs, though synthetic linear targets, absent significance tests, and missing linear baselines weaken the evidence (medium repro readiness).

Timeline

1. 2026Published first-authored arXiv preprint on autonomous model lifecycle management for digital twin-based manufacturing control, reporting Cpk gains of +28–45% in a deployed automotive setting.
2. 2026Released a hybrid PID–TD3 approach to the Industrial Benchmark, matching the best deep RL agent (-203.46 vs -204.92).
3. 2026Benchmarked hybrid deep learning architectures (LSTM-Transformer) for Industry 4.0 predictive maintenance across 700+ runs.

Papers in the Evolve catalog

• 2026-09-19first author

  Autonomous Model Lifecycle Management for Digital Twin-Based Manufacturing Control

• 2026-09-18first author

  Augmenting PID Control with Deep Reinforcement Learning: A Hybrid Approach to the Industrial Benchmark

• 2026-09-18first author

  Benchmarking Hybrid Deep Learning Architectures for Predictive Maintenance in Industry 4.0

Frequent co-authors