QHQunshan He

Qunshan He

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Qunshan He has a total of 36 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Explicit Timestamp Modeling for Long-term Time Series Forecasting via Pattern and Residual Fusion

    by Simengxu Qiao, Kaiwen Xu, Yichen Song, Qunshan He, Xinggao Liu

    Long-term time series forecasting remains challenging due to the complex entanglement of deterministic periodic patterns and irregular stochastic residuals. Existing methods often utilize timestamps merely as implicit embeddings, neglecting their potential to capture global multi-scale periodicities. To address this limitation, we propose TTW, a lightweight, plug-and-play framework designed to explicitly exploit temporal cues. Specifically, the framework employs a Time Index Generator to encode multi-scale temporal attributes, guiding a Temporal Pattern Extractor to rigorously disentangle basic patterns from irregular residuals. Furthermore, a Weight Constrained Combiner adaptively fuses these components under strict constraints to ensure interpretability. Instantiated with a simple Multilayer Perceptron, TTW-MLP achieves stateof-the-art or competitive performance on mainstream benchmarks, including Electricity and Weather, with only 0.42 million parameters. Empirical results validate the synergy among explicit timestamp modeling, residual-focused learning, and constrained fusion, demonstrating superior efficiency over complex Transformer-based architectures.

  • Open Access

    LMT-Net: A Hybrid Forecasting Framework with LLM-Guided Mechanistic Router Spatial Attention for Biochemical Industry Process

    by Yichen Song, Simengxu Qiao, Qunshan He, Zhaoran Liu, Changdi Li, Zijian Tian, Xinggao Liu

    The prediction of biochemical process indicators is a fundamental task in AI4Biochemical, as accurate forecasting improves process transparency and decision-making efficiency. Conventional methods often treat this as a purely data-driven multivariate time series forecasting problem, overlooking the domain-specific insights from mechanistic models. However, relying solely on either mechanistic or time series models limits automation and generalization. To bridge this gap, we propose LMT-Net, a hybrid framework that integrates Large Language Models (LLMs), mechanistic models, and time series models. A key innovation of our framework is the strategic use of the LLM as a mechanistic inference agent to generate a relational graph, rather than as a direct numerical predictor. This design decouples knowledge extraction from prediction and, when paired with our proposed mechanistic router spatial attention block, creates an inherently robust system. This synergy ensures that even if the LLM hallucinates, the data-driven attention mechanism can learn to correct for flawed priors, a capability we validate through simulation experiments. Extensive experiments on a real-world biochemical dataset demonstrate that LMT-Net significantly outperforms state-of-the-art models in prediction accuracy. Furthermore, our analysis confirms the framework's practical feasibility, showing that the LLM agent's one-time inference cost is minimal in both time and budget, overcoming a major barrier for industrial deployment. This combination of accuracy, robustness, and cost-effectiveness establishes LMT-Net as a novel and viable paradigm for integrating LLMs into industrial systems.

  • Toward Accessible AI Research for Complex Batch Biochemical Processes

    by Shifan Chen, Yinhong Zhou, Peiyan Tu, Leiyu Chen, Qunshan He, Simengxu Qiao, He Zhang, Xinggao Liu

    Batch biochemical processes are essential in industrial manufacturing, yet their intrinsic complexity poses major challenges for reliable measurement, monitoring, and modeling. Key issues include data scarcity, privacy constraints, and the absence of standardized preprocessing workflows, all of which hinder the application of deep learning (DL) in such settings. To address these challenges, we propose the Privacy-Preserving Preprocessing Pipeline (P2P)—a practical and reproducible framework designed to enhance the accessibility and robustness of DL for complex batch processes. First, a robust data preprocessing module supports batch-wise segmentation, temporal alignment, and variable selection. Second, a privacy-preserving mechanism generates synthetic datasets that retain essential physical characteristics while protecting sensitive information. Furthermore, we introduce Spatiotemporal Kinetic Interaction Modeling (SKIM), a graph-theoretic approach for identifying and characterizing causal dependencies among process variables, contributing directly to process monitoring and system identification. We construct four practical benchmark datasets and release three open-access datasets with graph structures derived via SKIM. Experimental results demonstrate that P2P significantly improves the fidelity and generalizability of DL-based modeling in batch biochemical systems, paving the way for broader adoption of data-driven methods in instrumentation and measurement. All code and datasets are available at: https://github.com/cynara517/Towards-Accessible-AI-Research-for-Complex-Batch-Biochemical-Processes.

  • DeepFilter: A Transformer-Style Framework for Accurate and Efficient Process Monitoring

    by Hao Wang, Zhichao Chen, Licheng Pan, Xiaoyu Jiang, Yichen Song, Qunshan He, Xinggao Liu

    Effective process monitoring is increasingly vital in industrial automation for ensuring operational safety, necessitating both high accuracy and efficiency. Although Transformers have demonstrated success in various fields, their canonical form based on the self-attention mechanism is inadequate for process monitoring due to two primary limitations: ❶ the step-wise correlations captured by the self-attention mechanism struggle to capture discriminative patterns in monitoring logs due to the lack of semantics of each step, thus compromising accuracy; ❷ the quadratic computational complexity of self-attention hampers efficiency. To address these issues, we propose DeepFilter, which modifies the Transformer for process monitoring. The core modification is replacing the self-attention layer with an efficient filtering layer that excels at capturing long-term discriminative patterns with reduced complexity. Equipped with the efficient filtering layer, DeepFilter enhances both accuracy and efficiency, meeting the stringent demands of process monitoring. Experimental results on real-world process monitoring datasets validate DeepFilterȉs superiority in terms of accuracy and efficiency compared to existing state-of-the-art models.


  • Controllable Mixture-of-Experts for Multivariate Soft Sensors

    by Licheng Pan, Hao Wang, Zhichao Chen, Yuxin Huang, Yunlong Niu, Zhaoran Liu, Qunshan He, Xinggao Liu

    Multivariate soft sensors are critical in industrial manufacturing for providing precise estimations of multiple quality variables and ensuring data reliability and completeness. Existing methods predominantly focus on parameter-sharing architectures while overlooking two fundamental issues: 1) catastrophic interference, where sharing parameters across all tasks leads to performance degradation; 2) uncontrollable optimization, where the optimizer lacks controllability over task priorities, misleading specific tasks converge to unexpected values. To handle these issues and enhance multivariate modeling, we propose a controllable Mixture-of-Experts (ControlMoE) framework for industrial soft sensors, consisting of a Mixture-of-Sequential-Experts (MoSE) module and a Proportional-Integral-Derivative Calibrating (PIDC) module. The MoSE module integrates sequential expert networks with task-specific gating networks, enabling parameter sharing while preserving task distinctions to mitigate catastrophic interference. The PIDC module employs an innovative PID controller to dynamically calibrate task learning weights, ensuring expected outcomes through feedback control and enhancing optimization controllability. Evaluations on two industrial datasets from Chinese nuclear monitoring stations demonstrate that ControlMoE achieves superior accuracy and controllability in multivariate soft sensor modeling. Note to Practitioners—This paper tackles the challenges in applying multivariate soft sensors within industrial manufacturing processes, particularly in environments such as nuclear monitoring where precision and control over multiple quality variables are critical. Traditional methods often suffer from performance degradation due to parameter sharing across all tasks-known as catastrophic interference–and lack control in optimizing multiple tasks simultaneously, leading to suboptimal outcomes. Our proposed framework, ControlMoE, introduces a novel architecture that combines sequential expert networks with a dynamic weight adjustment mechanism using PID controllers. This enables precise control over task-specific optimizations, ensuring that the desired importance and sequence of tasks are maintained, thereby enhancing both accuracy and operational control. We believe this framework could be widely applicable in various industrial settings where multivariate monitoring and control are necessary, improving not only performance but also the reliability of the data produced for critical decision-making.


  • Graph-Based Multiscale Stacked Autoencoder for Biopharmaceutical Processes Soft Sensor

    by Qunshan He, Zhichao Chen, Yichen Song, Changdi Li, Xinggao Liu

    Accurate monitoring of key process variables is essential for biopharmaceutical manufacturing, but the direct measurements are often costly and involve substantial time delays. Data-driven soft sensor methods provide an effective solution by leveraging readily available process data to predict key process variables. However, despite the progress in industrial applications, their implementation in biopharmaceutical processes faces significant challenges, including complex spatiotemporal modeling, oversight of batch-specific characteristics, and seamless fusion of multiscale features. To address these challenges, this paper proposes the Graph-based MultiScale Stacked Autoencoder (GMScaleSAE) framework. GMScaleSAE integrates a graph structure learning module to derive an adjacency matrix to explicitly represent complex interactions among process variables at the process scale. This adjacency matrix is then employed in a spatiotemporal graph convolution module, effectively capturing localized spatiotemporal dynamics at the segment scale. Subsequently, a batch-oriented autoencoder module is developed to account for batch-specific characteristics at the batch scale. Additionally, utilizing stacked autoencoder architecture with layer-wise pre-training and hierarchical structures, GMScaleSAE effectively fuses multiscale features, ensuring stable training and enhanced prediction accuracy. In comparison with five groups of baseline models, GMScaleSAE achieves state-of-the-art performance on both real-world and benchmark datasets, reducing RMSE and MAPE by at least 19.64% and 23.93% respectively. These results highlight the robustness and superior predictive capabilities of GMScaleSAE in process variables prediction.

  • Unified Pest Prevention and Control System Based on AIoT for Sustainable Agriculture

    by Changdi Li, Guangye Li, Yanfeng Wu, Yichen Song, Qunshan He, Zijian Tian, Hu Xu, Xinggao Liu

    Traditional agriculture often relies heavily on pesticides for distributed pest management, which increases production costs and can lead to environmental pollution. The inefficiency and environmental impact of these methods highlight the need for more sustainable solutions. Artificial Intelligence Internet of Things (AIoT) enables precise monitoring and prediction of pest infestations in agriculture, facilitating targeted interventions to effectively reduce crop loss rates. This article proposes a unified pest prevention and control system based on Agricultural Social Internet of Things (Agri-Social IoT) and artificial intelligence (AI) algorithm. By social relationships among devices, the system aims to enhance agricultural production efficiency through unified pest prediction and prevention. To achieve rapid and accurate detection of plant pests, this article introduces an innovative detection method named Ghost-YOLO-ShuffleAttention (GhostYOLOSA). This AI method improves the capture of spatial relationships and contextual information, enhancing the detection accuracy for small targets. It significantly reduces the computational load and model parameters while ensuring high detection accuracy, making it suitable for resource-limited devices. Experimental results from datasets and real-world applications demonstrate that the unified pest control system performs well in pest detection and prevention tasks, promoting sustainable agricultural development.