/Awesome_Data_Drift_and_Model_Adaptation_in_Industrial_Settings

Review: Data Drift and Model Adaptation in Industrial Settings

Data Drift and Model Adaptation in Industrial Settings

This repo contains the paper list and figures for A Survey of Data Drift and Model Adaptation in Industrial Settings.

Abstract

总体框架~~~

Scope and rationales

The scope of this survey is mainly defined by following aspects.

  • xx

Citation

@article{chen2024a,
    title = {A Survey of Data Drift and Model Adaptation in Industrial Settings},
    author = {Chen, Jiao and Liu, Qianmiao and Dai, Suyan and He, Jiayi
    and Lv, Zuohong},
    journal={arXiv preprint arXiv:240x.xxxxx},
    year = {2024}
}

Table of Contents

  • Industrial Applications

    • Industrial Fault Diagnosis
    • Remaining Life Prediction
    • Laser Micro/Nano Processing
    • Planing and Control for Autonomous Driving
    • Biomorphic Robotic Motion Control
    • Additive Manufacturing Process Monitoring and Control
    • Gesture Recognition Based on Hydrogel Electronic Skin

  • Definition of data drift

  • [Short-term Drift Adaptation Strategies](#Short-term Drift Adaptation Strategies)

    • Before Deployment
    • After Deployment

  • [Long-term Drift Adaptation Strategies](#Long-term Drift Adaptation Strategies)

    • Continual/Lifelong Learning
    • Learn from Model

  • Challenges and Directions

    • Framework/Platform Development
    • Datasets/Benchmarks
    • Integration with Large Models
    • Multi-Model Management
    • Knowledge Base Construction

Industrial Applications

Industrial Fault Diagnosis

（工业故障诊断）

Remaining Life Prediction

（剩余寿命预测）

Laser Micro/Nano Processing

（激光微纳加工)

Planing and Control for Autonomous Driving

（自动驾驶规划与控制，e.g., 自动清扫车）

Biomorphic Robotic Motion Control

（仿生机器人运动控制）

Additive Manufacturing Process Monitoring and Control

（增材制造过程监控及控制）

Gesture Recognition Based on Hydrogel Electronic Skin

（水凝胶电子皮肤应用、e.g., 手势识别）

Definition of data drift

The goal is to present a simple and intuitive overview of the definition, types, and case studies of data drift. Introducing related concepts: Out-Of-Distribution (OOD), Long-Tail Distribution, Non-IID, and Few-Shot Learning.

TODO： 给出一个示意图，代表不同的检测方法在cifar10数据集的检测效果。

Short-term Drift Adaptation Strategies

Before Deployment

Concept Drift Adaptation by exploiting Drift Type.[ACM Transactions on Knowledge Discovery from Data 2023][[paper](Concept Drift Adaptation by Exploiting Drift Type | ACM Transactions on Knowledge Discovery from Data)]

Transfer Learning

Domain Adaptation

Federated Learning

Flower: A friendly federated learning research framework. [arXiv'20] [Paper] [Code]

Fedml: A research library and benchmark for federated machine learning. [arXiv'20] [Paper] [Code]

Knowledge Distillation

After Deployment

[Serving on Edge]

Parameter-Efficient

e.g., Prompt Adapter, LoRA, Prefix tuning

Resource-Efficient

e.g., Edge-Cloud Collaboration (by Chen), Test-Time Adaptation, Transfer Learning

EdgeFM: Leveraging Foundation Model for Open-set Learning on the Edge. [SenSys'23] [Paper]

Data-Efficient

e.g., Data Selection Mechanisms Based on Gradients, Entropy, Data Annotation (Language Models, Pseudo-Label Generation)

Test Time Adaptation

Gradual Domain Adaptation

[Serving on Cloud]

Edge Cloud Collaboration

Towards Edge-Cloud Collaborative Machine Learning: A Quality-aware Task Partition Framework.[paper]

ECLM: Efficient Edge-Cloud Collaborative Learning with Continuous Environment Adaptation.[paper]

Long-term Drift Adaptation Strategies

Continual/Lifelong Learning

Memory efficient continual learning with transformers. [NeurIPS'22] [Paper]

Regularization Techniques for Model Adaptation

Parameter Isolation

Replay Methods

Analytic Learning

Learn from Model

Model reuse

Meta learning

Challenges and Directions

Framework/Platform Development

Datasets/Benchmarks

Integration with Large Models

Multi-Model Management

Knowledge Base Construction