A Systematic Approach for Evaluating Artificial Intelligence Models in Industrial Settings

Paul-Lou Benedick, Jérémy Robert, Yves Le Traon

TL;DR

This paper introduces a systematic approach to evaluate the robustness of AI models in industrial settings by artificially introducing realistic data perturbations. They found that model robustness varies significantly, and predicting this robustness based on dataset characteristics alone is challenging, highlighting the need for their systematic evaluation method.

What Problem Does This Paper Address?

The increasing adoption of AI in industry, particularly for analyzing sensor data in Industrial IoT (IIoT) environments, promises significant improvements in efficiency and decision-making. However, AI models, even those with high accuracy on clean, isolated datasets, can perform poorly when deployed in real-world conditions. This degradation is often due to "perturbed" or "noisy" data that can arise from various sources, such as aging sensors, network issues, or data processing bottlenecks. Existing research often evaluates AI models on clean benchmark datasets, failing to account for these real-world data quality challenges. This gap leaves a critical question unanswered: how reliable are these AI models when faced with imperfect data in an industrial setting?

Methods

The core contribution of this paper is a systematic approach for evaluating AI model robustness under perturbations. This approach involves:

  1. Artificial Perturbation Injection: They propose a method to artificially introduce two specific types of perturbations that are common in industrial data collection:
  2. Swapping Perturbation: Alters the order of data points within a time series. This is relevant in scenarios where data packets might be reordered during transmission (e.g., using UDP).
  3. Dropping Perturbation: Simulates missing data points in a time series. This can occur due to network losses or buffer overflows.
  4. These perturbations are parameterized (e.g., percentage of swapped/dropped data, range of swapping) to allow for systematic and reproducible testing.
  5. Model Evaluation: The robustness of seven state-of-the-art Machine Learning (ML) and Deep Learning (DL) algorithms is evaluated on these perturbed datasets.
  6. Predictability Analysis: The paper investigates whether the impact of these perturbations on model accuracy can be predicted using decision trees based on dataset characteristics.

Datasets

The study utilizes 20 univariate time-series datasets from the public University of California Riverside (UCR) repository. These datasets were chosen because they resemble sensor data commonly found in industrial scenarios. No new benchmarks were introduced; instead, the focus was on applying a novel evaluation methodology to existing, well-known datasets.

Key Insights and Ideas

  • Real-world data is messy: The assumption that AI models trained on "clean" data will perform reliably in production is flawed. Industrial environments are inherently prone to data quality degradation.
  • Robustness is not uniform: Different AI models exhibit vastly different levels of robustness to the same types of data perturbations. Some models are highly sensitive, while others remain relatively stable.
  • Predicting robustness is hard: Simply knowing the characteristics of a dataset (like its length or correlation) is insufficient to reliably predict how an AI model will perform under specific perturbations. This suggests that a "one-size-fits-all" approach to model selection based on dataset properties is not viable.
  • Mitigation matters: Even simple data imputation techniques (like linear regression for dropped values) can significantly improve model robustness, highlighting the importance of considering data preprocessing and handling strategies from the outset.

Key Results

  • Significant Robustness Disparities: The experimental study revealed a "great disparity" in the robustness of the evaluated ML/DL algorithms. Some models, like MLP, MCDCNN, and CNN, showed relatively limited accuracy drops (around 3-5%) even under significant swapping perturbations. In contrast, ResNet and FCN were much more susceptible, experiencing accuracy losses of up to 30% in worst-case scenarios.
  • Dropping Perturbations Less Impactful (with mitigation): With the proposed linear regression-based reconstruction mechanism, dropping perturbations had a less severe impact on model accuracy compared to swapping perturbations.
  • Decision Tree Predictability Limited: When attempting to predict model impact using decision trees based on dataset characteristics, the accuracy was generally low, especially for deep learning models. For swapping perturbations, decision tree accuracy ranged from 12% to 89%, with deep learning models performing poorly. Dropping perturbations showed better predictability (up to 100% accuracy for some models), attributed to the lower impact of these perturbations.

Limitations and Future Work

  • Limited Perturbation Types: The study focused on only two types of perturbations: swapping and dropping. Future work should explore other realistic perturbations relevant to industrial settings.
  • Parameter Range: The ranges and percentages of perturbations used were chosen based on practical considerations but could be widened for a more exhaustive analysis.
  • Uniform Distribution: The perturbations were generated using a uniform distribution, which might not reflect all real-world scenarios. Adapting to different distributions could be beneficial.
  • Predictability Model Size: The decision trees used for predictability analysis were trained on a limited set of datasets and algorithms. Expanding this dataset could lead to more robust predictive models.
  • Explainable AI: The difficulty in predicting robustness for deep learning models opens avenues for research in explainable AI (XAI) to better understand why certain models are more robust than others.

Summary updated

Abstract

Artificial Intelligence (AI) is one of the hottest topics in our society, especially when it comes to solving data-analysis problems. Industry are conducting their digital shifts, and AI is becoming a cornerstone technology for making decisions out of the huge amount of (sensors-based) data available in the production floor. However, such technology may be disappointing when deployed in real conditions. Despite good theoretical performances and high accuracy when trained and tested in isolation, a Machine-Learning (M-L) model may provide degraded performances in real conditions. One reason may be fragility in treating properly unexpected or perturbed data. The objective of the paper is therefore to study the robustness of seven M-L and Deep-Learning (D-L) algorithms, when classifying univariate time-series under perturbations. A systematic approach is proposed for artificially injecting perturbations in the data and for evaluating the robustness of the models. This approach focuses on two perturbations that are likely to happen during data collection. Our experimental study, conducted on twenty sensors’ datasets from the public University of California Riverside (UCR) repository, shows a great disparity of the models’ robustness under data quality degradation. Those results are used to analyse whether the impact of such robustness can be predictable—thanks to decision trees—which would prevent us from testing all perturbations scenarios. Our study shows that building such a predictor is not straightforward and suggests that such a systematic approach needs to be used for evaluating AI models’ robustness.
BibTeX
@inproceedings{benedick2021systematic,
  author = {Paul-Lou Benedick and Jérémy Robert and Yves Le Traon},
  title = {{A Systematic Approach for Evaluating Artificial Intelligence Models in Industrial Settings}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=SNgWmfa3Qgc},
}
Figure