📄 other

Conditional Diffusion-Based Data Augmentation for Imbalanced Multi-Mode Fault Diagnosis in Predictive Maintenance

This paper demonstrates that class-conditional denoising diffusion probabilistic models (DDPMs) effectively mitigate severe class imbalance in predictive maintenance by synthesizing high-fidelity minority-class samples, significantly outperforming traditional augmentation methods like SMOTE and GANs in detecting rare industrial failures.

Original authors: ABDELHAFID HAMDI ALAOUI, ANWAR MEDDAOUI

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: ABDELHAFID HAMDI ALAOUI, ANWAR MEDDAOUI

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but there's a catch: the crime scene is a massive, bustling city where 97% of the people are perfectly innocent, and only a tiny handful are the actual culprits. If you build a computer program to spot the bad guys, it will quickly learn the easiest trick: just guess "innocent" for everyone. It will be right 97% of the time, earning a perfect score, but it will miss every single criminal. This is the daily struggle of Predictive Maintenance in the modern industrial world. Factories are filled with sensors that constantly watch machines, but machines rarely break. When they do, it's a rare, expensive disaster. The data looks like a giant ocean of "everything is fine" with just a few drops of "something is wrong."

To fix this, scientists use Data Augmentation, which is like a digital photocopier for data. Instead of just waiting for more real breakdowns (which might take years), they try to create fake but realistic examples of failures to teach the computer what a breakdown looks like. For a long time, they used simple tricks like copying existing bad examples or stretching the space between them. But these methods are a bit like trying to learn what a tiger looks like by only looking at a single photo and then drawing a blurry line between it and another photo; you don't really learn the stripes, you just learn the average. Recently, a new type of AI called a Diffusion Model has become famous for creating amazing images. It works by starting with pure static (like TV snow) and slowly, step-by-step, cleaning it up until a clear picture emerges. This paper asks: Can we use this "snow-to-image" magic to create realistic fake machine failures and teach our detective better?

The researchers, Dr. Abdelhafid Hamdi Alaoui and Prof. Anwar Meddaoui, decided to put this idea to the test using a famous, publicly available dataset that simulates an industrial milling machine. They set up a fair fight: they took the tiny handful of real failure examples (only about 3% of the total data) and asked five different AI methods to create 500 new, fake examples for each type of failure. The competitors were: doing nothing, just copying the old data, the "stretching" method (SMOTE), a Generative Adversarial Network (a high-stakes game of cat-and-mouse between two AIs), and a Variational Autoencoder (a method that compresses and rebuilds data). The star of the show was the Conditional Diffusion Model, which acts like a master sculptor, chipping away noise to reveal a new, unique, but realistic failure.

The results were a clear victory for the diffusion model. When they tested these new "fake" datasets on two different types of computer brains (a Multi-Layer Perceptron and a Random Forest), the diffusion model helped them spot the rare failures much better than any other method. Without any help, the computer missed about 40% to 60% of the rare failures. With the diffusion model's help, it started catching up to 76% of them. In fact, the diffusion model was so good that in every single one of the eight different test runs they performed, it beat every other method. The paper explicitly rules out the idea that simply copying data or stretching it between points is enough; those methods failed to create the necessary variety to teach the computer how to spot the tricky, rare breakdowns.

However, there is a trade-off, much like a security guard who is told to catch every single thief. If the guard is hyper-vigilant, they might catch all the thieves, but they might also stop innocent people by mistake. The paper found that while the diffusion model caught more real failures, it also raised more "false alarms" (saying a machine was broken when it wasn't). The authors argue that in the world of predictive maintenance, this is actually a good thing. Missing a real failure costs a factory millions of dollars and can be dangerous, while a false alarm just costs a little bit of time to check the machine. So, the "mistake" of being too cautious is the safer bet.

The study also checked how much effort this took. Creating these fake failures with the diffusion model took about 15 seconds on a standard computer processor, which is only slightly slower than the simpler methods and much faster than the "cat-and-mouse" AI game. The researchers were careful to note that their results come from a simulated dataset, not a real factory floor yet, so while the numbers look fantastic, real-world sensors might be messier. But the core finding stands: by using this "snow-to-image" technique to generate diverse, realistic examples of rare failures, we can train our machines to be much better at spotting the disasters before they happen, turning a blind detective into a sharp-eyed one.

Technical Summary: Conditional Diffusion-Based Data Augmentation for Imbalanced Multi-Mode Fault Diagnosis

Problem Statement
Predictive maintenance (PdM) in Industry 4.0 relies on data-driven models to anticipate equipment failures. However, industrial data is structurally imbalanced: healthy operating conditions generate the vast majority of observations, while specific failure modes are rare and unevenly distributed. This class imbalance causes standard classifiers, optimized for overall accuracy, to bias toward the majority "healthy" class, systematically missing the rare, high-cost failures that PdM systems are designed to detect. Classical resampling techniques like random oversampling (adding no new information) and SMOTE (interpolating between existing points) often fail to capture the complex, multimodal nature of failure mechanisms or create implausible samples in low-density regions. While deep generative models like GANs and VAEs offer alternatives, GANs are notoriously unstable on small tabular datasets, and VAEs tend to produce over-smoothed, low-diversity samples.

Methodology
The authors propose a controlled benchmark to evaluate class-conditional denoising diffusion probabilistic models (DDPMs) as a data augmentation strategy for imbalanced fault diagnosis. The study utilizes the AI4I 2020 Predictive Maintenance dataset, a synthetic benchmark of 10,000 records simulating an industrial milling machine. The problem is formulated as a five-class, single-label fault diagnosis task involving one majority class (No Failure) and four minority failure modes: Tool Wear Failure (TWF), Heat Dissipation Failure (HDF), Power Failure (PWF), and Overstrain Failure (OSF).

The experimental pipeline isolates the augmentation strategy by keeping all other components fixed:

Data Preprocessing: Features are standardized, and categorical variables are one-hot encoded. The dataset is split into 70% training and 30% test sets, with the test set remaining untouched.
Augmentation Strategy: The four minority classes are augmented to a target of 500 samples each. The proposed method uses a conditional DDPM trained on pooled minority data to synthesize novel samples.
Baselines: The diffusion model is compared against:
- No augmentation.
- Random Oversampling (ROS).
- SMOTE.
- Conditional GAN (cGAN).
- Conditional VAE (cVAE).
Evaluation: Two structurally distinct downstream classifiers—a Multilayer Perceptron (MLP) and a Random Forest (RF)—are trained on the augmented data and evaluated on the real test set. The entire process is repeated across eight random seeds to ensure robustness.
Metrics: Performance is assessed on two axes:
- Intrinsic Fidelity: Maximum Mean Discrepancy (MMD) and Classifier Two-Sample Test (C2ST) to measure how well synthetic data mimics the real distribution.
- Extrinsic Utility: Macro-averaged F1, Balanced Accuracy, and specifically Minority Recall (the ability to detect rare failures), which is prioritized as the operational metric.

Key Contributions

Uniform Benchmarking: The paper provides a reproducible framework comparing classical resampling, GANs, VAEs, and diffusion models under identical protocols, targets, and classifiers.
Dual-Axis Evaluation: It evaluates methods not just on downstream classification scores but also on the distributional fidelity of the synthetic data.
Statistical Rigor: Results are validated across eight random seeds with non-parametric significance testing (Wilcoxon signed-rank test), ensuring findings are not artifacts of specific data splits or initializations.
Transparent Trade-off Analysis: The study explicitly analyzes the precision-recall trade-off induced by diffusion augmentation, clarifying its operational implications.

Results

Downstream Performance: The diffusion-based augmentation consistently achieved the highest minority-class recall and balanced accuracy for both the MLP and Random Forest classifiers.
- For the MLP, minority recall improved from 0.61 (no augmentation) to 0.76 with diffusion.
- For the Random Forest, minority recall improved from 0.41 to 0.76.
- These improvements were statistically significant ( $p = 0.008$ ) and consistent across all eight seeds against all five baselines.
Distributional Fidelity: While SMOTE achieved the lowest MMD (due to its interpolation nature), the diffusion model produced the most realistic samples among generative models, evidenced by a C2ST AUC of 0.70 (closer to 0.5, indicating indistinguishability from real data) compared to 0.85 for cVAE and 0.88 for cGAN.
Precision-Recall Trade-off: The high recall of the diffusion model came with a reduction in minority precision (increased false alarms). The authors note this is a deliberate and acceptable trade-off in safety-critical PdM contexts where missing a failure is far costlier than a false alarm.
Per-Class Analysis: Diffusion augmentation improved recall across all failure modes, including the most difficult (TWF), where recall increased from 0.06 to 0.21.
Computational Cost: The diffusion model was computationally efficient, requiring less training time than the cGAN and comparable to the cVAE on a single CPU core.

Significance and Claims
The paper claims that conditional diffusion models are a competitive and reproducible tool for mitigating class imbalance in predictive maintenance, particularly when the operational priority is maximizing the detection of rare failures. The authors argue that the most useful augmentation is not the one that copies training data most faithfully (like SMOTE), but the one that adds informative, plausible diversity to the minority manifold.

The study concludes that diffusion augmentation improves the training distribution itself rather than exploiting quirks of a specific classifier, as evidenced by consistent gains across two different model families. While the dataset is synthetic and the study is limited to tabular data, the authors posit that the approach offers a practical default for new assets, provided the augmentation target is tuned to balance the cost of missed failures against false alarms. Future work is identified as extending this approach to time-series data and real-world field data.

Technical Summary: Conditional Diffusion-Based Data Augmentation for Imbalanced Multi-Mode Fault Diagnosis

More like this