ProtoX-AD: Self-Explainable Time Series Anomaly Detection and Characterization

In today’s data-rich environment, the ability to detect anomalies in time series data is crucial across industries such as finance, manufacturing, and healthcare. Traditional statistical methods and shallow machine learning models often fall short when faced with complex, nonlinear patterns inherent in real-world data. Deep learning approaches, especially those leveraging self-supervised learning, have recently gained prominence due to their superior detection capabilities. These models typically learn to recognize normal patterns and flag deviations as anomalies, often using transformation-based tasks to generate surrogate labels. However, despite their high performance, many of these models operate as black boxes, providing limited insight into the nature of detected anomalies.

This opacity hampers trust and hinders actionable insights, especially in high-stakes domains where understanding the root cause of anomalies is vital. To bridge this gap, the paper introduces ProtoX-AD, a novel framework that combines the strengths of self-supervised learning with interpretability through prototype-based explanations. The core idea is to learn transformation-aware latent representations that are structured around multiple interpretable prototypes, each capturing distinct normal or anomalous behaviors induced by specific transformations.

ProtoX-AD’s architecture is composed of five interconnected modules. The transformation module applies diverse transformations—both manually designed based on domain knowledge and neural network-based—to generate augmented views of normal samples. The feature extraction module employs a variational autoencoder (VAE) to encode these views into a structured latent space, where each class is associated with multiple prototypes initialized via K-means clustering. The dual reconstruction module ensures semantic fidelity by reconstructing both the augmented views and original samples, preserving interpretability. The prototype module computes similarity matrices between latent representations and prototypes, which are then used by a linear classifier to perform self-supervised anomaly detection.

Training involves optimizing a composite loss function that balances classification accuracy, reconstruction fidelity, and prototype regularization. Once trained, the model evaluates new samples by measuring the cross-entropy loss of their identity view, with higher scores indicating potential anomalies. Importantly, the learned prototypes can be decoded into human-interpretable concepts, enabling visual explanations of detected anomalies. This approach not only maintains detection performance comparable to black-box models but also provides meaningful insights into the nature of anomalies, such as identifying specific abnormal patterns or behaviors.

Extensive experiments on synthetic datasets and real-world applications—including the UMD dataset, global temperature anomalies, and water leak detection—demonstrate that ProtoX-AD achieves state-of-the-art results. Its AUROC scores average around 0.92, surpassing traditional methods and existing explainable baselines. The interpretability is validated through qualitative visualizations and quantitative metrics, with explanation errors consistently below 0.05. Moreover, the systematic analysis of transformation design reveals that tailored transformations significantly enhance both detection accuracy and explanation quality.

This research marks a significant step forward in making deep anomaly detection models more transparent and trustworthy. By integrating transformation-aware latent representations with prototype-based explanations, ProtoX-AD offers a comprehensive solution that addresses both performance and interpretability. Its ability to characterize different types of anomalies semantically is particularly valuable for critical applications requiring human-in-the-loop decision-making. Looking ahead, future work will explore automating transformation design, improving scalability for real-time deployment, and extending the framework to handle point anomalies and multivariate data, further broadening its impact across diverse domains.