Unsupervised Anomaly Detection in Process-Complex Industrial Time Series: A Real-World Case Study

Abstract: Industrial time-series data from real production environments exhibits substantially higher complexity than commonly used benchmark datasets, primarily due to heterogeneous, multi-stage operational processes. As a result, anomaly detection methods validated under simplified conditions often fail to generalize to industrial settings. This work presents an empirical study on a unique dataset collected from fully operational industrial machinery, explicitly capturing pronounced process-induced variability. We evaluate which model classes are capable of capturing this complexity, starting with a classical Isolation Forest baseline and extending to multiple autoencoder architectures. Experimental results show that Isolation Forest is insufficient for modeling the non-periodic, multi-scale dynamics present in the data, whereas autoencoders consistently perform better. Among them, temporal convolutional autoencoders achieve the most robust performance, while recurrent and variational variants require more careful tuning.

• The paper demonstrates that TCN-AE significantly outperforms classical models (F1-score ~0.991) in detecting anomalies.
• It systematically evaluates multiple autoencoder variants, highlighting the impact of architectural inductive biases on detection stability.
• The study provides actionable insights for industrial deployments by aligning model choice with process-induced data complexity.

Unsupervised Anomaly Detection in Process-Complex Industrial Time Series: An Expert Analysis

Dataset Complexity and Study Scope

The paper addresses anomaly detection in industrial time-series, leveraging a proprietary dataset collected from over 100 machines operating in production environments. Unlike benchmark datasets characterized by periodicity or monotonic degradation, this dataset contains pronounced process-induced complexity: sensor heterogeneity, non-periodicity, multi-scale dynamics, and variable phase orderings. Such characteristics pose significant challenges for anomaly detection due to non-stationarity and noise inherent in realistic machine health monitoring scenarios.

A representative sample from the dataset emphasizes the diversity and non-repeatability of sensor behavior during a process execution, showing stable operational phases, gradual drifts, and abrupt transitions frequently associated with anomalies or process changes:

Figure 1: Plot of four sensor readings exemplifying variability, drift, and abrupt transitions during a single production cycle.

Model Architectures and Experimental Methodology

The research evaluates the anomaly detection capabilities of both classical and deep learning models:

• Classical Baseline: Isolation Forest (IF), selected for its simplicity and scalability, was tasked with detecting anomalies from aggregated sequence-level features.
• Representation-Learning Models: Six autoencoder-based architectures, covering deterministic and variational approaches:
  • Temporal Convolutional Network Autoencoder (TCN-AE, TCN-VAE)
  • Long Short-Term Memory Autoencoder (LSTM-AE, LSTM-VAE)
  • Gated Recurrent Unit Autoencoder (GRU-AE, GRU-VAE)

The models were trained exclusively on normal operation data and subjected to rigorous hyperparameter optimization. Evaluation followed a two-stage protocol: first, assessing reconstruction fidelity; second, benchmarking anomaly detection on withheld, manually labeled instances encompassing real failures and irregular behaviors.

Quantitative Results and Performance Analysis

Classical IF was categorically insufficient, yielding an average F1-score of $0.12 \pm 0.13$; in most runs, it missed the majority of anomalies and exhibited high variance. In contrast, all autoencoder variants significantly surpassed the baseline. TCN-AE achieved the highest and most consistent anomaly detection performance with an average F1-score of $0.991 \pm 0.009$, frequently detecting all anomalies without false positives—a claim substantiated with strong numerical evidence.

Recurrent variants (LSTM-AE, GRU-AE) were competent but less robust to varying process dynamics, displaying higher sensitivity to hyperparameter selection and training stability. Among variational approaches, TCN-VAE and LSTM-VAE performed well but systematically underperformed their deterministic counterparts, with GRU-VAE showing the most pronounced drop.

The F1-score distributions across architectures further reinforce the superior performance of TCN-based models:

Figure 2: F1-score distributions for top 5 configurations per model, highlighting consistent superiority of TCN variants.

Notably, standard autoencoders consistently outperformed VAEs. Stochasticity from variational inference introduced detrimental artifacts in reconstruction, decreasing threshold stability for anomaly classification.

Architectural Implications and Robustness Considerations

The convolutional inductive bias of TCN architectures provides intrinsic robustness to multi-scale, non-periodic dynamics, outperforming recurrent models that struggle with variable phase lengths and process ordering. This alignment with process structure proved critical in process-complex datasets, indicating that architectural choices supersede raw model complexity in efficacy.

These findings offer practical guidance for model selection in industrial deployment scenarios: convolutional autoencoders, particularly TCN-based, deliver superior stability and robustness, minimizing performance variability across configuration space. The empirical evidence supports a strong architectural preference in favor of temporal convolution for unsupervised industrial anomaly detection.

Limitations and Future Directions

The case study focuses on a single proprietary dataset, constraining direct comparison with public benchmarks and limiting generality. Additionally, only classical and autoencoder-based models are systematically evaluated; transformer-based and attention architectures, while promising, require further study given industrial constraints on training data size, complexity, and resource availability.

Future research should address broader generalization by validating convolutional autoencoder architectures across diverse process-complex industrial datasets and exploring hybrid temporal models incorporating attention mechanisms for enhanced sequence modeling without forfeiting computational practicality.

Conclusion

This paper systematically compares classical and autoencoder-based anomaly detection architectures in real-world industrial time-series with process-induced complexity. Feature-agnostic models are categorically inadequate for such settings; convolutional autoencoders, specifically TCN-AE, demonstrate marked improvement in anomaly detection, substantiated by consistently high F1-scores. The evidence underscores the importance of architectural inductive alignment with temporal process structure and provides actionable recommendations for practical industrial AI deployments. Further advances will require complexity-aware evaluations across broader datasets and architectural innovations tailored to industrial constraints and variability.