Publicação
Anomaly detection in gas sensor data using LSTM autoencoder and latent space analysis
| Resumo: | Anomaly detection in gas sensor data is crucial for food quality control, environmental monitoring, and industrial safety, yet traditional supervised approaches require labeled anomalous data that is often impossible to obtain. This paper presents a single-class LSTM autoencoder for BME688 gas sensor anomaly detection using latent space distance analysis. Training exclusively on normal samples (Anis estrellado), the model detects anomalies by measuring Euclidean distances in the learned 8-dimensional latent space. Treating the 10-step heater profile as a temporal sequence enables the capture of sequential dependencies in gas resistance patterns. Evaluation across seven compounds achieves 100% detection for olive oil and 50.4% for air while maintaining false positive rates at or below 5% for normal classes (coffee: 0.0%, tea: 0.4%, cocoa: 5.0%). Compared to reconstruction-based methods, our approach provides 3.7× better separation, faster inference (6.8ms vs 12.3ms), and improved interpretability, offering an efficient solution for real-time anomaly detection where only normal operational data is available. |
|---|---|
| Autores principais: | Ahmadi, Mahdia |
| Outros Autores: | Izidorio, Felipe; Igrejas, Getúlio; Rodrigues, Pedro João; Lopes, Rui Pedro |
| Assunto: | anomaly detection LSTM autoencoder gas sensors BME688 latent space analysis single-class learning timeseries |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | comunicação em conferência |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Instituto Politécnico de Bragança |
| Idioma: | inglês |
| Origem: | Biblioteca Digital do IPB |
| Resumo: | Anomaly detection in gas sensor data is crucial for food quality control, environmental monitoring, and industrial safety, yet traditional supervised approaches require labeled anomalous data that is often impossible to obtain. This paper presents a single-class LSTM autoencoder for BME688 gas sensor anomaly detection using latent space distance analysis. Training exclusively on normal samples (Anis estrellado), the model detects anomalies by measuring Euclidean distances in the learned 8-dimensional latent space. Treating the 10-step heater profile as a temporal sequence enables the capture of sequential dependencies in gas resistance patterns. Evaluation across seven compounds achieves 100% detection for olive oil and 50.4% for air while maintaining false positive rates at or below 5% for normal classes (coffee: 0.0%, tea: 0.4%, cocoa: 5.0%). Compared to reconstruction-based methods, our approach provides 3.7× better separation, faster inference (6.8ms vs 12.3ms), and improved interpretability, offering an efficient solution for real-time anomaly detection where only normal operational data is available. |
|---|
Atividades financiadas
Carregando projetos financiados...