← Back to Sessions

Enhancing Classification Reliability With Anomaly Detection for Operational Monitoring of Continuous Seismic Data

Description: 

With the ever-growing amount of seismic data collected, the role of artificial intelligence (AI) for processing and analyzing those recordings becomes indispensable. Recent years have seen a surge in seismic detection and classification applications. However, the robustness of classification models poses a challenge: the complexity of continuous incoming stream of data is never fully represented in a training set, and a classification model does not have the ability to detect unexpected behavior and to say “I do not know”. Consequently, it will erroneously classify unknown samples in one of the predefined classes. In current work, we improve the reliability of seismic data detection and classification in two different aspects. Using a 32-month seismic dataset from 43 French stations, identifying six classes, a single convolutional neural network (CNN) analyzes the entire multi-stational dataset. Firstly, we explored strategies to mitigate bias in a multi-class classification model due to class-size imbalances. Best results are obtained by optimizing the loss function regarding class-size, enhancing balanced accuracy by 47%. Secondly, we propose the implementation of an anomaly detection module to mitigate the problem of misclassification of unknown samples by providing the model the opportunity to say “I do not know where this sample belongs”. We considered four semi-supervised anomaly detection algorithms, namely one-class SVM, elliptic envelope, isolation forest and local outlier factor. Our findings reveal that incorporating an anomaly detection module effectively mitigates the misclassification of most novelties as predefined classes. Furthermore, it allows for detecting unexpected behaviors such as anomalies, class drifts and ambiguous samples. Thereby enhancing the trustworthiness and reliability of predictions. Our study indicates that the isolation forest method yields best results for our specific application.

Session: Special Applications in Seismology - I

Type: Oral

Date: 5/1/2024

Presentation Time: 05:15 PM (local time)

Presenting Author: Marielle

Student Presenter: No

Invited Presentation: 

Authors