An unsupervised rockfall monitoring method based on memory-augmented network

By using unsupervised learning-based memory enhancement networks, the problem of detecting rapid movement and short-term anomalies of small targets in rockfall monitoring was solved, achieving efficient and reliable rockfall monitoring, reducing false alarm rates and ensuring the long-term stability of the model.

CN122156719APending Publication Date: 2026-06-05CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHINA UNIV OF GEOSCIENCES (WUHAN)
Filing Date
2026-01-28
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

Existing rockfall monitoring methods struggle to effectively monitor the rapid movement of small targets when labeled data is scarce, and they lack robustness to short-term anomalies, resulting in high rates of missed detections and false alarms.

Method used

An unsupervised learning paradigm is adopted, and an anomaly detection model is constructed by using a memory-enhanced network. The network is trained using a normal mountain monitoring video dataset. By combining motion weighting and dual-threshold temporal gating strategies, the sensitivity of monitoring rapid movement of small targets is enhanced and memory bank contamination is prevented.

Benefits of technology

It achieves highly sensitive monitoring of rapid movement of small targets even when labeled data is scarce, reduces false alarm rate, and ensures the long-term stability and robustness of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122156719A_ABST
    Figure CN122156719A_ABST
Patent Text Reader

Abstract

The application discloses a kind of based on memory enhancement network's unsupervised rockfall monitoring method, it is related to rockfall monitoring technical field, mainly includes: the memory enhancement network and total loss function including encoder, memory module, motion weighting module and decoder are constructed, memory enhancement network is trained using total loss function and normal mountain monitoring video dataset to obtain rockfall monitoring model;The video frame sequence after pre-processing is predicted using the rockfall monitoring model to obtain target prediction frame, according to this, calculate abnormal score, and dynamically update memory bank using double threshold time series gating strategy, according to target mountain monitoring video, pixel-level reconstruction error and abnormal score generate multi-modal monitoring result.The implementation based on memory enhancement network's unsupervised rockfall monitoring method provided in the application can improve the monitoring sensitivity of small target fast motion and enhance the robustness of short-time anomaly under the condition of labeled rockfall data scarcity.
Need to check novelty before this filing date? Find Prior Art