Options
Case Study: AI-Driven Log Extraction and Trace Outlier Detection for Efficient Post-Silicon Validation
Citation Link: https://doi.org/10.15480/882.14870
Publikationstyp
Conference Paper
Date Issued
2025
Sprache
English
Author(s)
Akash, Kowshic A.
Wulf, Tobias
Valentin, Torsten
Jose, John
TORE-DOI
Citation
26th IEEE Latin-American Test Symposium, LATS 2025
Contribution to Conference
Peer Reviewed
true
In the post-silicon validation process, various functionalities and boundaries of a system-on-chip (SoC) are tested, generating a large amount of data in the form of log files, trace data, and oscilloscope images. Log files provide essential information regarding a test run, such as test setup, while trace files offer insights into internal register statuses and sweep parameters like voltage, frequency, and temperature. Manually analyzing and debugging these files is time-consuming, inefficient, costly, and prone to errors. To address these challenges, we propose an AI-powered approach to automatically extract critical log messages from extensive datasets, generating concise log files with only the most pivotal information. Our method utilizes a multi-class Long Short Term Memory (LSTM) neural network. Our primary focus is to minimize false negatives (high recall) to ensure that critical anomalies are not overlooked, thus delivering more reliable SoCs. Simultaneously, we aim to minimize false positives (high precision) to reduce manual debugging efforts. Our proposed method achieves high recall/precision of 94%/99%
for normal, 99%/99% for information, 92%/64% for error, and 98%/88% for warning log categories. Additionally, for outlier detection in trace data, we propose an unsupervised method based on Isolation Forest, which achieves high recall/precision of 95%/100% and 92%/73% for anomalous data points across two distinct datasets, and nearly 100% for normal data points.
for normal, 99%/99% for information, 92%/64% for error, and 98%/88% for warning log categories. Additionally, for outlier detection in trace data, we propose an unsupervised method based on Isolation Forest, which achieves high recall/precision of 95%/100% and 92%/73% for anomalous data points across two distinct datasets, and nearly 100% for normal data points.
Subjects
Post-silicon validation
Anomaly detection
Machine learning
Log extraction
Trace outlier detection
DDC Class
621.3: Electrical Engineering, Electronic Engineering
Publication version
publishedVersion
Loading...
Name
Akash_LATS2025.pdf
Size
662.01 KB
Format
Adobe PDF