<p>Industrial automation systems generate vast amounts of log data that are crucial for monitoring operations, detecting faults, and ensuring system reliability. These logs often contain context-dependent patterns and irregular structures that make real-time anomaly detection challenging. Traditional anomaly detection methods often struggle with unstructured log data, high-dimensional feature spaces, and real-time processing constraints. This paper proposes a novel anomaly detection framework that combines bidirectional encoder representations from transformers (BERT) with edge sensor computing for efficient and accurate log analysis. Unlike conventional BERT-based models that rely on centralized inference, our method supports lightweight deployment at the edge to enable in-situ analysis. BERT is utilized to extract contextual embeddings from log entries, enabling the model to understand complex system behaviors and detect subtle anomalies. To enhance computational efficiency and reduce latency, edge computing is integrated, allowing real-time anomaly detection at the sensor level. We evaluate the system using an industrial log dataset collected from a smart factory setting, which includes timestamped event sequences and known fault cases. Experimental results highlight improvements in detection accuracy, processing speed, and system scalability over traditional baselines. This work demonstrates how transformer-based models and edge inference jointly improve the reliability of industrial monitoring in smart manufacturing environments.</p>