Enhancing Power System Component Classification with Transformer-Based Text Embeddings and OCR

Power system analysis and studies rely on good data collection and model development. Data collection and model development typically takes 50 % of the time allotted for power system studies. Classifying images and engineering data collected from various equipment and facilities is traditionally a labor-intensive and time-consuming task. This paper proposes an automated approach to streamline this classification process using advanced optical character recognition (OCR) and text embedding techniques. We employ natural scene text recognition for images captured with a phone camera at a waste-to-energy facility, utilizing Tesseract OCR with the LSTM (Long Short-Term Memory) engine for precise text extraction. The extracted text is then converted into features using the Sentence Transformer model-a state-of-the-art text embedding technique that captures the semantic meaning of the text. These features are subsequently classified using a Support Vector Classifier (SVC). This paper highlights the role of transformer-based embeddings in understanding text semantics and enhancing image classification accuracy in power system studies.