TEMPORAL DEEP LEARNING FOR JOB COMPETENCY DEVIATION DETECTION IN COAL CHEMICAL VOCATIONAL TRAINING

Abstract

Conventional assessment in coal chemical vocational training relies heavily on final scores and provides limited information about how operational competency evolves throughout the training process. To address this limitation, this study proposes a temporal deep-learning approach for detecting job-competency deviations from process-level training data. A multidimensional competency framework is established to characterize standardized operation, process regulation, abnormality recognition, fault handling, and safety and emergency response. Time-stamped operating events, action intervals, equipment states, and key process variables are organized into multivariate temporal sequences. A BiLSTM-Attention model is then developed to extract dynamic behavioral representations and estimate competency states, while competency deviations are quantified by comparing the predicted states with job-oriented reference levels. The proposed model outperforms Random Forest, XGBoost, LSTM, GRU, and Transformer baselines, achieving an R2 of 0.879 and an F1-score of 0.876. Dimension-specific analysis indicates relatively stable performance in standardized operation and safety-related tasks, whereas larger gaps occur in abnormality recognition and fault handling. Attention analysis further identifies alarm occurrence, critical process adjustment, and fault-handling stages as important contributors to competency estimation. The framework demonstrates the potential of process-level temporal data for fine-grained, data-driven competency diagnosis in coal chemical vocational training.