Staged anomaly detection in temperature field data of concrete dams driven by time-varying characteristics

TO ENHANCE THE STAGE ADAPTABILITY AND INTERPRETABILITY OF ANOMALY DETECTION IN TEMPERATURE MONITORING DATA FOR CONCRETE DAMS, THIS STUDY ADDRESSES THE CHALLENGES POSED BY LIMITED DATA AVAILABILITY FOR SMALL AND MEDIUM-SIZED DAMS. A SPARSE AUTO-ENCODER IS EMPLOYED TO EXTRACT LATENT FEATURES FROM MULTI-MODAL DATA, INCLUDING TEMPERATURE AND WATER LEVEL MEASUREMENTS. THE PELT (PRUNED EXACT LINEAR TIME) CHANGE-POINT DETECTION ALGORITHM IS INTEGRATED TO ENABLE ADAPTIVE PARTITIONING OF HYDRATION HEAT, COOLING, AND STABILITY STAGES. BY ANALYZING ENVIRONMENTAL FACTORS AND THEIR INFLUENCE ON THE TEMPERATURE OF CONCRETE DAM BODIES, A HYBRID AUTO-ENCODER MODEL COMBINING CONVOLUTIONAL NEURAL NETWORKS (CNN), BIDIRECTIONAL LONG SHORT-TERM MEMORY NETWORKS (LSTM), AND ATTENTION MECHANISMS IS DEVELOPED. THIS MODEL CAPTURES STAGE-SPECIFIC CHARACTERISTICS OF DAM TEMPERATURE DATA WHILE ACCOUNTING FOR ITS COUPLING RELATIONSHIP WITH ENVIRONMENTAL VARIABLES. A POINT-LEVEL DYNAMIC THRESHOLD MODEL INCORPORATING MULTI-FACTOR FUSION IS DESIGNED, ENABLING THRESHOLD ADJUSTMENTS BASED ON BOTH STAGE-SPECIFIC CHARACTERISTICS AND LOCAL ATTRIBUTES. ADDITIONALLY, STRATEGIES FOR AMPLIFYING ANOMALY SCORES ARE PROPOSED TO MINIMIZE FALSE POSITIVES IN DETECTION OUTCOMES. THE PERFORMANCE OF THE METHOD IS COMPREHENSIVELY EVALUATED THROUGH FEATURE RETENTION TESTS USING REAL-WORLD DATA AND ANOMALY ADDITION TESTS CONDUCTED WITHIN FINITE ELEMENT SIMULATIONS. THE RESULTS DEMONSTRATE THAT THE HYBRID APPROACH ACHIEVES AN AVERAGE FEATURE RETENTION RATE OF 98.57%, WITH A RECALL RATE OF 93.3%, PRECISION RATE OF 77.8%, AND AN F1 SCORE OF 0.848—SIGNIFICANTLY OUTPERFORMING OTHER METHODS. THIS RESEARCH HIGHLIGHTS THE ROBUST PERFORMANCE OF THE PROPOSED ANOMALY DETECTION TECHNIQUE IN "DATA-DEFICIENT" ENVIRONMENTS, PROVIDING A NOVEL PATHWAY FOR MULTIFACTORIAL ANALYSIS OF TEMPERATURE FIELDS IN CONCRETE DAMS.