A multi-class anomaly signal classification method for fixed-point deformation combining PSO–VMD multiscale adaptive sample entropy and ALO–SOM model

To address the challenge of identifying five predefined categories of disturbance-related abnormal signals (lightning interference, artificial blasting, power-supply failure, collapse-related events, and meteorological anomalies) in fixed-point deformation monitoring, this paper proposes an identification framework integrating parameter-optimized variational mode decomposition (VMD), multiscale entropy analysis, and an intelligent classification model. Firstly, particle swarm optimization is adopted to adaptively determine the mode number and penalty factor for VMD to realize adaptive signal decomposition. Then, an adaptive sample entropy algorithm with self-adjustable similarity thresholds is applied to extract multiscale complexity-oriented features from each decomposed modal component. Finally, the ant lion optimizer optimizes the competitive-layer topology and training hyperparameters of a self-organizing map network to build the target classifier. Comparative experiments against 18 widely-used classification models (including AdaBoost ensemble learners, generalized regression neural networks, long short-term memory networks, and gated recurrent units) are carried out on the real-world observation dataset from the Yixian seismic station. The experimental results show that the proposed framework achieves competitive classification performance for the five anomaly categories. The mean test-set accuracy over 100 independent repeated-run experiments reaches 99.62%, with a standard deviation of 0.6361. This study offers an effective technical reference for identifying deformation-related abnormal waveforms under complex observational conditions.