DLinear KAN Model for Ore Crusher Liner Wear Estimation Based on Motor Power and Vibration History
Authors
Miloslav Sládek, Dušan RychlýFiles
Abstract
Wear of the ore crusher liner alters the crushing-chamber profile and can change energy demand, vibration transmission, and product-size consistency. This study develops a DLinear-KAN estimator that uses 48 h histories of motor power, power variance, vibration root-mean-square value, envelope kurtosis, high-frequency band energy, and a load descriptor to estimate liner wear. The DLinear path retains slowly varying operating information, whereas the KAN path represents nonlinear residual interactions before the two embeddings are fused by a regression head. A constructed single-campaign evaluation contains 1,248 hourly records, six synchronized variables, and 14 inspection-style checkpoints spanning 2.1– 24.7 mm of wear. Using a chronological 70/10/20 training, validation, and test partition, the proposed model achieved a mean absolute error of 1.42 mm, a root means square error of 1.86 mm, and an R2 value of 0.946. At an illustrative 18 mm warning boundary, the threshold error was 4.1% of all test windows. These results support the use of fused electrical and vibration histories as advisory evidence for inspection planning; they do not establish a universal replacement threshold or cross-site deployment performance. The study therefore provides a transparent demonstration of trend-residual fusion under sparse inspection-style labels and identifies the need for validation across liners, crushers, and operating conditions.
