基于FEDformer-MTFA的轻工装备多源时序数据异常检测OA

Under the strategy of industrial intelligence and the digital transformation of manufacturing,traditional frequency-domain models often exhibit insensitivity to localized anomalies such as sudden temperature spikes,while direct concatenation of multi-source data leads to the loss of spatiotemporal correlations.To address these issues,a multi-source time-series anomaly de-tection model for light industrial equipment based on an improved Frequency Enhanced Decomposition Transformer(FED-former)model,termed FEDformer-MTFA,is proposed.The model incorporates a multi-scale time-frequency joint attention mech-anism,which integrates time-domain features at different temporal scales with multi-band frequency-domain information,combined with an adaptive weight allocation strategy.This approach overcomes the limitations of single-scale frequency-domain attention in traditional models like FEDformer.The mechanism enables the simultaneous capture of dynamic characteristics of equipment op-erating states across multiple temporal granularities and frequency distributions,thereby significantly enhancing sensitivity and representational capability towards weak transient anomalous signals.Additionally,the model employs a multi-task detection head design,simultaneously predicting reconstruction errors and anomaly probability distributions.By dynamically balancing the optimi-zation weights of the two tasks through adjustable hyperparameters,it effectively mitigates overfitting caused by a single optimiza-tion objective,thereby improving detection robustness.Furthermore,to tackle sample imbalance,the model incorporates a joint weighted loss function that dynamically adjusts sample weights,focusing training on potential anomaly regions and further optimi-zing detection performance under complex working conditions.Experimental results show that there are 0.186 0 and 0.298 2 im-provements in accuracy and AUC,respectively,compared to FEDformer.