Research On Temporal Anomaly Detection Methods Based On Adversarial Training And Frequency Domain Enhanced Attention Mechanism

Our country’s industrial development has been rapid,accompanied by the gradual popularization of intelligent industrial production.The industrial sector is the most widely applied field for time series data anomaly detection.With the assistance of intelligent sensors,production parameters and indicators of industrial production equipment are continuously monitored,which helps monitor and analyze the parameters collected by the sensors.It enables early detection of anomalies in the operational status of industrial equipment and ensures their normal functioning.In recent years,with the rapid development of deep learning,the number of solutions in the field of time series data anomaly detection has increased.However,existing methods mostly use recurrent neural networks or convolutional neural networks to model the features of time series data.These models are limited in their ability to extract temporal relationships in long sequences and have poor robustness,making them prone to overfitting.They easily lead to misjudgment when the distribution of the data to be tested differs slightly from the training data distribution.Additionally,most methods do not consider the impact of temporal data noise on anomaly detection,making it difficult to distinguish between anomalies and noise,thereby limiting the success rate of anomaly detection.Furthermore,real-world industrial production environments are complex,and sensor data transmission is prone to interference.The collected data often suffers from data quality issues and frequently contains missing values,which greatly affect subsequent anomaly detection tasks.To address the aforementioned problems,corresponding solutions have been proposed.Firstly,to tackle the issue of missing values that frequently occur in industrial scenarios,the model is improved based on the generative adversarial network.By introducing a bidirectional long short-term memory network,the time series data features are extracted as comprehensively as possible.Additionally,an attention mechanism is applied to the hidden memory units of the long short-term memory network to enhance the model’s ability to model long sequences and improve its performance.The thesis manually constructs different proportions of anomalous data in two time series datasets for comparative experiments,and the results demonstrate the superior performance of the model in missing value imputation.Secondly,to address the problem of the vulnerability of time series anomaly detection models to noise and their poor robustness,an improved anomaly detection model is constructed based on a simplified Trans Former architecture.This model combines the advantages of adversarial training and attention mechanisms with frequency domain enhancement.The use of wavelet decomposition and Fourier transform in the frequency domain enables the model to better withstand the influence of noise on detection performance.Comparative experiments and ablation experiments are conducted on four anomaly detection datasets,comparing the model with several state-of-the-art baseline models,and the effectiveness of the model is demonstrated.