针对故障诊断系统的数据潜在特征过滤防御策略OA

Data-driven fault diagnosis models are widely applied in modern industry,significantly enhancing the accuracy of fault diagnosis systems.However,data-driven fault diagnosis models are vulnerable to adversarial attacks,that is minor disturbances on samples can easily lead to incorrect prediction results output by the model.The existing defense methods mainly focus on the fields of images,sounds and texts,and are difficult to effectively deal with the adversarial attack problems in industrial fault diagnosis systems.A defense method based on an auto-encoder is proposed to resist adversarial perturbations by learning the latent classification features of samples.An autoencoder structure is introduced to extract the potential classification features of samples through the adversarial learning mechanism of the encoder and the attack discriminator in the encoding stage.In the decoding stage,an ad-versarial learning mechanism is formed by using the decoder and the sample discriminator to restore the sample to a clean one.Through experiments on the Tennessee Isman dataset,it can be known that the diagnostic accuracy rate of the fault diagnosis system has increased by 46.49 percentage points when facing eight types of attack disturbanc-es,verifying the effectiveness and applicability of the proposed method,and providing new ideas and directions for improving the security of the fault diagnosis system.