MPHLDAE‐1DCNN: A Novel Denoising Method for Improved Fault Diagnosis

Fault diagnosis of rotating machines has undergone significant advancements through the use of deep learning models. However, the effectiveness of these models is often compromised by noisy raw vibration data collected from industrial machines, which can negatively impact accuracy rates. To address this challenge, we present an improved fault diagnosis method that combines a multiple parallel hidden layer denoising autoencoder (MPHLDAE) with a one‐dimensional convolutional neural network (1DCNN). In this method, the MPHLDAE is used to effectively denoise the raw vibration signals, whereas the 1DCNN is responsible for diagnosing faults in the denoised data. Hyperparameter optimization was carried out through a grid search with k ‐fold cross‐validation. Experimental evaluations were conducted using rolling bearing and gearbox vibration datasets intentionally mixed with varying levels of Gaussian noise. The results validate the superior denoising capability of the MPHLDAE compared with existing denoising techniques. The proposed method consistently achieves fault diagnosis accuracies exceeding 95.30%, even under low signal‐to‐noise ratio conditions. These findings prove the superiority of the proposed MPHLDAE‐1DCNN method over other advanced methods for rotating machines.