Deep Fault Recognizer: An Integrated Model to Denoise and Extract Features for Fault Diagnosis in Rotating Machinery

Abstract

Fault diagnosis in rotating machinery is significant to avoid serious accidents; thus, an accurate and timely diagnosis method is necessary. With the breakthrough in deep learning algorithm, some intelligent methods, such as deep belief network (DBN) and deep convolution neural network (DCNN), have been developed with satisfactory performances to conduct machinery fault diagnosis. However, only a few of these methods consider properly dealing with noises that exist in practical situations and the denoising methods are in need of extensive professional experiences. Accordingly, rethinking the fault diagnosis method based on deep architectures is essential. Hence, this study proposes an automatic denoising and feature extraction method that inherently considers spatial and temporal correlations. In this study, an integrated deep fault recognizer model based on the stacked denoising autoencoder (SDAE) is applied to both denoise random noises in the raw signals and represent fault features in fault pattern diagnosis for both bearing rolling fault and gearbox fault, and trained in a greedy layer-wise fashion. Finally, the experimental validation demonstrates that the proposed method has better diagnosis accuracy than DBN, particularly in the existing situation of noises with superiority of approximately 7% in fault diagnosis accuracy.

Deep Learning Algorithms for Bearing Fault Diagnostics—A Comprehensive Review

“…The average classification result reveals the proposed SDA is able to achieve a worst case accuracy of 91.79%, which is 3% to 10% higher when compared to the conventional SAE without the denoising capability, and classical ML algorithms such as SVM and random forest (RF). Similar to [124], another form of SDA is utilized in [125] with three hidden layers of (500, 500, 500) units. Signals from the CWRU dataset are mixed with different levels of artificially induced noise in the time domain, and later transformed to the frequency domain.…”

“…It is also reported in [106] that the conventional CNN has a better built-in denoising mechanism compared to other classical DL algorithms such as AE. Due to this limitation, some papers applied the stacked denoising AE (SDAE) [124], [125], [132] to increase AE's noise resilience under a small SNR, i.e., SNR = 5 or 10. 2) Unbalanced Sampling: Regarding the selection of training samples from the CWRU dataset, many papers did not guarantee a balanced sampling, which means the ratio of data samples selected from the healthy condition and the faulty condition is not close to 1:1.…”

An Encoder-Decoder Based Convolution Neural Network (CNN) for Future Advanced Driver Assistance System (ADAS)

“…Current models offer methods to classify either a single object's label for a bounding box of a few objects or a whole input window in each scene. The CNNs have also been applied to a wide diversity of other tasks, e.g., stereo-depth [22], pose estimation [23], instance segmentation [24], and much more [25,26]. In such research methodologies, CNNs are either applied to discover local features or to produce descriptors of discrete proposal regions.…”

An Improved Second-Order Blind Identification (SOBI) Signal De-Noising Method for Dynamic Deflection Measurements of Bridges Using Ground-Based Synthetic Aperture Radar (GBSAR)

“…In particular, the used GBSAR equipment should be placed under the monitored bridge, and the instantaneous vibrations of the equipment itself will be inevitably caused by passing vehicles under the monitored bridge, which will reduce the accuracy of the obtained dynamic time-series displacement. Therefore, it is of great importance to reduce the influence of noise information in the time-series displacements of bridges obtained using GBSAR-especially for the instantaneous vibrations of the equipment itself.Currently, the primary signal de-noising methods for the time-series data include the filtering method [9][10][11][12], the wavelet transform method [13,14], the singular value decomposition method [15,16], and the empirical mode decomposition (EMD) method [17][18][19]. Filtering methods use statistical features to derive some estimation algorithms, and further estimate the useful signals or filter the signals with certain statistical features from the mixed signal, which can improve the signal-to-noise ratio (SNR).…”

“…Currently, the primary signal de-noising methods for the time-series data include the filtering method [9][10][11][12], the wavelet transform method [13,14], the singular value decomposition method [15,16], and the empirical mode decomposition (EMD) method [17][18][19]. Filtering methods use statistical features to derive some estimation algorithms, and further estimate the useful signals or filter the signals with certain statistical features from the mixed signal, which can improve the signal-to-noise ratio (SNR).…”