Pre-Processing-Free Gear Fault Diagnosis Using Small Datasets with Deep Convolutional Neural Network-Based Transfer Learning (1710.08904v1)

Abstract: Early fault diagnosis in complex mechanical systems such as gearbox has always been a great challenge, even with the recent development in deep neural networks. The performance of a classic fault diagnosis system predominantly depends on the features extracted and the classifier subsequently applied. Although a large number of attempts have been made regarding feature extraction techniques, the methods require great human involvements are heavily depend on domain expertise and may thus be non-representative and biased from application to application. On the other hand, while the deep neural networks based approaches feature adaptive feature extractions and inherent classifications, they usually require a substantial set of training data and thus hinder their usage for engineering applications with limited training data such as gearbox fault diagnosis. This paper develops a deep convolutional neural network-based transfer learning approach that not only entertains pre-processing free adaptive feature extractions, but also requires only a small set of training data. The proposed approach performs gear fault diagnosis using pre-processing free raw accelerometer data and experiments with various sizes of training data were conducted. The superiority of the proposed approach is revealed by comparing the performance with other methods such as locally trained convolution neural network and angle-frequency analysis based support vector machine. The achieved accuracy indicates that the proposed approach is not only viable and robust, but also has the potential to be readily applicable to other fault diagnosis practices.

Pre-Processing-Free Gear Fault Diagnosis Using Small Datasets with DCNN-Based Transfer Learning

The paper presents a deep convolutional neural network (DCNN)-based transfer learning approach aimed at the pre-processing-free diagnosis of gear faults using limited size datasets. This method addresses some enduring challenges in the field of early fault diagnosis in mechanical systems, particularly gearboxes, where traditional methods have been significantly reliant on domain-specific feature extraction and have required substantial volumes of training data.

Methodology

The paper introduces a novel transfer learning architecture consisting of two core components: a pre-trained DCNN for feature extraction and a fully connected stage for classification. The former leverages a model initially trained on a large dataset (1.2 million images from ImageNet), which is repurposed for the gear fault diagnosis task. The second component is specifically trained using the experimental data from gear fault conditions, which requires significantly fewer samples compared to traditional deep learning methodologies.

Experiments and Results

The experiments utilized raw accelerometer data from a gearbox system, which was analyzed without pre-processing—a common requirement in many alternatives. The experimental setup involved testing nine different gear conditions, including various severities of faults such as root cracking and tooth chipping, to rigorously evaluate the approach.

The proposed method demonstrated a superior classification accuracy of up to 99.47%, significantly outperforming other traditional techniques such as locally trained CNNs and angle-frequency analysis followed by Support Vector Machine (SVM) classification. This performance held true even when training datasets were highly reduced, with the method outperforming alternatives with only 10% of the available data used for training.

Implications

The implications of this paper are manifold both from a theoretical and practical standpoint. By effectively eliminating the need for pre-processing and utilizing transfer learning, the approach simplifies the diagnostic process considerably, reducing biases associated with manual feature extraction. Practically, the reduced data requirement and pre-processing-free nature bode well for industries where obtaining large datasets is a challenge.

Furthermore, the employment of high-level feature extraction from pre-trained models (originally designed for image classification) in a mechanical fault diagnosis context opens new avenues for cross-domain applications of deep learning models. This could potentially expedite the adoption of AI in traditional engineering domains, facilitating more robust predictive maintenance frameworks.

Future Directions

There are several intriguing directions for future research stemming from this work. Extending the applicability of such transfer learning approaches to other fault-diagnosis realms within mechanical engineering can significantly broaden the impact. Additionally, refining the approach to handle real-time diagnostics and integrating it with IoT devices could harness continuous monitoring and live fault detection. Exploring more sophisticated architectures could also push the boundaries in achieving even greater generalization across multiple mechanical systems.

In summary, this paper introduces a DCNN-based transfer learning methodology that pioneers the efficient diagnosis of gear faults using small datasets without pre-processing. Its demonstrated superiority over traditional methods underscores the growing potential of AI-driven approaches in industrial applications and sets the stage for broader adoption of transfer learning practices.