Mechanical fault diagnosis based on combination of sparsely connected neural networks and a modified version of social network search

Abstract

Swift and precise fault diagnosis is a significant category to guarantee machinery operates reliably and avoid major failures. Conventional methods for monitoring bearing health rely on large datasets of labeled faulty samples, which can be time-consuming and costly to gain. Intelligent fault identification techniques are limited by a lack of defective samples. Although Convolutional Neural Networks (CNNs) are effective tools for diagnosing mechanical faults, they may perform poorly on unseen data due to overfitting when trained with few faulty samples. The small sample problem poses a significant difficulty in mechanical fault diagnosis, as insufficient faulty samples can induce overfitting in models such as CNNs that leads to inadequate generalization. For instance, conventional CNNs may extremely adapt to the unique traits of limited training data, whereas SCNNs, characterized by their sparse connectivity, reduce this concern. Additionally, optimizing the hyperparameters of SCNNs can be complex. However, the M-SNS algorithm, using Lévy flight and self-adjusting population mechanisms proficiently solves this issue by enhancing exploitation and exploration. This study suggests a novel approach to solve the small sample problem Sparsely Connected Neural Networks (SCNNs) enhanced by optimizing its hyperparameters based on an improved version of Social Network Search (M-SNS). While standard SNS-based optimizers struggle with local optima, the improved version incorporates Lévy flight to meaningly improve global search performance, guaranteeing better generalization even in small sample scenarios. The proposed SCNN/M-SNS is employed to use a tool for the fault diagnosis. To guarantee the efficiency of the model, its results are applied to a benchmark, called Case Western Reserve University (CWRU) Bearing Dataset which includes a flexible test rig with a 2 hp motor to simulate various load conditions (0, 1, 2, and 3 hp) and controlled fault introduction using Electrical Discharge Machining (EDM) to precisely introduce faults in bearings with different diameters (7, 14, 28, and 40 mils) and at different locations (roller, inner, and outer race faults). Accelerometers are used to collect data at two sampling rates (12 kHz and 48 kHz). The results are compared with some similar state of the art models based on SCNN, including fuzzy neural network model (FNN), Convolution Neural Network (CNN), and LSTM neural network. The results show that the proposed model achieves a mean squared error (MSE) of 0.021 on the test dataset which is higher than the fully connected network with an MSE of 0.035. When compared to other metaheuristic algorithms, the M-SNS-based SCNN model achieves an MSE of 0.021, while Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) achieve MSE values of 0.023 and 0.025, respectively. The proposed model also outperforms other techniques on the CWRU dataset, with a classification accuracy of 98.2 %, compared to 95.5 % for CNN, 94.3 % for LSTM, and 90.1 % for FNN. These results demonstrate the effectiveness of the proposed model in fault diagnosis, particularly in scenarios with limited training data.