14 November 2023 in Structural Health Monitoring
A convolutional neural network fault diagnosis method based on frequency attention mechanism was designed for the problem that the traditional method cannot adaptively extract effective feature information in rolling bearing fault diagnosis and the diagnosis effect of rolling bearing is poor under strong environmental noise interference. Firs, the Mel-frequency cepstral coefficient (MFCC) of the bearing vibration signal was extracted. Second, to solve the problem of the channel attention mechanism adopting global average pooling (GAP) and neglecting channel internal characteristic information, the GAP was extended in the frequency domain, and a two-stage frequency component selection criterion was designed. The results show that the MFCC method can extract fault-sensitive features in industrial noise environments, improve the existing channel attention mechanism using frequency domain attention mechanism, and overcome the information loss caused by GAP of convolutional layer features in channel attention mechanism. Identification accuracy, recall rate, and F1-score are 100% on the rolling bearing simulation fault datasets of Case Western Reserve University and Central South University. Compared with the convolutional block attention module, the accuracy of the method combining spatial attention mechanism and channel attention mechanism is improved by 0.34 and 0.24%, respectively, and compared with other front-bearing fault diagnosis methods, it also offers significant improvement.
11 January 2024 in IEEE Transactions on Artificial Intelligence
Spiking neural networks (SNNs) are deeply inspired by biological neural information systems. Compared to convolutional neural networks (CNNs), SNNs are low power consumption because of their spike based information processing mechanism. However, most of the current structures of SNNs are fully-connected or converted from deep CNNs which poses redundancy connections. While the structure and topology in human brain systems are sparse and efficient. This paper aims at taking full advantage of sparse structure and low power consumption which lie in human brain and proposed efficient structure slimming methods. Inspired by the development of biological neural network structures, this paper designed types of structure slimming methods including neuron pruning and channel pruning. In addition to pruning, this paper also considers the growth and development of the nervous system. Through iterative application of the proposed neural pruning and rewiring algorithms, experimental evaluations on CIFAR-10, CIFAR-100, and DVS-Gesture datasets demonstrate the effectiveness of the structure slimming methods. When the parameter count is reduced to only about 10% of the original, the performance decreases by less than 1%.
25 September 2024 in IEEE Transactions on Emerging Topics in Computational Intelligence
Underwater object detection is of great significance for various applications in underwater scenes. However, class imbalance issue is still an unsolved bottleneck for current underwater object detection algorithms. It leads to large discrepancies in the detection precision among different classes that the dominant classes with more training data achieve higher precision while the minority classes with less training data achieve much lower precision. In this paper, we propose a balanced underwater object detection network (BAUODNET) to address the class imbalance issue by exploiting two techniques, i.e., the style augmentation technique and the example re-weighting technique. Firstly, we propose a class-wise style augmentation (CWSA) algorithm to augment the training data for the minority classes that generates different colors, textures and contrasts for the minority classes whilst preserving geometry. The augmented dataset possesses more balanced data distribution; Secondly, we exploit the the focal loss to re-weight the examples during the training of the deep detector, it down-weights the loss assigned to the well-detected examples from the dominant classes and focuses on learning undetected hard examples from the minority classes. Extensive experiments show the effectiveness of CWSA and focal loss for addressing the class imbalance problem in underwater scenes, BAUODNET obtains 49.5% mAP on URPC2017 and 66.8% mAP on URPC2018, achieving state-of-the-art or comparable performance on URPC2017 and URPC2018.