计算机科学
卷积神经网络
噪音(视频)
人工智能
残余物
模式识别(心理学)
断层(地质)
滤波器(信号处理)
特征提取
方位(导航)
特征(语言学)
光学(聚焦)
核(代数)
深度学习
一般化
灵活性(工程)
算法
特征学习
参数化复杂度
代表(政治)
故障检测与隔离
自适应滤波器
降噪
调制(音乐)
信号处理
瞬态(计算机编程)
人工神经网络
包络线(雷达)
信号(编程语言)
能量(信号处理)
网络体系结构
滚动轴承
机器学习
编码(集合论)
卷积(计算机科学)
频道(广播)
理论(学习稳定性)
作者
Jing Kang,Taiyong Wang,Ye Wei,Ying Zhong Tian
标识
DOI:10.1088/1361-6501/aebaa6
摘要
Rolling bearing fault diagnosis is essential for ensuring the safe and reliable operation of rotating machinery. In industrial scenarios, vibration signals are often contaminated by strong noise and affected by variable operating conditions, which obscure incipient fault signatures, aggravate signal non-stationarity, and hinder stable feature extraction and accurate health-state identification. To address the limited noise immunity, insufficient physical interpretability, and inadequate nonlinear representation capability of existing deep learning methods under low signal-to-noise ratio (SNR) conditions, this paper proposes a physics-informed adaptive multiscale convolutional network, termed PAMCN. The proposed method embeds parameterized time-frequency analysis kernels into the convolutional architecture to construct a trainable complex-valued band-pass finite impulse response filter bank, enabling the network to adaptively focus on fault-sensitive resonance bands while suppressing out-of-band noise. The complex modulus operation is further introduced to transform complex responses into energy-domain representations, thereby preserving fault-related energy fluctuations and reducing the influence of random phase perturbations. This design supports robust feature learning with physically interpretable time-frequency modeling. In addition, a residual adaptive multiscale fusion block, namely the RAM Block, is developed to capture transient impacts, envelope modulation patterns, and long-range temporal dependencies through parallel convolutions with diverse receptive fields. A channel attention mechanism is incorporated to adaptively reweight fault-sensitive features, enhancing discriminability and generalization under noisy and varying working conditions. Experiments on the CWRU, HUST, and Tianjin University bearing datasets show that PAMCN consistently outperforms representative comparison methods across different noise levels and changing operating conditions, particularly under low-SNR scenarios. The results demonstrate the effectiveness of PAMCN in terms of anti-noise robustness, physical interpretability, and cross-condition generalization, providing an engineering-oriented solution for reliable rolling bearing fault diagnosis in complex industrial environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI