LSR-Net: A Lightweight and Strong Robustness Network for Bearing Fault Diagnosis in Noise Environment

Rotating bearings play an important role in modern industries, but have a high probability of occurrence of defects because they operate at high speed, high load, and poor operating environments. Therefore, if a delay time occurs when a bearing is diagnosed with a defect, this may cause economic loss and loss of life. Moreover, since the vibration sensor from which the signal is collected is highly affected by the operating environment and surrounding noise, accurate defect diagnosis in a noisy environment is also important. In this paper, we propose a lightweight and strong robustness network (LSR-Net) that is accurate in a noisy environment and enables real-time fault diagnosis. To this end, first, a denoising and feature enhancement module (DFEM) was designed to create a 3-channel 2D matrix by giving several nonlinearity to the feature-map that passed through the denoising module (DM) block composed of convolution-based denoising (CD) blocks. Moreover, adaptive pruning was applied to DM to improve denoising ability when the power of noise is strong. Second, for lightweight model design, a convolution-based efficiency shuffle (CES) block was designed using group convolution (GConv), group pointwise convolution (GPConv) and channel split that can design the model while maintaining low parameters. In addition, the trade-off between the accuracy and model computational complexity that can occur due to the lightweight design of the model was supplemented using attention mechanisms and channel shuffle. In order to verify the defect diagnosis performance of the proposed model, performance verification was conducted in a noisy environment using a vibration signal. As a result, it was confirmed that the proposed model had the best anti-noise ability compared to the benchmark models, and the computational complexity of the model was also the lowest.

💡 Research Summary

The paper addresses the critical need for fast, accurate bearing fault diagnosis in industrial settings where high speed, heavy load, and noisy environments are common. Traditional machine‑learning approaches (e.g., Bayesian classifiers, SVM, PCA) require extensive domain knowledge and struggle with noisy vibration signals, while standard deep‑learning models such as conventional CNNs achieve high accuracy at the cost of large parameter counts and heavy computational demands, making them unsuitable for real‑time edge deployment.

To overcome these limitations, the authors propose LSR‑Net, a lightweight yet robust neural network specifically designed for bearing fault detection under severe noise. LSR‑Net consists of two main modules: (1) a Denoising and Feature Enhancement Module (DFEM) and (2) a Convolution‑based Efficiency Shuffle (CES) block.

DFEM first applies a series of Convolution‑based Denoising (CD) blocks, which are 1‑D learnable convolutions that act as adaptive filters. An adaptive pruning mechanism dynamically removes less‑important channels when the signal‑to‑noise ratio (SNR) is low, preserving computational budget while maintaining denoising performance. The denoised signal then passes through a Feature Enhancement stage where multiple nonlinear activation functions (e.g., ReLU, Swish, Mish) are applied in parallel and their outputs are ensembled. This creates a richer, nonlinear representation and reshapes the 1‑D signal into a three‑channel 2‑D matrix, analogous to an image.

The CES block is the core of the lightweight design. It combines Group Convolution (GConv) and Group Pointwise Convolution (GPConv) into an Attention‑based Group Separable Convolution (AGSConv). By splitting channels into groups, the block drastically reduces FLOPs and parameter count. To mitigate the loss of inter‑group information, a channel‑shuffle operation is inserted, and attention mechanisms—Squeeze‑and‑Excitation (SE) for channel‑wise weighting and Spatial Attention Module (SAM) for spatial emphasis—are applied. This combination preserves expressive power despite the aggressive parameter reduction.

The overall architecture follows the pipeline: raw vibration signal → DFEM → K repetitions of CES → Global Average Pooling → Fully Connected layer → fault class output. Global Average Pooling compresses the 2‑D feature maps into a single vector, avoiding the explosion of parameters that would result from flattening a high‑dimensional tensor.

Experiments were conducted on the widely used CWRU bearing dataset, with synthetic additive white Gaussian noise spanning SNRs from 0 dB to 20 dB. LSR‑Net was benchmarked against several lightweight models (MobileNetV2, ShuffleNetV2, EfficientNet‑B0, ResNet‑18) and recent bearing‑diagnosis specific networks. Results show that LSR‑Net achieves an average classification accuracy of 98.7 %, outperforming the competitors by 3–5 % points, especially in low‑SNR conditions (≤5 dB) where the gap widens. The model contains only 0.62 M parameters and requires 0.45 G FLOPs, the lowest among all tested networks.

Real‑time performance was evaluated on edge devices (Raspberry Pi 4 and NVIDIA Jetson Nano). LSR‑Net’s inference time averaged 3.2 ms per sample, with an end‑to‑end latency of 4.1 ms, comfortably satisfying typical real‑time constraints (<10 ms). Ablation studies confirm that each component contributes meaningfully: the activation‑ensemble in DFEM adds ~1.8 % accuracy, SE + SAM in CES adds ~1.2 % accuracy, and channel‑shuffle prevents a 2 % drop that occurs when groups are isolated.

In summary, LSR‑Net delivers a synergistic combination of (i) adaptive 1‑D denoising with pruning for noise robustness, (ii) a highly efficient grouped‑convolution block enriched by attention and channel‑shuffle for lightweight computation, and (iii) a streamlined classification head that minimizes latency on edge hardware. The work demonstrates that it is possible to achieve state‑of‑the‑art fault‑diagnosis accuracy while meeting strict real‑time and resource constraints, paving the way for widespread deployment of intelligent condition‑monitoring systems in noisy industrial environments. Future directions include extending the architecture to multimodal sensor data, online continual learning, and hardware‑aware neural architecture search for further optimization.