Research on the Application of VMD-CNN-LSTM Hybrid Deep Learning Framework in Mechanical Fault Diagnosis of Pump Station Units

This study proposes a hybrid deep learning framework combining Variational Mode Decomposition (VMD), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) network, specifically for mechanical fault diagnosis of pump station units. This framework effectively addresses issues such as noise interference and mode mixing in rotating machinery signal processing, achieving a diagnostic accuracy of 97.50% under complex operating conditions, and providing efficient and reliable technical support for intelligent operation and maintenance of pump station units.

Pump station units are core equipment in water conservancy projects and industrial water supply systems, and their operating status directly affects system safety and reliability. Under long-term high-load continuous operation, components such as rotors and bearings are prone to faults like wear and misalignment, which may lead to serious accidents if not handled in time. Traditional fault diagnosis relies on manual inspection and regular maintenance, which has limitations such as low efficiency, strong subjectivity, and difficulty in real-time monitoring. Automatic diagnosis based on vibration signal analysis has become a hot topic, but the vibration signals of pump stations are nonlinear, non-stationary, and noisy, posing challenges to signal processing and feature extraction.

VMD Signal Preprocessing

Variational Mode Decomposition (VMD) decomposes complex vibration signals into several Intrinsic Mode Functions (IMF) by constructing a variational optimization problem, avoiding mode mixing and improving decomposition stability. As a front-end module, it adaptively decomposes signals into multiple IMF components, reconstructs and denoises to improve signal-to-noise ratio, and provides structured input for subsequent models.

CNN Spatial Feature Extraction

Applying CNN to one-dimensional vibration signal feature extraction, it automatically learns local features and hierarchical representations through sliding operations of multi-layer convolution kernels, captures deep fault features that are difficult to find with traditional methods, and avoids the limitations of manually designed features.

LSTM Temporal Modeling

LSTM controls information flow through a gating mechanism, solving the gradient vanishing problem of traditional RNNs, memorizing signal evolution patterns within a long time window, which is crucial for identifying progressive and intermittent faults. It receives spatial features extracted by CNN and models temporal dependencies.

Adaptive Parameter Optimization Strategy

To address the problem of VMD parameters relying on expert experience, an adaptive parameter optimization method based on signal characteristics is introduced. It automatically determines the optimal decomposition parameters by analyzing spectral features and complexity indicators, improving engineering practicality.

Multi-layer Feature Fusion Mechanism

Organically integrating VMD time-frequency domain features, CNN spatial features, and LSTM temporal features, using multi-perspective complementary information, significantly improves diagnostic robustness.

End-to-End Joint Training

The framework adopts end-to-end joint training, where the parameters of the three modules are optimized collaboratively under a unified loss function, coordinating information transmission to achieve global optimization, which is superior to phased independent training.

Experimental Dataset

A dataset covering typical faults such as rotor imbalance, bearing wear, shaft misalignment, and dynamic-static rubbing is constructed, covering different load and speed conditions to simulate the actual operating environment.

Key Performance Indicators

1. Diagnostic accuracy reaches 97.50%, which is better than single CNN or LSTM models;
2. Signal-to-noise ratio is improved by 12.72 dB;
3. The loss function converges stably within 20 epochs;
4. The accuracy of cross-condition testing remains above 80.95%.

Comparative Experiments

• Compared with traditional time-frequency analysis methods (wavelet transform, Hilbert-Huang transform), it does not require manual selection of basis functions and has stronger adaptability;
• Compared with single deep learning models, the hybrid architecture captures both spatial features and temporal dependencies, making diagnosis more comprehensive;
• Compared with other preprocessing methods, VMD effectively suppresses mode mixing and has higher feature quality.

Integration into Intelligent Operation and Maintenance Systems

The results can be integrated into the intelligent operation and maintenance system of pump stations to realize online fault monitoring and early warning. Deployed on edge devices, it processes vibration signals locally in real time, identifies anomalies in time, and triggers maintenance processes.

Cross-domain Migration Potential

The framework has strong universality and can be migrated to fault diagnosis of rotating equipment such as wind turbines, gas turbines, and compressors, and can be quickly deployed with only minor adjustments.

Future Research Directions

• Introduce attention mechanisms to enhance the weight of key features;
• Explore lightweight networks to adapt to resource-constrained scenarios;
• Combine digital twins to build a closed-loop system for equipment health management;
• Study few-shot learning to reduce dependence on labeled data.

The VMD-CNN-LSTM hybrid framework proposed in this study provides an efficient and reliable solution for fault diagnosis of pump station units through the organic integration of signal decomposition, spatial feature extraction, and temporal modeling. Experimental verification shows its superior performance under complex operating conditions, providing technical support for intelligent operation and maintenance of water conservancy engineering equipment, and having demonstrative significance for promoting the application of deep learning in the field of industrial equipment condition monitoring.