A DPHD model for compressor fault diagnosis based on the fusion of multi-scale deep features and physical priors

ObjectiveExisting intelligent diagnosis methods predominantly rely on purely data-driven models, which are inherently “black-box” and often lack generalization due to the absence of physical information when handling the variable-length, non-stationary vibration signals of natural gas compressors. Therefore, developing a hybrid fault diagnosis model that offers high accuracy, robustness, and interpretability is essential for the safe and intelligent operation and maintenance of compressors. MethodsA Dual-Path Hybrid Diagnostic (DPHD) model for compressor faults was proposed. In the data-driven path, the Multi-Scale Temporal Residual Block (MTRB) served as the core, capturing dynamic signal characteristics across multiple time scales using parallel convolution kernels of varying sizes. A length-adaptive pooling layer then unified variable-length feature sequences into fixed-length vectors, effectively addressing variable-length signal processing. Concurrently, the physical prior path employed a Multilayer Perceptron (MLP) to perform deep nonlinear extraction of nine-dimensional physical statistical features, such as kurtosis and spectral entropy, thereby incorporating domain knowledge. Finally, the heterogeneous feature vectors from both paths were concatenated and fused before being input to a fully connected classifier for collaborative fault diagnosis. ResultsThe proposed DPHD model was validated on an industrial compressor dataset encompassing 10 operating conditions. Results demonstrated an overall diagnostic accuracy of 99.60％, significantly outperforming baseline models such as 1D Convolutional Neural Network (1D CNN) and Support Vector Machine (SVM). The t-Distributed Stochastic Neighbor Embedding (t-SNE) algorithm was employed to reduce dimensionality and visualize the high-dimensional features, confirming that the fused features learned by the DPHD model exhibited superior intra-class compactness and inter-class separability. Additionally, ablation studies confirmed the necessity and effectiveness of the dual-path architecture, multi-scale design, and integration of physical prior knowledge. ConclusionThe proposed DPHD model addresses challenges in variable-length signal processing and interpretability by integrating multi-scale deep features driven by data and physical prior knowledge. It delivers high-precision compressor fault diagnosis and offers a hybrid paradigm combining strong performance with physical insight for intelligent operation and maintenance of complex industrial equipment.