Research on Fault Diagnosis of Water-Based Dynamic Rodless Oil Pumping Machine Based on Manifold Learning and Genetic Algorithm

Technical Summary: Fault Diagnosis of Water-Based Dynamic Rodless Oil Pumping Machines via ISOMAP-GA

Problem Statement
Water-based dynamic rodless oil pumping machines offer significant advantages over traditional rod-type pumps, including improved efficiency, reduced costs, and the elimination of rod-tube wear. However, these systems operate under complex, time-varying nonlinear conditions where fault signals exhibit nonlinearity, non-stationarity, and high redundancy. Existing diagnostic methods face specific limitations: expert systems often lack precise fault type identification; kinematic equation-based methods struggle with parameter mapping under high noise; and fractal theory approaches lose differentiation when noise ratios exceed 0.4. Furthermore, traditional manifold learning algorithms, while effective at capturing intrinsic geometric structures without preset data distributions, suffer from insufficient discriminative power, sensitivity to noise and outliers, reliance on manual trial-and-error for core parameter selection, and high computational costs on large datasets.

Methodology
The paper proposes a fault diagnosis framework that integrates ISOMAP (Isometric Mapping) manifold learning with a Genetic Algorithm (GA) for global optimization. The methodology proceeds through the following stages:

1. Data Acquisition and Feature Extraction:

   • Data is collected from a fault simulation test bench equipped with sensors measuring oil inlet pressure, power fluid flow, produced fluid flow, outlet temperature, cylinder displacement, and vibration signals.
   • Raw vibration signals are denoised.
   • A 20-dimensional feature vector is constructed for each sample, comprising 13 time-domain features (e.g., RMS, peak-to-peak value) and 7 frequency-domain features (e.g., average frequency, spectral peak stability).
   • Four fault categories are defined: Normal operation, Oil outlet blockage, Power fluid leakage, and Fixed valve blockage.

2. Comparative Manifold Learning:

   • Initial experiments compared ISOMAP, Locally Linear Embedding (LLE), and Laplacian Eigenmaps (LE).
   • Results indicated that while LLE preserves local neighborhoods, it loses long-range correlations; LE excessively compresses feature space, reducing inter-class discrimination. ISOMAP was selected as the base method for its ability to preserve global topological structure and form distinct clusters for different fault modes.

3. ISOMAP-GA Collaborative Optimization:

   • To address ISOMAP's reliance on manual parameter tuning and sensitivity to noise, a Genetic Algorithm is employed to optimize the ISOMAP process.
   • Encoding: The GA uses real-number encoding for chromosomes representing the neighborhood size ($k$), dimensionality reduction dimension ($d$), geodesic distance weight ($w$), and a feature screening mask.
   • Fitness Function: A multi-objective function ($F$) is constructed to balance feature separability ($J$), computational efficiency ($T$), and physical meaning correlation ($P$), weighted as $F = 0.6J + 0.2T + 0.2P$.
   • Optimization Process: The GA performs selection (roulette wheel with elite retention), crossover (arithmetic, $p=0.8$), and mutation (Gaussian, $p=0.1$) to find the optimal parameters that maximize the distance between fault classes while minimizing intra-class variance.

4. Classification:

   • The optimized low-dimensional manifold features are input into a K-Nearest Neighbors (KNN) classifier for final fault mode recognition.

Key Contributions

• Algorithmic Integration: The paper introduces a synergistic model where a Genetic Algorithm automates the selection of ISOMAP core parameters ($k$ and $d$) and optimizes feature subsets, eliminating the need for manual trial-and-error.
• Enhanced Discriminative Power: By optimizing for maximum inter-class distance and minimum intra-class distance, the method significantly improves the separability of fault features in low-dimensional space compared to standard ISOMAP.
• Robustness: The integration of fault mechanism constraints into the fitness function and the selection of noise-resistant features enhance the method's stability against industrial noise and outliers.

Experimental Results

• Feature Separability: The optimized fault feature separability index ($J$) increased from 2.3141 (unoptimized) to 79.7555, representing an improvement of approximately 33.5 times.
• Clustering Performance: In the 2D manifold space, the four fault categories (Normal, Leakage, Outlet Blockage, Valve Blockage) formed compact, independent clusters with clear boundaries and no significant overlap.
• Diagnostic Accuracy: When combined with a KNN classifier, the proposed method achieved a fault diagnosis accuracy of 94.44%.
• Confusion Matrix Analysis: The confusion matrix indicated that all samples were correctly classified with no cross-class misclassifications, demonstrating the method's reliability in multi-fault mode recognition.

Significance and Claims
The paper claims that the proposed ISOMAP-GA method provides an effective solution for the intelligent fault diagnosis of water-based rodless oil pumping units. It addresses the specific challenges of high-dimensional, nonlinear, and noisy fault signals inherent in these systems. The authors assert that the method offers strong anti-interference capabilities and engineering applicability, as it requires no complex manual parameter tuning and maintains computational stability. By achieving high diagnostic accuracy and clear feature visualization, the study provides reliable technical support for the condition monitoring and maintenance of rodless oil pumping equipment in oilfield environments.