Abstract
Early wear of injector nozzles can cause unstable fuel injection, poor atomization, and increased exhaust emissions. This study proposes a vibration-signal-based diagnostic method for early nozzle wear detection. Miniature accelerometers are mounted on the injector body and fuel line to capture high-frequency vibration signals. A hybrid feature extraction scheme combining time-domain, frequency-domain, and time–frequency analysis—such as RMS, kurtosis, envelope spectrum, instantaneous energy, short-time Fourier transform (STFT), and wavelet packet decomposition—is used to characterize wear-induced changes. Selected features are then fed into machine learning classifiers (Support Vector Machine, Random Forest, or Convolutional Neural Network) to identify early wear states. Accelerated bench tests and vehicle validation show that the proposed method can detect micro-level nozzle wear before measurable flow deviation appears, achieving over 92% diagnostic accuracy and enabling real-time condition monitoring for predictive maintenance.
1. Introduction
In modern common-rail diesel systems, fuel injection precision and stability are crucial to engine performance and emission control. Even minor wear in the injector nozzle—such as orifice edge rounding or needle–seat clearance enlargement—can generate transient flow disturbances that induce distinctive vibration signatures. Compared to traditional flow or pressure-based diagnostics, vibration signal analysis offers higher sensitivity and suitability for online monitoring. Therefore, developing a vibration-based early wear detection approach provides valuable support for intelligent fault prediction and injector reliability enhancement.
2. Methodology
(1) Sensing and Data Acquisition
High-bandwidth MEMS or piezoelectric accelerometers (bandwidth ≥ 100 kHz) are mounted near the injector tip and fuel line. A sampling rate of ≥ 200 kHz is adopted to capture micro-impacts and high-frequency harmonics. The acquisition system is synchronized with the injector trigger pulse for time alignment.
(2) Signal Preprocessing
Raw signals are processed through DC removal, band-pass filtering (1–80 kHz), normalization, and independent component analysis (ICA) to eliminate background vibration and engine noise.
(3) Feature Extraction
-
Time-domain features: RMS, mean value, kurtosis, skewness, crest factor.
-
Frequency-domain features: power spectral density, main harmonic amplitude, spectral energy ratio.
-
Envelope analysis: Hilbert envelope spectrum to detect repetitive impact frequencies.
-
Time–frequency features: STFT and wavelet packet decomposition to describe transient energy distribution.
-
Nonlinear features: approximate entropy, multiscale entropy, and Hilbert–Huang transform (HHT) for instantaneous frequency and nonstationary dynamics.
(4) Feature Selection
Mutual information, principal component analysis (PCA), and recursive feature elimination (RFE) are used to select the most wear-sensitive features—typically high-frequency energy ratio, kurtosis, and envelope energy.
(5) Diagnostic Modeling
Supervised learning models such as SVM and Random Forest are trained for binary (healthy/worn) or multiclass (mild/moderate/severe wear) classification. For large datasets, 1D-CNN or LSTM architectures can perform end-to-end feature learning. Model generalization is evaluated through k-fold cross-validation.
(6) Threshold and Warning Strategy
Dynamic thresholds are established using statistical control methods (CUSUM, EWMA). Exceeding thresholds triggers early warning or intensified data sampling to support predictive maintenance.
3. Experimental Setup and Validation
A high-pressure common-rail test bench was constructed with interchangeable injector samples subjected to controlled wear (mechanical polishing or particle erosion). Tests were performed under varying injection pressures (100–200 MPa), fuel temperatures, and engine speeds. Microscopic and coordinate-measuring techniques recorded wear geometry, including orifice diameter changes and needle clearance.
Results indicate that when the nozzle orifice diameter change reaches only 5–10 μm, significant increases appear in high-frequency vibration energy and kurtosis, while envelope energy clearly rises. Machine learning models successfully identified these early changes within fewer than 10⁶ injection cycles. Field vehicle tests further confirmed the robustness of the proposed method under complex engine vibration backgrounds.
4. Evaluation Metrics
Diagnostic performance was evaluated using accuracy, recall, F1-score, and ROC-AUC. The method provided an average early-warning lead time of 2–12 hours before conventional flow or pressure measurements detected anomalies, achieving F1 = 0.91 and AUC > 0.95.
5. Conclusions
The vibration-signal-based diagnostic method enables high-sensitivity, non-intrusive detection of early injector nozzle wear. It can be integrated into an electronic control unit (ECU) or on-board diagnostic system using low-cost MEMS sensors and edge computing for real-time monitoring. Future work will focus on model robustness under multi-injection and multi-hole conditions, sensor fusion with pressure and flow data, and quantitative prediction of wear severity and remaining useful life (RUL) through physics-informed machine learning.
