In modern automotive systems, the electronically controlled gear pump plays a critical role in ensuring the reliable operation of transmission and lubrication systems. As a key component, it supplies lubricating oil to reduce friction and wear, while maintaining optimal operating temperatures. The integration of a motor control unit allows for precise regulation of pump speed and performance, adapting to varying vehicle demands. However, under complex on-board conditions, such as high temperatures and dynamic loads, gear pumps are prone to failures like tooth damage or missing teeth, which can lead to catastrophic system failures. Therefore, real-time monitoring and accurate fault diagnosis of these pumps are essential for vehicle safety and efficiency. In this study, I explore advanced vibration signal analysis techniques to diagnose faults in automotive electronically controlled gear pumps, with a focus on improving decomposition methods and feature extraction under diverse operating scenarios.
The importance of fault diagnosis stems from the need to prevent unexpected downtimes and ensure operational integrity. Traditional methods, such as current signal analysis or acoustic monitoring, often suffer from limitations like susceptibility to noise or load variations. For instance, current signals can be influenced by changes in the motor control unit load, while audio signals may be contaminated by ambient noise, making it difficult to isolate fault-specific features. Vibration signal analysis, in contrast, offers a direct insight into mechanical conditions, but extracting meaningful features from these signals—especially in low-speed or high-noise environments—remains challenging. My research addresses these gaps by developing a robust framework that combines mathematical modeling, experimental validation, and enhanced signal processing techniques. This approach not only improves fault detection accuracy but also adapts to the real-time constraints of automotive applications, where the motor control unit must respond swiftly to diagnostic outputs.
To lay the groundwork, I first establish a dynamic mathematical model for the gear pump. This model helps elucidate the vibration sources and characteristics under both normal and fault conditions. The gear pump, typically an involute internal gear type, involves key parameters such as tooth count, eccentricity, and rotor thickness. By deriving equations for displacement and flow rate, I can relate geometric changes to vibrational behavior. For example, the displacement \(V\) is given by:
$$V = \frac{ABZ_1}{2\pi} \approx \frac{(R_{e1}^2 – R_{i1}^2)B}{2}$$
where \(A\) is the area, \(B\) is the rotor thickness, \(Z_1\) is the number of teeth on the inner rotor, and \(R_{e1}\) and \(R_{i1}\) are the outer and inner radii of the inner rotor, respectively. The flow rate \(q\) depends on the rotational speed \(n_1\) and volumetric efficiency \(\eta_v\):
$$q \approx \frac{\pi n_1 (R_{e1}^2 – R_{i1}^2) B \eta_v}{60}$$
These equations show that faults altering the gear geometry—such as tooth damage—can affect pump performance and induce unbalanced centrifugal forces, leading to distinctive vibration patterns. The motor control unit regulates \(n_1\), making it a pivotal element in monitoring these changes. By integrating this model, I can predict how faults manifest in vibration signals, providing a theoretical basis for subsequent analysis.
Next, I design and construct an experimental platform to simulate real-world operating conditions and fault modes. This platform includes an electronically controlled pump system managed by a motor control unit, which allows precise control over speed and torque via CAN communication. The oil temperature control system maintains temperatures from -40°C to 130°C, simulating extreme automotive environments. A PLC-based electrical control system handles data acquisition from flow and pressure sensors, ensuring comprehensive monitoring. The pump is installed in a test chamber with controlled oil lines, and vibration signals are captured using accelerometers. This setup enables the simulation of four fault states: normal, damaged tooth, missing tooth, and combined damaged-missing tooth, across various temperatures (30°C, 50°C, 80°C, 100°C, 120°C) and speeds (800, 1000, 1500, 2000, 3000 rpm), resulting in 100 distinct operating modes. The integration of the motor control unit ensures that speed variations are accurately replicated, mimicking on-board scenarios where the pump operates under dynamic loads.
With the experimental data collected, I proceed to analyze the vibration signals in both time and frequency domains. Time-domain analysis involves calculating standard indicators to quantify signal characteristics. For a vibration signal \(x(n)\) with \(N\) samples, key parameters include variance \(\delta\), standard deviation \(\sigma_x\), kurtosis \(K\), and skewness \(S\), defined as:
$$\delta = \frac{1}{N} \sum_{n=1}^{N} X^2_n$$
$$\sigma_x = \sqrt{\frac{1}{N} \sum_{n=1}^{N} [X(n) – \bar{x}]^2}$$
$$K = \frac{\sum_{n=1}^{N} [x(n) – \bar{x}]^4}{(N-1) \sigma_x^4}$$
$$S = \frac{\sum_{n=1}^{N} [x(n) – \bar{x}]^3}{(N-1) \sigma_x^3}$$
Additionally, root mean square (RMS) \(X_{rms}\), peak factor \(C\), and impulse factor \(I\) are computed to assess signal intensity and impulsivity. I summarize these parameters for different fault states under specific conditions, such as 120°C oil temperature and 800 rpm speed. For instance, the kurtosis values show significant variations, with normal state at 26.39, damaged tooth at 18.45, missing tooth at 23.33, and combined fault at 24.26, indicating that kurtosis is a sensitive indicator for fault discrimination. This sensitivity is crucial for the motor control unit to trigger alerts based on threshold breaches. Similarly, in the frequency domain, I apply Fast Fourier Transform (FFT) to obtain spectral features. Metrics like spectral centroid \(F_1\), spectral kurtosis \(F_6\), and mean spectral kurtosis \(F_7\) are derived:
$$F_1 = \frac{\sum_{k=1}^{K} f_k S(k)}{\sum_{k=1}^{K} S(k)}$$
$$F_6 = \frac{\sum_{k=1}^{K} (f_k – F_2)^4 S(k)}{\sum_{k=1}^{K} S(k)}$$
$$F_7 = \frac{1}{K} \sum_{k=1}^{K} F_6$$
where \(S(k)\) is the spectral amplitude at frequency \(f_k\). The mean spectral kurtosis, in particular, exhibits marked differences between fault types, e.g., increasing by over 30 times for combined faults at high speeds, highlighting its utility in feature extraction. To visualize time-frequency complexities, I employ Continuous Wavelet Transform (CWT), which reveals energy distributions across scales. For example, normal signals show uniform energy, while combined faults display concentrated low-frequency impulses. These analyses collectively identify optimal parameters—kurtosis in time domain and mean spectral kurtosis in frequency domain—that best reflect fault characteristics, enabling efficient data reduction for the motor control unit diagnostics.
To further enhance feature extraction, I compare several signal decomposition methods: Empirical Mode Decomposition (EMD), Ensemble EMD (EEMD), Complete EEMD with Adaptive Noise (CEEMDAN), and Multivariate EEMD (MEEMD). These techniques decompose signals into Intrinsic Mode Functions (IMFs), facilitating the isolation of fault-related components. The decomposition process for a signal \(x(t)\) can be generalized as:
$$x(t) = \sum_{i=1}^{M} IMF_i(t) + r(t)$$
where \(IMF_i\) are the IMFs and \(r(t)\) is the residual. I evaluate these methods based on decomposition time, clustering performance using SSE and silhouette coefficient, and accuracy when integrated with a Backpropagation (BP) neural network. The BP network, with its forward and backward propagation, adjusts weights to minimize error, expressed as:
$$\Delta w = -\eta \frac{\partial E}{\partial w}$$
where \(\eta\) is the learning rate and \(E\) is the error function. Results show that CEEMDAN-BP achieves the highest accuracy (e.g., 98.58% at 800 rpm) but with long decomposition times, while MEEMD-BP offers a balance with reduced time. However, to address efficiency concerns, I propose an improved method called Modified CEEMDAN (MCEEMDAN). This approach first applies EMD to the raw signal, filters IMFs using permutation entropy to remove noisy components, reconstructs the signal, and then performs CEEMDAN. Permutation entropy \(H_p\) for an IMF is calculated as:
$$H_p = -\sum_{j=1}^{J} p_j \log_2 p_j$$
where \(p_j\) is the probability of ordinal patterns. By thresholding \(H_p\), I eliminate complex, random IMFs, simplifying the signal before CEEMDAN. This reduces decomposition time while preserving accuracy. For instance, at 3000 rpm, MCEEMDAN improves accuracy by 1.29% over CEEMDAN and cuts decomposition time by 28.45% on average. The motor control unit benefits from this efficiency, as faster diagnostics allow quicker adjustments to pump operations.
The performance of MCEEMDAN is validated across various operating modes. I present detailed tables summarizing key metrics. For example, the decomposition times for different methods under varied conditions are compared:
| Condition Type | EEMD (s) | CEEMDAN (s) | MCEEMDAN (s) | Time Reduction (%) |
|---|---|---|---|---|
| Normal Mode | 364.03 | 522.24 | 340.70 | 6.41 / 34.76 |
| Damaged Tooth | 359.81 | 499.68 | 348.79 | 3.06 / 30.20 |
| Missing Tooth | 356.56 | 441.75 | 326.76 | 8.36 / 26.03 |
| Combined Fault | 357.49 | 396.81 | 312.34 | 12.63 / 21.29 |
| Oil Temperature 30°C | 350.18 | 452.59 | 304.28 | 13.11 / 32.77 |
| Oil Temperature 120°C | 356.55 | 466.39 | 345.86 | 3.00 / 25.84 |
| Speed 800 rpm | 362.18 | 497.54 | 357.75 | 1.22 / 28.10 |
| Speed 3000 rpm | 354.34 | 423.32 | 306.67 | 13.45 / 27.56 |
This table illustrates that MCEEMDAN consistently reduces time while maintaining diagnostic precision. Additionally, the accuracy of fault identification using BP neural networks is summarized for different methods:
| Decomposition Method | Training Accuracy (%) | Testing Accuracy (%) | Total Time (s) | Iterations |
|---|---|---|---|---|
| EEMD-BP | 83.77 | 83.58 | 450.24 | 469 |
| CEEMDAN-BP | 98.63 | 98.58 | 672.99 | 981 |
| MEEMD-BP | 85.22 | 85.28 | 113.15 | 574 |
| MCEEMDAN-BP | 87.40 | 87.31 | 426.56 | 834 |
At high speeds, MCEEMDAN-BP achieves 93.66% testing accuracy, outperforming others. The improved efficiency stems from the permutation entropy step, which effectively filters out irrelevant IMFs, reducing computational load. This is particularly advantageous for the motor control unit, which must process data in real-time to make swift decisions. Moreover, the robustness of MCEEMDAN is tested under low-speed conditions, where fault signals are weaker. Although accuracy slightly decreases compared to CEEMDAN, the time savings (e.g., 29.6% at 800 rpm) make it viable for practical applications. The motor control unit can leverage this method to monitor gear pumps across diverse scenarios, from idle to high-speed operations, ensuring continuous protection.
Beyond decomposition methods, I delve into the theoretical implications of the vibration model. The gear pump’s dynamics can be expressed through differential equations accounting for fluid-structure interactions. For instance, the vibration acceleration \(a(t)\) due to tooth meshing can be modeled as:
$$a(t) = \sum_{m=1}^{M} A_m \cos(2\pi f_m t + \phi_m) + n(t)$$
where \(A_m\), \(f_m\), and \(\phi_m\) are the amplitude, frequency, and phase of the m-th meshing harmonic, and \(n(t)\) is noise. Faults modulate these parameters, e.g., tooth damage introduces sidebands around \(f_m\). By analyzing these modulations, the motor control unit can infer fault severity. I also explore the impact of oil temperature on signal properties. As temperature rises, oil viscosity decreases, altering damping characteristics and vibration signatures. This relationship can be quantified using empirical formulas, such as:
$$\eta_v = \eta_0 e^{-\beta (T – T_0)}$$
where \(\eta_v\) is the viscosity, \(T\) is temperature, and \(\beta\) is a constant. Integrating such models into the diagnostic framework enhances adaptability, allowing the motor control unit to compensate for environmental variations. Furthermore, I investigate the role of sampling frequency and sensor placement in signal quality. High-frequency sampling (e.g., 20 kHz) captures transient features, while optimal sensor locations—near the gear mesh—maximize signal-to-noise ratio. These factors are critical for reliable data acquisition, which the motor control unit depends on for accurate fault detection.
To address potential challenges in real-world deployment, I consider issues like sensor drift, electromagnetic interference, and computational constraints. For example, vibration sensors may degrade over time, requiring calibration routines integrated with the motor control unit. Similarly, interference from other vehicle systems can be mitigated using shielded cabling and digital filtering. Computational limits are tackled by optimizing algorithm parameters, such as reducing the number of IMFs in MCEEMDAN or employing lightweight neural networks. I also propose a hybrid approach combining vibration analysis with other signals, such as pressure or flow data from the pump, to cross-validate diagnoses. This multi-sensor fusion, managed by the motor control unit, increases robustness against false alarms. For instance, a drop in flow rate coupled with specific vibration patterns can confirm a tooth fault, enhancing diagnostic confidence.
In terms of implementation, the proposed methods can be embedded into automotive electronic control units (ECUs). The motor control unit would run the MCEEMDAN algorithm in real-time, using pre-trained BP models for classification. This requires efficient code optimization and hardware acceleration, such as using DSP chips or FPGA-based solutions. I outline a potential workflow: the motor control unit acquires vibration data, preprocesses it (e.g., filtering), applies MCEEMDAN for decomposition, extracts features like kurtosis and spectral kurtosis, and feeds them into the BP network for fault prediction. Results are then logged or transmitted for maintenance alerts. This automated process reduces human intervention and enables proactive maintenance, aligning with Industry 4.0 trends. Additionally, I discuss scalability to other pump types or rotating machinery, emphasizing the versatility of the approach.
The economic and safety benefits of this research are substantial. By preventing pump failures, vehicle downtime is minimized, reducing repair costs and improving reliability. For automotive manufacturers, integrating such diagnostics enhances product value and customer satisfaction. The motor control unit becomes not just a control device but a health monitoring system, contributing to overall vehicle intelligence. From a safety perspective, early detection of gear pump faults avoids lubrication loss, which could lead to transmission seizure and accidents. Thus, this work supports advancements in autonomous vehicles, where system health monitoring is paramount. I also touch on environmental aspects, as efficient pumps reduce energy consumption and emissions, further underscoring the importance of reliable operation.
Looking ahead, future research directions include exploring deep learning techniques, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), for direct feature learning from raw signals. This could bypass manual feature extraction, potentially improving accuracy. Additionally, integrating physics-informed neural networks could combine model-based and data-driven approaches, leveraging the gear pump equations I derived earlier. Another avenue is cloud-based diagnostics, where the motor control unit uploads data for centralized analysis, enabling fleet-wide monitoring and predictive analytics. However, challenges like data security and latency must be addressed. Furthermore, extending the method to multi-fault scenarios or progressive fault degradation could provide deeper insights into pump lifecycle management. Collaborations with automotive OEMs would facilitate real-world testing and validation under diverse driving conditions.
In conclusion, my study presents a comprehensive fault diagnosis framework for automotive electronically controlled gear pumps, emphasizing vibration signal analysis under complex on-board conditions. Through mathematical modeling, experimental validation, and advanced signal processing—particularly the improved MCEEMDAN method—I demonstrate significant enhancements in decomposition efficiency and fault identification accuracy. The integration with BP neural networks and the focus on key parameters like kurtosis and spectral kurtosis ensure robust performance across temperatures and speeds. The motor control unit plays a central role in this ecosystem, enabling real-time monitoring and adaptive control. By addressing limitations of existing methods and proposing practical solutions, this research contributes to safer, more efficient automotive systems. As vehicles evolve towards electrification and autonomy, such diagnostic capabilities will become increasingly vital, paving the way for smarter maintenance strategies and enhanced operational reliability.