1. Introduction
Gearboxes, as main power conversion components of automotives [
1], wind turbines [
2], marine devices [
3], and mining equipment [
4], are prone to wear failure due to their complex structures and poor working conditions. Gearbox failures are multifaceted and can stem from diverse causes [
5,
6,
7]. These failures might emerge due to factors, e.g., inadequate lubrication, which can escalate friction between gear teeth, leading to wear and eventual damage. Wear and tear on gear teeth or bearings over time due to cyclic loading, misalignment, or overloading can also contribute significantly to gearbox failures. Moreover, issues like fatigue failure arising from repeated stress on gear components or the presence of contaminants within the gearbox can compromise its efficiency and functionality. Each factor, alone or in combination, can precipitate gearbox failures, emphasizing the importance of regular maintenance, proper lubrication, alignment checks, and operational care to mitigate such risks.
Fatigue wear phenomenon in gearboxes generally occurs in combined rolling and sliding friction under elastohydrodynamic lubrication (EHL) or mix EHL/boundary lubrication conditions [
8,
9]. Some operation conditions of gearboxes, such as loads, speeds, film thickness, and flash temperatures, also have influences on their wear phenomenon [
10,
11]. More than 80% of scrap parts in gearboxes are caused by severe wear. Therefore, prediction and monitoring of worn parts of gearboxes in real time are important to evaluate the reliability and lifespans of gearboxes. To minimize wear failures in gearboxes, it is important to follow proper maintenance practices, including regular inspection, lubrication, and alignment checks [
12]. Additionally, using high-quality lubricants, ensuring proper load distribution, and addressing any operational issues promptly can help extend lifespans of gearboxes. The wear phenomenon of gearboxes has been studied from several perspectives [
13], including vibration-based condition monitoring [
14,
15,
16], tribo-dynamic behaviors [
17,
18], metallurgy [
19], surface finish [
20,
21], sliding friction [
22], assembly [
23], transmission error [
24,
25], and lubrication [
26,
27,
28,
29]. Mechanisms of the wear phenomenon are not completely understood, but it appears clear that it is affected by operation conditions, surface roughness, and lubricants.
Oil-related condition monitoring technologies are common solutions to identify gearbox faults, evaluate operation conditions, and define oil maintenance intervals, including oil debris analysis and oil condition monitoring (OCM) [
30]. Oil debris analysis is mainly performed by using oil sampling or sensors that measure some parameters, e.g., wear particle counting and oil contaminant elements [
31]. Other studies of OCM focused on sensor validation of oil analysis methods to identify gearbox faults based on valid data on additive depletion and oil degradation due to oxidation, oil contamination, temperature, and viscosity changes [
32]. In addition, several OCM methods of gearboxes based on multi-variable oil analysis methods, e.g., fluorescence spectroscopy, and oil viscosity measurements, were developed to improve analysis accuracy and measurement sensitivity of oil-based condition monitoring systems of gearboxes [
32,
33].
An oil debris analysis method of wear evaluation was developed to monitor wear conditions of engines by analyzing wear particles in engine lubrication oil [
34]. This oil debris analysis method exhibited strong sensitivity to identification of engine wear failure. Spectral analysis [
35] is one of the main solutions for oil debris analysis for mechanical equipment. This technique is used to monitor variations in wear element concentrations to predict wear conditions of mechanical equipment. For many years, numerous failure analysis methods have been proposed to evaluate equipment wear failure based on spectral analysis. Huang et al. [
36] evaluated dynamic performances of a hypoid gear pair in a tidal energy converter using a coupled torsional vibration model across diverse tidal speed profiles. developed a two-stage residual grey model to predict spectral data of wear particles in lubrication oil of marine engines, which improved the accuracy of failure prediction with large fluctuations of wear conditions. A remaining useful life prediction model of power-shift steering transmission systems was developed to analyze wear failure by monitoring the oil spectral data [
37]. Zhang et al. [
38] proposed an unequal interval GM(1, 1) model and used a lead (Pb) element as the wear characteristic index to predict wear failure of marine diesel engines. Valis et al. [
39] used Wiener failure model for iron (Fe) and Pb elements in lubrication oil to enhance wear failure identification of vehicle engines. To improve the prediction accuracy of wear failure analysis based on spectral analysis, some works studied oil debris analysis with multi-variable wear elements. Lara et al. [
40] combined principal component and sequential discriminant analyses to investigate multi-variable wear elements. A subspace reconstruction-based robust kernel principal component analysis model was developed [
41] for fault detection of wind turbines using supervisory control and data acquisition (SCADA) data to extract nonlinear features under discontinuous interference. However, only the overall wear conditions of gearboxes can be obtained based on the above oil debris analysis methods with the Fe element as the main component, and the specific wear parts of gearboxes remained obscured.
Multi-variable gray prediction models can identify correlations among
N variables [
42]. The GM(0,
N) and GM(1,
N) models are multi-variable gray prediction models for the zero- and first-order equations, respectively. The GM(1,
N) model can identify correlations between
independent variables and a dependent variable by using the first-order non-derivative series [
43]. The principle of the GM(1,
N) model is to accumulate and generate the first-order accumulated generating operation (1-AGO) sequence with increasing regularity based on the original sequence with unstable and irregular original sequence data [
44]. Consequently, the GM(1,
N) model can effectively describe regular changes in correlations among multiple variables. To improve the prediction accuracy of the GM(1,
N) model, Tien [
45] proposed an optimized GM(1,
N) model with an improved analysis structure. A multi-variable GM(1,
N) model was proposed in [
46], incorporating additional terms, e.g., dependent variable lag term and linear correction term. Li et al. [
47] developed a time-lagged logistic grey prediction model with harvesting terms.
Wear debris and elements in lubrication oil of gearboxes come from wear behaviors of meshing gear tooth surfaces, friction couplings, and bearings, and components of each wear part are complex and different [
48]. Therefore, this study aims to predict wear parts of a gearbox by studying multi-variable relationships among different wear elements in lubrication oil. An oil debris analysis method is developed based on a multi-objective genetic algorithm-GM(1,
N) (MOGA-GM(1,
N)) model, which combines the reliability of describing multi-variable parameter relations of the GM(1,
N) model [
42], and a multi-objective optimization function of the MOGA, to predict wear locations with the Fe element. The main source of the Fe element was examined based on the weight coefficient of each related element to the Fe element. Wear experimental results of a gearbox verifies feasibility and accuracy of the proposed oil debris analysis method.
The remaining part of the paper is organized as follows: the oil debris analysis method of the gearbox is introduced in
Section 2. The MOGA-GM(1,
N) model is developed to evaluate correlations between the Fe element and other elements of parts of the gearbox in
Section 3. Wear experiments and discussions of the proposed oil debris analysis method with the MOGA-GM(1,
N) model of a gearbox are presented in
Section 4. Finally, some conclusions from this study are presented in
Section 5.
2. Oil Debris Analysis Method
Main wear parts in a gearbox consist of gears, bearings, shafts, friction plates, seals, gear pumps, and other friction parts. Identifying varied element compositions of different parts in the gearbox is crucial for analyzing its wear conditions. Material and element compositions of parts in the gearbox are listed in
Table 1. According to
Table 1, most parts in the gearbox contain the Fe element, i.e., planetary gear sets, planet carriers, external teeth friction plates, bearings, ring gears, and the housing. Few parts in the gearbox do not contain the Fe element, i.e., internal toothed friction plates, bearing needle roller diaphragms, and gasket rings.
Since percentages of elements of material composition of parts in the gearbox are constant, the wear condition of the Fe element is always accompanied by other related elements of oil debris or particles, such as copper (Cu), nickel (Ni), chromium (Cr), molybdenum (Mo), aluminum (Al), and manganese (Mn), in the corresponding proportion, excluding carbon (C), and silicon (Si) elements that can be greatly affected by external pollution. Therefore, changes in the concentration of the Fe element in the lubrication oil can be analyzed by concentrations of other related elements. Since contents of different elements of the same part in the gearbox are different, it is difficult to obtain the correlation between the concentration of the Fe element and concentrations of other related elements in the lubrication oil. In this study, the wear growth rate (WGR) of each element in an oil sample is used to serve as the wear characteristic index of the element for oil debris analysis. The WGR of an element in the oil sample can be represented as
where
y is the concentration of the element, and the subscript
i is the index of the oil sample.
If the wear phenomenon occurs on any part of the gearbox, the WGR of each element increases with wear progression in the gearbox. While there are multiple worn parts in the gearbox and contents of elements of different parts are different, the WGR of each element of the same part is constant. This constancy principle is fundamentally attributed to the invariant elemental composition percentages within each component’s metallurgical matrix, ensuring that wear debris liberation occurs in proportional accordance with the predetermined material stoichiometry. To analyze multivariate relationships between the concentration of the Fe element and those of other related elements, an oil debris analysis method based on the MOGA-GM(1,
N) model is proposed to analyze wear conditions of main parts of the gearbox using weight coefficients of other related elements to the WGR of the Fe element. The WGR of the Fe element of a ring gear in the gearbox can be represented as
where
is the proportion of the Fe element content of the ring gear, and
is the wear degree of the ring gear in the
oil sample. The geometric factor
represents the sectoral wear analysis approach, where
denotes the effective radius reduction from the original unworn state, and the equivalence relationship validates that co-located elements exhibit identical wear progression rates within the same component.
3. Multi-Objective Genetic Algorithm
Considering the limited number of oil samples and nonlinear distribution of oil monitoring data, the GM(1,
N) model is more practical and accurate than other traditional multi-variable linear methods for analyzing the relationship between multivariate parameters of small data samples. The MOGA has high adaptability to solve multi-objective optimization problems with enhanced parallelism and stability [
49]. Therefore, the oil debris analysis method is proposed for multi-objective and multi-variable optimization based on a MOGA-GM(1,
N) model. Detailed procedures of the oil debris analysis method are shown in
Figure 1. The MOGA-GM(1,
N) model is developed, in which the concentration of a common element of Fe serves as a dependent variable, and concentrations of other related elements, i.e., Ni, Cr, Al, Mn, and Mo, serve as independent variables. An optimal weight coefficient of the concentration of each related element to the Fe element can be obtained by the MOGA with the minimum posterior difference ratio
C and the maximum small error probability
P as optimization objectives.
3.1. GM(1, N) Model
The system characteristic sequence of the GM(1,
N) model can be represented as
where
is a sequence of dependent variables,
are sequences of independent variables, and
n is the number of data samples. The 1-AGO sequence of
can be represented as
The GM(1,
N) model can be represented as
where
is the weight coefficient of the
ith independent variable to the dependent variable, and
a is a non-random constant. Equation (
5) can be rewritten as
where
,
, and
Using the least square method, weight coefficients of independent variables in the GM(1,
N) model can be obtained by
Based on Equation (
5), the GM(1,
N) model of the oil debris analysis method, in which the concentration of the Fe element serves as the dependent variable, and concentrations of Ni, Cr, Mo, Al, and Mn elements serve as independent variables, can be represented as
3.2. MOGA-GM(1, N) Model
The genetic algorithm (GA) is an objective optimization method, which is developed based on the theory of biological evolution and genetics [
50]. The MOGA is an improved multi-objective optimization method based on the GA. The posterior difference ratio
C and the small error probability
P are two key indicators for the accuracy test of the GM(1,
N) model. Therefore, the posterior difference ratio
C and the small error probability
P work as objective functions to realize effective oil debris analysis. A weighting method for the MOGA is used here to transform multi-objective optimization functions into a single-objective optimization problem. To improve the optimization efficiency and accuracy of weight coefficients of the MOGA, an adaptive crossover and a mutation operator are used in this study. Moreover, an elite strategy for the MOGA is added to ensure the population diversity of the MOGA and accelerate the convergence rate. The flowchart of the MOGA for oil debris analysis is shown in
Figure 2. Algorithm 1 presents the systematic implementation procedure for the MOGA-GM(1,
N) model applied to oil debris analysis, outlining the sequential steps from variable definition through optimization convergence.
Algorithm 1 MOGA-GM(1, N) model for oil debris analysis |
- 1:
Define the dependent variable (Fe concentration) and independent variables (Cr, Ni, Al, Mn, Mo concentrations) of the MOGA-GM(1, N) model - 2:
Perform 1-AGO accumulative processing of variables - 3:
Establish the GM(1, N) model - 4:
Establish objective functions: - 5:
Posterior difference ratio C - 6:
Small error probability P - 7:
Establish the MOGA-GM(1, N) model: - 8:
Initialize population - 9:
while not converged do - 10:
Evaluate fitness (minimize C, maximize P) - 11:
Select parents - 12:
Apply crossover and mutation - 13:
Generate new population - 14:
end while - 15:
Obtain optimal weight coefficients for independent variables - 16:
Analyze wear conditions based on weight coefficients
|
(1) Adaptive crossover-mutation probability: selection ranges of the crossover and the mutation probability in the GA model are (0.4, 0.9) and (0.001, 0.1), respectively. An adaptive crossover-mutation probability
p is proposed for adaptive target search to improve the searching capability of MOGA and ensure the diversity of population evolution. If the individual fitness is greater than the average fitness value, a large adaptive crossover-mutation probability
p enables individuals to search in the direction near the optimal fitness value. Suppose the individual fitness is less than the average fitness value. In that case, the search trend is reasonable, and the adaptive crossover-mutation probability
p can be selected to improve the local search accuracy of the MOGA. The adaptive crossover-mutation probability can be represented as
where
and
are the maximum and the minimum probabilities of crossover and mutation factors, respectively,
is the fitness value of the current individual,
and
are the maximum and the minimum fitness values, respectively, and
is the average fitness value.
(2) Optimization variables: weight coefficients of concentrations of Ni, Mn, Al, Cr, and Mo elements to the concentration of the Fe element need to be optimized, which serve as optimization variables of the MOGA-GA(1, N) model. The constraint of each weight coefficient of the concentration of any other element to the concentration of the Fe element is .
(3) Fitness function: the least squares method is used to analyze relationships of weight coefficients between the concentration of the Fe element and concentrations of other elements in the GM(1,
N) model. However, the accuracy is insufficient to guarantee that the posterior difference ratio
C and the small error probability posterior
P of the GM(1,
N) model are of the first-order accuracies. Weight coefficients of the MOGA for oil debris analysis are also unconstrained. Therefore, negative values of these weight coefficients may occur. The proposed MOGA-GM(1,
N) model aims to ensure that the posterior difference ratio
C and the small error probability posterior
P exhibit first-order accuracies, as shown in
Figure 3.
The GM(1, N) model selects the minimum posterior difference ratio C and the maximum small error probability posterior P as fitness functions and uses MOGA instead of the least squares algorithm to obtain the optimal multi-variable relationship between the concentration of the common element of Fe and those of other elements. According to optimization constraints, the main wear sources of the Fe element are examined by their weight coefficients.
The posterior difference ratio
C can be represented as
where
is the residual variance of the concentration of the Fe element,
is the variance of the concentration of the Fe element,
is the residual mean value of residual errors that is
is the residual error of the concentration of the Fe element that can be represented as
is the sequence of the concentration of the Fe element,
is the simulated sequence of the concentration of the Fe element, and
is the mean value that is
Based on Equations (
11) and (
12), the small error probability
P can be represented as
The fitness function
Y of the MOGA-GM(1,
N) model can be represented as
where
and
are weight coefficients of the posterior difference ratio and the small error probability, respectively.
5. Conclusions
This work presents an oil debris analysis method based on the MOGA-GM(1, N) model for gearboxes to predict their wear conditions by using lubrication oil spectral analysis. The proposed MOGA-GM(1, N) model provides a new approach for the wear location analysis of parts of the gearbox. The MOGA-GM(1, N) model can analyze wear conditions of parts, including different materials with multiple elements. Based on prediction results obtained by the proposed MOGA-GM(1, N) model for oil debris analysis of gearboxes, some meaningful conclusions are drawn below.
(1) Wear growth rates of concentrations of elements in lubrication oil of the gearbox are used for wear prediction of the gearbox. Wear location analysis of parts is realized by studying weight coefficients between the concentration of the common element and those of other related elements.
(2) The MOGA-GM(1, N) model is developed to analyze complex relationships between the concentration of the Fe element and those of other elements in oil samples. The posterior difference ratio and the small error probability of the accuracy test of the GM(1, N) model are used as objective functions of the MOGA-GM(1, N) model.
(3) According to the oil debris analysis method based on the MOGA-GM(1, N) model, numerical results show that the main wear part of the Fe element is the gear pump of the gearbox, and some slight wear also occurs on the sun gear, planetary gears, and the gear shaft. Wear experimental results verify the feasibility and accuracy of the MOGA-GM(1, N) model. By comparing prediction results of the MOGA-GM(1, N) model and the GM(1, N) model, the MOGA-GM(1, N) model can improve the accuracy of oil debris analysis, and describe complex relationships of WGRs of multiple elements of the gearbox.