Fault diagnosis method for engine lubrication system based on FRS-IMPA-XGBoost

Abstract

The invention discloses an engine lubrication system fault diagnosis method based on FRS-IMPA-XGBoost, and relates to the field of engine lubrication systems. Oil liquid data including abrasive particle information and various physicochemical information of lubricating oil are collected through an integrated sensor. Using a fuzzy rough set to reduce oil data attributes, and retaining features with high fault correlation degree to construct a data set; sobol sequences and dynamic reverse learning are introduced in the initialization stage of a marine predator algorithm, and the problems that initial solutions of the algorithm in the initialization stage are uneven in distribution and prone to falling into local optimum are solved; secondly, a self-adaptive parameter control strategy is introduced in the middle and later periods of the algorithm to further improve the balance of exploration and development, and an improved marine predator algorithm IMPA is obtained; the XGBoost model is optimized by using the improved IMPA (International Messaging and Presence Amplification); and the hyper-parameter obtained after optimization is finished is input into the XGBoost model. According to the method, the global search capability is enhanced, the conversion of algorithm exploration and development is balanced, and the accuracy of model fault diagnosis is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of engine lubrication, and in particular to a fault diagnosis method for an engine lubrication system based on FRS-IMPA-XGBoost. Background Art

[0002] The performance of the engine directly determines the maneuverability of the vehicle, and the lubrication system, as the "lifeline" of the engine, is crucial to the performance of the engine. Its core function is to reduce the friction between the moving parts inside the engine, reduce mechanical wear, and effectively control the operating temperature of the system to ensure that the engine operates stably and reliably under high-intensity and complex working conditions. Therefore, accurate evaluation and diagnosis of the health status of the engine lubrication system plays a vital role in the predictive maintenance of the vehicle.

[0003] In the field of engine fault diagnosis, traditional methods mainly rely on physical models and empirical rules, but these methods often have problems such as long time consumption, complex models, and high investment costs. Especially in engine lubrication system fault diagnosis, due to its complexity, traditional methods face challenges such as low diagnostic accuracy and high data dimensionality.

[0004] Therefore, the prior art still needs to be improved and enhanced. Summary of the invention

[0005] In view of the above-mentioned shortcomings of the prior art, the object of the present invention is to provide an engine lubrication system fault diagnosis method based on FRS-IMPA-XGBoost, aiming to solve the problem that in the prior art, in the engine lubrication system fault diagnosis, the traditional method relies on physical models and empirical rules, resulting in the traditional method facing the problems of low diagnostic accuracy and high data dimension.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for diagnosing engine lubrication system faults based on FRS-IMPA-XGBoost comprises the following steps:

[0008] Step S01, using the oil data collected by the integrated sensor on the engine lubrication system as the original data, and dividing the samples into a training set and a test set;

[0009] Step S02: normalize the original data, use the fuzzy rough set (FRS) algorithm to simplify the attributes, and retain the features with high fault correlation to construct a data set;

[0010] Step S03, introducing the Sobol sequence in the initialization stage of the marine predator algorithm (MPA) to generate uniformly distributed solutions, and adding a dynamic reverse learning strategy (DOL) to expand the initial search range and speed up the solution speed; in the middle and late stages of the MPA algorithm iteration, introducing the adaptive parameter CF to further improve the balance between exploration and development, and obtaining an improved marine predator algorithm (IMPA);

[0011] Step S04, using the improved marine predator algorithm IMPA to optimize the hyperparameters in the XGBoost model, such as learning rate, maximum depth of the tree, subsample ratio, and regularization parameters (λ, γ), etc.;

[0012] Step S05: input the parameters obtained after the optimization into the XGBoost model to establish an IMPA-XGBoost fault diagnosis model;

[0013] Step S06: Use the trained diagnostic model to perform fault diagnosis and output the diagnostic result.

[0014] Furthermore, in step S03:

[0015] Initialization stage: Introduce the Sobol sequence to produce uniformly distributed solutions; introduce the dynamic reverse learning strategy (DOL) to speed up the convergence of the algorithm.

[0016] The Sobol sequence is a low-discrepancy sequence that replaces traditional random sampling with uniformly distributed sample points. Each point of the Sobol sequence is generated by a recursive formula, and each generated Sobol sequence value x can be expressed as a binary decimal, thereby increasing the diversity of the initial population and making its distribution more uniform, avoiding falling into local optimality at an early stage, and improving the convergence speed of the algorithm. Therefore, the algorithm description of its generation process is:

[0017]

[0018] In the formula, x n is the nth point (vector) of the Sobol sequence, each component of which ranges from [0,1); n represents the nth point in the sequence; b i (n) represents the value of the i-th bit in the binary representation of n (0 or 1); v i is the value of the i-th direction vector; Represents a bitwise exclusive OR (XOR) operation.

[0019] The single-dimensional result of the sobol sequence can be expressed as:

[0020]

[0021] Where m represents the number of direction vectors; b i (n) is the ith bit of the binary representation of n, v i is the real value of the i-th direction vector.

[0022] The introduction of dynamic reverse learning includes:

[0023] In the formula, x ij is the current value of the variable at position (i,j), is the updated variable value; a ij =min(x i,j (t)) represents the minimum value of the current search space, b ij =max(x i,j (t)) is its maximum value, t is the current number of iterations; i∈{1,N], j∈{1,D]; N is the population size, D is the dimension of the solution space; k is the reverse coefficient, and its value range is (0,1).

[0024] Furthermore, the initialization phase of step S03 further defines an elite matrix Elite, corresponding to the fitness value in each prey matrix. During the iteration process, the fitness of each individual in the prey matrix is ​​calculated, and if the fitness is better than the fitness of the corresponding position in the Elite matrix, the Elite matrix is ​​updated.

[0025] The technical solution adopted by the present invention has the following beneficial effects:

[0026] 1. The present invention uses a fuzzy rough set (FRS) algorithm to preprocess lubricating oil data. The fuzzy rough set is suitable for processing uncertainty, ambiguity and incomplete information. The method is also suitable for processing high-dimensional features and retaining feature data with high fault correlation.

[0027] 2. The present invention optimizes the extreme gradient boosting tree (XGBoost) by using the improved marine predator optimization algorithm (IMPA) to construct an IMPA-XGBoost fault diagnosis model. The experimental results show that this method effectively improves the accuracy and efficiency of the model in fault diagnosis of the engine lubrication system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the engine lubrication system fault diagnosis method based on FRS-IMPA-XGBoost of the present invention;

[0029] Figure 2 This is a result graph of attribute simplification using fuzzy rough sets in the present invention;

[0030] Figure 3 This is a schematic diagram of the improved marine predator algorithm of the present invention;

[0031] Figure 4 is the fitness curve of the IMPA-XGBoost model and the MPA-XGBoost model of the present invention;

[0032] Figure 5 This is a comparison chart of the predicted value and the true value of the PSO-XGBoost model of the present invention;

[0033] Figure 6 It is a comparison chart of the predicted value and the true value of the MPA-XGBoost model of the present invention;

[0034] Figure 7 This is a comparison chart of the predicted value and the true value of the IMPA-XGBoost model of the present invention. DETAILED DESCRIPTION

[0035] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0036] The purpose of the present invention is to provide a method for diagnosing engine lubrication system faults, aiming to improve the accuracy of engine lubrication system fault diagnosis.

[0037] The present invention simplifies the oil data attributes through fuzzy rough set (FRS), retains the characteristics of the mansion under the fault height as the input of the diagnosis model; then, introduces the Sobol sequence to initialize the population, and combines dynamic reverse learning (DOL) to expand the initial search range and improve the solution speed. In order to balance the ability of exploration and development, the adaptive parameter CF is introduced, and the hyperparameters of the XGBoost model are optimized through the IMPA algorithm to establish a fault diagnosis model, and compared with the PSO-XGBoost and MPA-XGBoost models. The experimental results show that this method can effectively diagnose faults in the engine lubrication system, and the performance is improved in terms of accuracy and diagnostic precision.

[0038] Example

[0039] A fault diagnosis method for engine lubrication system based on FRS-IMPA-XGBoost, such as Figure 1 As shown, the following steps are included:

[0040] Step S01, the oil data collected by the integrated sensor on the engine lubrication system, a total of 500 groups of engine fault data are selected, including wear mechanical fault, sensor bias fault, lubricating oil cooling system fault, sensor failure fault and oil pump fault. Each type of data is 100 groups, of which 70 groups are divided into training set data and 30 groups are test set data, and the category label is set to 0-4.

[0041] Step S02: Perform maximum-minimum normalization on the data to make it fall within [0,1]; then use fuzzy rough set (FRS) to simplify its attributes and retain the features with strong correlation. After fuzzy rough set simplification, the features finally retained are: number of abrasive particles (<100um), contamination, temperature, moisture, viscosity, and density. The significance results of each feature are as follows: Figure 2 shown.

[0042] Step S03, in the initialization stage of the marine predator algorithm (MPA), because the randomly generated initial solutions are unevenly distributed and prone to local optimal problems, the Sobol sequence is introduced to make the initial solutions evenly distributed in the search space, and dynamic reverse learning is added to speed up the convergence of the algorithm and avoid it from falling into the local optimal solution; in the middle and late stages of the MPA algorithm iteration, the adaptive parameter CF is introduced to further improve the balance between exploration and development, and the improved marine predator algorithm (IMPA) is obtained. The improved marine predator algorithm (IMPA) is as follows Figure 3 shown.

[0043] The step S03 specifically includes:

[0044] (1) Initialization phase: Since the initial solution randomly generated by the marine predator algorithm will lead to uneven distribution of solutions, it is easy to fall into the problem of local optimal solution. In order to overcome this difficulty, the Sobol sequence is introduced to ensure that the initial solution is evenly distributed in the search space. The Sobol sequence can be expressed as:

[0045]

[0046] In the formula, x n Represents the nth point of the Sobol sequence, the interval range is [0,1); d i is the ith binary coefficient of n (0 or 1); v i is the direction vector used to adjust the weight of the i-th dimension; m represents the number of square vectors. The larger m is, the denser the generated points are and the better the uniformity is.

[0047] The initial solution of the algorithm can be expressed as follows after adding the Sobol sequence:

[0048] X ij =X min +Sobol j (X max -X min );

[0049] Where, X min To find the minimum value of the variable in the problem, X max is its maximum value; Sobol jis the jth low discrepancy value generated by the Sobol sequence, ranging from [0,1].

[0050] On the basis of introducing the Sobol sequence, combined with the dynamic reverse learning strategy, the convergence speed of the algorithm is accelerated, effectively avoiding the emergence of local optimal solutions. The formula of dynamic reverse learning is:

[0051]

[0052] In the formula, x ij is the current value of the variable at position (i,j), is the updated variable value; a ij =min(x i,j (t)) represents the minimum value of the current search space, b ij =max(x i,j (t)) is its maximum value, t is the current number of iterations; i∈{1,N], j∈{1,D]; N is the population size, D is the dimension of the solution space; k is the reverse coefficient, and its value range is (0,1).

[0053] In addition, a prey matrix Prey and an elite matrix Elite need to be constructed in the initialization phase. The Prey matrix is ​​mainly used to store the prey individual X in each iteration. ij Position and calculate its optimal fitness value. The Elite matrix is ​​composed of n copies of the individual with the best fitness.

[0054] The Prey matrix and Elite matrix are defined as follows:

[0055]

[0056] Where, X I Represents the optimal predator vector generating the elite matrix by repeating n times; n represents the population size; d is the dimension of the solution to the problem.

[0057] (2) MPA optimization stage

[0058] The marine predator algorithm MPA considers the velocity ratio (v) of prey and predator and defines a three-stage position update formula.

[0059] ① High speed ratio (v>10): In the initial stage of algorithm iteration, the predator stops chasing the prey due to its slow speed and remains stationary, while the prey completes the search through Brownian motion and adds a dynamic reverse learning strategy. The mathematical model of this stage is:

[0060] when hour,

[0061]

[0062] In the formula, i represents the i-th prey individual in the population; is the moving step length; For an elite matrix constructed by top predators; is the position vector of the i-th prey individual; is a normally distributed random vector of Brownian motion, which simulates the Brownian motion of the prey after being dot-multiplied with the prey position; P = 0.5; R is a uniform random number in the interval [0,1]; Iter represents the current iteration number, and Max_Iter represents the maximum iteration number.

[0063] In the middle and late stages of the MPA algorithm iteration, the step size control parameter CF is crucial to balance exploration and exploitation. CF gradually decreases from 1 to 0 as the iteration proceeds. A larger CF enhances global exploration capabilities, while a smaller CF promotes local exploitation and rapid convergence. Therefore, in order to further improve the balance between exploration and exploitation, enhance global search capabilities and accelerate local convergence.

[0064] The new adaptive parameter control strategy CF is:

[0065]

[0066] Where CF is the adaptive parameter; Iter represents the current iteration number, and Max_Iter represents the maximum iteration number.

[0067] ② Same speed ratio (v≈1): This stage is in the middle of the algorithm iteration, and the algorithm transitions from exploration to development. At this time, the predator and the prey each look for their target, and the entire population is divided into two parts: the predator is responsible for searching (Brownian motion), while the prey is responsible for development (Lévy flight). The adaptive parameter control strategy CF is introduced at this stage to improve the balance between exploration and development. The behavior of this stage is described by the mathematical model:

[0068] when hour,

[0069] The first half of the population Prey Development:

[0070]

[0071] The second half of the population Predator Exploration:

[0072]

[0073] In the formula, i represents the i-th prey individual in the population; is the moving step length; an elite matrix constructed for top predators; is the position vector of the i-th prey individual; P = 0.5; R is a uniform random number in the interval [0,1]; is a random number vector simulating Lévy motion; and Multiplication simulates prey performing Lévy motion; is a normally distributed random vector of Brownian motion, Multiplying with Elite represents the Brownian motion of the predator; CF is the adaptive parameter; Iter represents the current iteration number, and Max_Iter represents the maximum iteration number.

[0074] ③ Low speed ratio (v = 0.1): This stage is in the late stage of algorithm iteration, and is mainly to improve the development ability of the algorithm. During this period, the predator's strategy is Lévy movement:

[0075] when hour,

[0076]

[0077] In the formula, i represents the i-th prey individual in the population; is the moving step length; an elite matrix constructed for top predators; is the position vector of the i-th prey individual; P = 0.5; is a random number vector simulating Lévy motion, and Multiplication simulates the predator's Lévy motion; CF is an adaptive parameter; Iter represents the current iteration number, and Max_Iter represents the maximum iteration number.

[0078] (3) FADs effect: This stage shows that in addition to the influence of its own factors, the behavior of predators is also constrained by external environmental factors (such as eddy currents or FADs aggregation). These factors cause predators to adjust their hunting paths in higher dimensions and longer periods of time to search for new areas of prey. FADs are the local optimal exploration areas. By increasing the length of the jump, the algorithm can avoid falling into the local optimal solution during the iteration process.

[0079]

[0080] In the formula, is the position vector of the i-th prey individual; is the minimum vector of the search space, is its maximum value; is a random vector with uniform distribution between [0,1]; FADs is a parameter for controlling the optimization probability of the algorithm, which is set to 0.2; U∈[0,1] is a random number, if U<0.2, U is constructed from 0; if U≥0.2, U is constructed from 1; r∈[0,1] is a random number; r1 and r2 are random indices of the prey matrix.

[0081] (4) Ocean memory stage: This stage is mainly used to update and calculate the Elite matrix and find the fitness of the best individual in the Elite matrix. For each individual in the prey matrix, its fitness is calculated. If the fitness is better than the fitness of the corresponding position in the Elite matrix, the Elite matrix is ​​updated.

[0082] Step S04: Use the improved marine predator algorithm IMPA to optimize the hyperparameters in the XGBoost model, such as learning rate, maximum depth, subsample ratio, and regularization parameters (λ, γ).

[0083] Step S05: input the parameters obtained after the optimization into the XGBoost model to establish the IMPA-XGBoost fault diagnosis model.

[0084] Through the simulation comparison experiment of IMPA-XGBoost and MPA-XGBoost fault diagnosis models, the error rate on the cross-validation set is selected as the fitness value, the population size is set to 20, the maximum number of iterations is set to 100, and the number of hyperparameters is 5. The results are as follows Figure 4 As shown. Figure 4 It can be seen that the improved IMPA algorithm is significantly better than the MPA algorithm in performance. In the early stage of the algorithm, the introduction of dynamic reverse learning increased the convergence speed of the algorithm to avoid it falling into the local optimum. The IMPA algorithm converged to the optimum before 20 iterations, while the MPA algorithm converged to the optimum after 40 iterations because it fell into the local optimum.

[0085] Depend on Figure 2 It can be seen that the attributes of oil data are simplified by fuzzy rough set, and 6 groups of data with higher comprehensive scores are retained to establish a feature signal data set. The final selected features are: contamination, viscosity, density, temperature, moisture and total number of wear particles as simplified data. A total of 500 groups of engine fault data were selected, including five types of faults: wear mechanical fault, sensor bias fault, lubricating oil cooling system fault, sensor failure fault and oil pump fault. Each type of data is 100 groups, of which 70 groups are divided into training set data and 30 groups are test set data. The category label is set to 0~4. The details are shown in Table 1.

[0086] Table 1 Classification of five types of faults

[0087]

[0088] In order to verify the effectiveness of the proposed model, the data after fuzzy rough set attribute reduction is used as the initial training sample, and the IMPA-XGBoost fault diagnosis model is compared with the PSO-XGBoost and MPA-XGBoost fault diagnosis models. The results are shown in Figure 5-Figure 7 As shown. Figure 5-Figure 7 It can be seen that the classification accuracy of the improved IMPA-XGBoost fault diagnosis model reaches 93.333%, which is 10.003% and 6.668% higher than the traditional PSO-XGBoost and MPA-XGBoost models. In order to avoid errors and the contingency of experimental results, 10 fault diagnosis experiments were conducted on several algorithms, and the maximum and minimum accuracy of each algorithm were taken, and then the mean was taken as the final diagnosis result. The comparison results are shown in Table 2.

[0089] Table 2 Comparison of four fault diagnosis models

[0090]

[0091] Comprehensive data analysis shows that the IMPA-XGBoost fault diagnosis model is superior to several other diagnosis models in terms of accuracy, verifying that the improved IMPA-XGBoost model has better diagnostic accuracy.

[0092] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and cannot be regarded as limiting the present invention. Those skilled in the art may make changes, modifications, substitutions and variations to the above embodiments, which are all within the scope of the present invention.

Claims

1. A fault diagnosis method for engine lubrication system based on FRS-IMPA-XGBoost, characterized in that: The steps include: Step S01, using the oil data collected by the integrated sensor on the engine lubrication system as the original data, and dividing the samples into a training set and a test set; Step S02: normalize the original data, use the fuzzy rough set (FRS) algorithm to simplify the attributes, and retain the features with high fault correlation to construct a data set; Step S03, introducing the Sobol sequence in the initialization stage of the marine predator algorithm MPA to generate uniformly distributed solutions, and adding the dynamic reverse learning strategy DOL to expand the initial search range and speed up the solution speed; in the middle and late stages of the MPA algorithm iteration, introducing the adaptive parameter CF to further improve the balance between exploration and development, and obtaining the improved marine predator algorithm IMPA; Step S04, using the improved marine predator algorithm IMPA to optimize the hyperparameters in the XGBoost model, such as the learning rate, the maximum depth of the tree, the subsample ratio, and the regularization parameters λ, γ, to obtain the optimized parameters; Step S05: input the optimized parameters into the XGBoost model to establish an IMPA-XGBoost fault diagnosis model; Step S06: train the IMPA-XGBoost fault diagnosis model using the training set, and use the trained IMPA-XGBoost fault diagnosis model to perform fault diagnosis and output the diagnosis result.

2. The nuclear fuel rod out-of-pile heating test device according to claim 1, characterized in that: In step 03: Initialization stage: The Sobol sequence is introduced into the marine predator algorithm to ensure that the initial solution is evenly distributed in the search space. The Sobol sequence can be expressed as: Where m represents the number of direction vectors; b i (n) is the ith bit of the binary representation of n, v i is the real value of the i-th direction vector; The initial solution of the algorithm can be expressed as follows after adding the Sobol sequence: X ij =X min +Sobol j (X max -X min ); Where, X min To find the minimum value of the variable in the problem, X max is its maximum value; Sobol j is the jth low discrepancy value generated by the Sobol sequence, ranging from [0,1].

3. The nuclear fuel rod out-of-pile heating test device according to claim 1, characterized in that: In step 03, on the basis of introducing the Sobol sequence, a dynamic reverse learning strategy is combined to accelerate the convergence speed of the algorithm. The formula of dynamic reverse learning is: In the formula, x ij is the current value of the variable at position (i,j), is the updated variable value; a ij =min(x i,j (t)) represents the minimum value of the current search space, b ij =max(x i,j (t)) is its maximum value, t is the current number of iterations; i∈{1,N], j∈{1,D]; N is the population size, D is the dimension of the solution space; k is the reverse coefficient, and its value range is (0,1).

4. The nuclear fuel rod out-of-pile heating test device according to claim 1, characterized in that: In step 03, in the MPA optimization stage, the step size control parameter CF is crucial to balance exploration and development. CF gradually decreases from 1 to 0 as the iteration proceeds. The new adaptive parameter control strategy formula is as follows: Where: CF is the adaptive parameter; Iter is the current number of iterations, Max_Iter is the maximum number of iterations; The improved position update is as follows: when hour, The second half of the population Predator Exploration: In the formula, i represents the i-th prey individual in the population; is the moving step length; an elite matrix constructed for top predators; is the position vector of the i-th prey individual; P = 0.5; is a normally distributed random vector of Brownian motion, Multiplying with Elite represents the Brownian motion of the predator; CF is the adaptive parameter; Iter represents the current iteration number, and Max_Iter represents the maximum iteration number; when hour, In the formula, i represents the i-th prey individual in the population; is the moving step length; an elite matrix constructed for top predators; is the position vector of the i-th prey individual; P = 0.5; is a random number vector simulating Lévy motion, and Multiplication simulates the predator's Lévy motion; CF is an adaptive parameter; Iter represents the current iteration number, and Max_Iter represents the maximum iteration number.

5. The nuclear fuel rod out-of-pile heating test device according to claim 1, characterized in that: In step 03, in the FADs effect stage, this stage shows that in addition to the influence of its own factors, the behavior of the predator is also restricted by external environmental factors. FADs is the local optimal exploration area. By increasing the jump length, the algorithm can avoid falling into the local optimal solution during the iteration process: In the formula, is the position vector of the i-th prey individual; is the minimum vector of the search space, is its maximum value; is a random vector with uniform distribution between [0,1]; FADs is a parameter for controlling the optimization probability of the algorithm, which is set to 0.2; U∈[0,1] is a random number, if U<0.2, U is constructed from 0; if U≥0.2, U is constructed from 1; r∈[0,1] is a random number; r1 and r2 are random indices of the prey matrix; In the ocean memory stage: update and calculate the Elite matrix, and find the fitness of the best individual in the Elite matrix; for each individual in the prey matrix, calculate its fitness, and when the fitness is better than the fitness of the corresponding position in the Elite matrix, update the Elite matrix.

6. The nuclear fuel rod out-of-pile heating test device according to claim 2, characterized in that: The initialization phase of step S03 also constructs a prey matrix Prey and an elite matrix Elite; wherein the Prey matrix is ​​mainly used to store the prey individual X in each iteration process. ij Position and calculate its optimal fitness value. The Elite matrix is ​​composed of n copies of the individual with the best fitness.

Images