CN119167213A

CN119167213A - A gyroscope group fault prediction method based on IHBA-RF

Info

Publication number: CN119167213A
Application number: CN202411686931.1A
Authority: CN
Inventors: 李英顺; 王德彪; 刘海洋; 史琳; 丛天龙; 张杨; 于浚豪; 赵玉鑫
Original assignee: Shenyang Shunyi Technology Co ltd
Current assignee: Shenyang Shunyi Technology Co ltd
Priority date: 2024-11-25
Filing date: 2024-11-25
Publication date: 2024-12-20
Anticipated expiration: 2044-11-25
Also published as: CN119167213B

Abstract

The invention relates to the technical field of artificial intelligent fault diagnosis, and discloses a fault prediction method of a gyroscope group based on IHBA-RF, which comprises the steps of performing degradation and normalization preprocessing on collected data through a KPCA algorithm, screening pin signal data with high association degree as a model input data set, improving a mele optimization algorithm HBA, introducing a Sine chaotic map in an HBA population initialization stage, calculating the adaptation degree, introducing a segmented optimal neighborhood strategy in an HBA mining stage and a honey searching stage to enhance global searching capability, performing parameter optimization on a random forest algorithm RF by adopting an improved mele optimization algorithm IHBA, constructing a IHBA-RF fault diagnosis model, training a training data set, inputting a test data set into the IHBA-RF prediction model, performing fault diagnosis on the gyroscope group, and outputting a prediction result. According to the invention, the main parameters of the Random Forest (RF) are optimized through the improved badger optimization algorithm (IHBA), so that the defect of blindness of parameter selection in the training process is overcome, and the prediction precision of the regression prediction model is improved.

Description

IHBA-RF-based gyroscope group fault prediction method

Technical Field

The invention relates to the technical field of artificial intelligence fault diagnosis, in particular to a method for predicting faults of a gyroscope group based on IHBA-RF.

Background

The gyroscope group is an important component of a gun control system and comprises a rate gyroscope, a temperature control plate, a power supply board, a detection board and the like, is used for accurately measuring the angular positions and the angular speeds of a gun turret and an artillery in the system and providing driving signals endowing the system with new positions of the artillery, plays an important role in battle field operation, increases the failure occurrence rate after frequent use, and can improve battle field operation capacity and firepower output and maximize the playing capability.

At present, various algorithms based on artificial intelligence are widely applied to the field of fault diagnosis, the randomness of kernel function selection and the limitation of large-scale training of a support vector machine lead to the lack of accuracy of a prediction result, the defects of difficult acquisition of expert system knowledge and limitation of a knowledge base to storage cannot guarantee the efficiency and the accuracy of the prediction result, a fault tree analysis method is large in calculation scale and complex in work, excessive memory occupation also leads to the slowing of the running speed in fault prediction, and compared with other algorithms, the Random Forest algorithm (RF) has the advantages of interpretation and robustness, the problem of complexity of an internal mechanism is perfectly solved, the RF has higher generalization and expression capability, the limitation of other machine learning algorithms in the field of fault diagnosis is improved to a certain extent through an integrated learning method, and the problems of weak generalization capability and the like are solved.

Disclosure of Invention

Aiming at the defects of large calculation scale, high complexity, complex work, lack of accuracy of a prediction result and the like of a fault analysis algorithm in the prior art, the invention adopts the technical scheme that the method for predicting the faults of the gyroscope group based on IHBA-RF is characterized by comprising the following steps:

s1, acquiring data of a gyroscope group pin signal;

S2, performing reduction and normalization preprocessing on the acquired data through a KPCA algorithm, and screening pin signal data with high association degree as a model input data set;

dividing the model input dataset into a test dataset and a training dataset;

S3, improving an optimization algorithm HBA of the badger, wherein the optimization algorithm HBA comprises the steps of introducing a fine chaotic map in an HBA population initialization stage, and calculating the fitness;

the method specifically comprises the following steps:

s301, introducing a fine chaotic map in an HBA population initialization stage,

,(2-1)

,(2-2)

Wherein: And The chaos number generated for the sine chaos mapping is 0.99,Is the position of the ith mel,AndThe lower boundary and the upper boundary of the search are respectively;

s302, calculating a density factor alpha and an intensity factor I, wherein the intensity is related to the concentration force of the prey and the distance between the prey and the meles, I is the smell intensity of the prey, if the smell is high, the movement speed is high, and vice versa, and the updating formula is as follows:

,(2-3)

where I is an intensity factor, S is a source or concentration intensity, Represents the distance between the prey and the current badger individual,In order to be a location of a prey,Is the position of the ith mel,Is the position of the (i+1) th badger;

density factor update-density factor alpha controls time-varying randomization to ensure a smooth transition from mining to honey production, the formula is as follows:

,(2-4)

wherein alpha is a density factor, Represented as the maximum number of iterations,;

S303, introducing a segmentation optimal neighborhood strategy in the HBA mining stage and the honey searching stage to enhance the global searching capability, namely setting a direction correction coefficient F for changing the searching direction so as to enable a searching individual to strictly scan the searching space by using a large number of opportunities,

During the digging phase, the meles perform actions similar to the shape of heart lines, and heart-shaped movements can be simulated by the following formulas (2-5):

(2-5)

In the formula, The updated position of the badger individual; In order to be a location of a prey, As a result of the random parameters,Indicating the ability of the badger to obtain food, I is intensity factor, alpha is density factor,Is the distance between the prey and the current badger individual,Three different random numbers between [0,1] are represented, F is a direction correction coefficient, and the specific expression is:

,(2-6)

in the honey searching stage, the condition formula that the badgers reach the honeycomb along with the honey is as follows:

,(2-7)

In the formula, For the updated individual position of the meles,In order to be a location of a prey,Is a random parameter, F is a direction correction coefficient, and alpha is a density factor,Is a random number between 0 and 1, and is based on distance informationBadger with honey at prey positionThe search is performed nearby, and at this stage the search is affected by the search behavior a, which varies with iteration.

At this time, a piecewise optimal neighborhood strategy is introduced in the mining stage and the honey searching stage to enhance the global searching capability, so that the situation that a better individual exists nearby an optimal individual is processed, and piecewise nonlinear decreasing weights are introduced in neighborhood parameters, wherein the piecewise nonlinear decreasing weights are defined as follows:,(2-8)

In the formula, For piecewise non-linear decreasing parameters, delta is a neighborhood parameter,Is a random number uniformly distributed in the interval [0,1 ];

three different decreasing trends are indicated for the non-linear decreasing coefficients respectively, For the current number of iterations,Is the maximum number of iterations.

The location update at the mining stage after the improvement of the segmentation optimal neighborhood strategy is adopted is as follows:

,(2-9)

In the formula, For a piecewise non-linear decrementing parameter,Is a prey location; in order to update the position of the badger, F is a direction correction coefficient, Indicating the ability of the badger to obtain food, I is intensity factor, alpha is density factor,Is the distance between the prey and the current badger individual,Representing three different random numbers between [0,1 ];

the position update of the honey searching stage after the improvement of the segmentation optimal neighborhood strategy is as follows:

,(2-10)

In the formula, A piecewise nonlinear decremental parameter; is a prey location; in order to update the position of the badger, F is a direction correction coefficient, Is a random parameter in [0,1], alpha is a density factor,Is the distance between the ith badger and its prey

S4, carrying out parameter optimization on a random forest algorithm RF by adopting an improved badger optimization algorithm IHBA, constructing a IHBA-RF fault diagnosis model, and training a training data set;

S5, inputting the test data set into a IHBA-RF prediction model, performing fault diagnosis on the gyroscope group, and outputting a prediction result.

As shown in fig. 2, in the step S2, the performing the down-scaling and normalization preprocessing on the collected data by using the KPCA algorithm specifically includes:

s201, constructing a decision matrix, namely setting n sample data with m dimensions to obtain an n X m decision matrix X,

,(1-1)

Wherein the method comprises the steps ofIs a vector, n is a positive integer;

mapping the matrix X to a d-dimensional high-dimensional space, wherein the mapping relation is as follows:

,(1-2)

Wherein R ^m is the low-dimensional space before mapping, R ^d is the high-dimensional space after mapping, and the new kernel matrix after mapping is:

,(1-3)。

S202, centering the kernel matrix, namely classifying sample data into zero vectors, and enabling a covariance matrix after dimension rising to be:

,(1-4)

Wherein C is a matrix of d dimensions; Is that Is a transposed matrix of (a); For the i-th mapping function, 。

S203, solving the eigenvalue, namely solving a formula according to the eigenvalue:

,(1-5)

wherein p is a feature vector, Substituting the formula (1-4) into the formula (1-5) for the characteristic coefficient, and omitting the coefficient to obtain the formula (1-6):

,(1-6)

divided by both sides of equation (1-6) Obtaining the formula (1-7)

;

In the formula,Is represented by a feature vector:

,(1-8)

In the formula, For an n-dimensional column vector, formula (1-8) is substituted into formula (1-6), and both sides are multiplied simultaneouslyObtaining formula (1-9)

,(1-9)

Order theIs an n-dimensional symmetrical semi-regular square matrix,The above formula can be simplified as:

,(1-10)

The nonzero eigenvalue obtained by the solution of the formula (1-10) corresponds to the nonzero eigenvalue of the formula (1-5), and the unitizing process is performed on p in the formula (1-5), and then:

,(1-11)。

s204, projecting the data processed in the step S203 onto the selected principal component to form a new low-dimensional representation, and obtaining projection t of the data on the feature vector p to obtain the reduced-dimensional data:

,(1-12)。

S205, carrying out normalization processing on the obtained reduced data by using a normalization formula, wherein the formula is as follows:

,(1-13)

In the formula, The data after normalization is represented and,The data that is to be normalized is represented,Represents the maximum value of the normalized data,Representing the minimum value of the normalized data.

The step S4 adopts an improved meles optimization algorithm IHBA to carry out parameter optimization on a random forest algorithm RF, and specifically comprises the steps of judging whether a random parameter r is smaller than 0.5, and updating the position of the meles by utilizing a segmentation optimal neighborhood strategy;

If the random parameter r is less than 0.5, the position information of the badger individual is updated according to the formula (2-9) in the mining stage, and if the random parameter r is more than or equal to 0.5, the position information of the badger individual is updated according to the formula (2-10) in the honey searching stage;

Judging whether the maximum iteration times are reached, if so, outputting the individual position and fitness value of the optimal middle group, and if not, re-executing S3, and updating the individual position of the badger again.

Compared with the prior art, the invention has the following beneficial technical effects and advantages:

1. Compared with other comprehensive evaluation algorithms (analytic hierarchy process, gray correlation analysis and the like), the KPCA algorithm is used for preprocessing the data, nonlinear data can be effectively processed, the data is mapped to a high-dimensional space through a kernel skill, a complex structure is captured, the dimension can be reduced, redundant characteristics are removed, the performance of a subsequent model is improved, and the method is good in noise reduction and is beneficial to improving the data quality;

2. compared with a Support Vector Machine (SVM), a decision tree and the like, the Random Forest (RF) can process various types of data, including discrete values and continuous values, can use some robust loss functions, can process nonlinear data, and has higher prediction accuracy under the condition of relatively less parameter adjustment;

3. The main parameters of the Random Forest (RF) are optimized through an improved badger optimization algorithm (IHBA), the defect of blindness of parameter selection in the training process is overcome, the prediction precision of the regression prediction model is improved, and compared with decision trees, support vector machines and GBDT algorithm experiments, IHBA-RF has higher prediction precision and practicability.

Drawings

FIG. 1 is a main body frame diagram of a method for predicting failure of a group of gyroscopes based on IHBA-RF of the present invention;

fig. 2 is a flowchart of the KPCA algorithm of the invention.

Detailed Description

The present invention will be described in detail with reference to the accompanying drawings, but the scope of the present invention is not limited by the accompanying drawings.

The invention provides a method for predicting failure of a gyroscope group based on IHBA-RF, which is shown in figure 1 and comprises the following steps:

S1, collecting data values of pin signals of a gyroscope group through an equipment test bed, and collecting port number data values as initial data of experiments.

dividing the model input dataset into a test dataset and a training dataset;

Further, S201, constructing a decision matrix, namely setting n sample data with m dimensions to obtain a decision matrix X with n X m,

,(1-1)

Wherein the method comprises the steps ofIs a vector, n is a positive integer;

,(1-2)

,(1-3)。

,(1-4)

,(1-5)

,(1-6)

divided by both sides of equation (1-6) Obtaining the formula (1-7)

,(1-7)

In the formula,Is represented by a feature vector:

,(1-8)

,(1-9)

,(1-10)

,(1-11)。

,(1-12)。

,(1-13)

s301, diversity of an initial population plays a vital role in convergence speed and convergence accuracy of an algorithm, and in order to further expand the range of the initial population to improve local searching capability and enable the initial population to find the position of an optimal solution more easily, a fine chaotic mapping and population filtering mechanism is adopted to optimize the initialization stage of the HBA, and the method is specifically shown as the expression:

,(2-1)

,(2-2)

Wherein: And The chaos number generated for the sine chaos mapping is 0.99,Is the position of the ith mel,AndThe lower and upper boundaries of the search, respectively.

,(2-3)

,(2-4)

And (3) jumping out of the local optimum, namely, jumping out of the local optimum region by the step and the next two steps. In this case, the proposed algorithm uses a flag F that changes the direction of the search to allow the searching individual to scan the search space strictly with a large number of opportunities.

(2-5)

,(2-6)

,(2-7)

In the formula, For the updated individual position of the meles,In order to be a location of a prey,Is a random parameter, F is a direction correction coefficient, and alpha is a density factor,Is a random number between 0 and 1, and is based on distance informationBadger with honey at prey positionThe search is performed nearby, and at this stage, the search is affected by the search behavior α that varies with iteration, and furthermore, a mel may be disturbed by F.

At this time, the piecewise optimal neighborhood strategy is introduced in the mining stage and the honey searching stage to enhance the global searching capability, so that the situation that better individuals exist nearby the optimal individuals is processed, and the problem that an effective processing algorithm is easy to trap in a local part can be solved, and therefore piecewise nonlinear decremental weights are introduced in neighborhood parameters. The definition is as follows:

,(2-8)

,(2-9)

,(2-10)

In the formula, A piecewise nonlinear decremental parameter; is a prey location; in order to update the position of the badger, F is a direction correction coefficient, Is a random parameter in [0,1], alpha is a density factor,Is the distance between the ith meles and their prey.

Further, a Random Forest (RF) algorithm is an intelligent integrated learning algorithm composed of a plurality of decision trees. Decision trees (Classification and Regression Trees, CART) are a statistical model that can be used to derive different classes or values after features are input. CART is used in RF as a base learner, and Gini index minimization criteria are used to select features and divide. Set up input sample set . In the classification problem, for a given sample D, it is assumed that there are K categories, the probability of which isThe base-Ni index is: . The variance of the leaf node P is defined as: , cut-off to leaf node p Average value of (2). Then dividing the next stage leaf node in the same mode until the preset node threshold value is reached, generating a final decision tree, building an estimation function after training is finished, and obtaining an estimation value Y by a new X through S. The random forest combines a plurality of decision trees together to realize data classification, and has the advantages of few adjustment parameters, high training speed, high estimation precision, strong generalization capability and the like. The algorithm extracts a plurality of sample data from the original data by a Bootstrap sampling method, and a new training sample set is constructed. Based on CART thought, establishing decision tree for each training set, and finally, averaging according to q decision tree results to obtain a final estimated value Y as follows: . The parameters involved in the random forest algorithm modeling mainly comprise learning rate (learning_rate) for controlling the step length of parameter updating during learning, if the step length is too large, the learning process may diverge, otherwise, too many iterations of the model may be caused, and the learning time is greatly increased; the maximum iteration number (n_ estimators) represents the number of basic learners, the number of the basic learners is required to be increased when the number of the basic learners is smaller and the learning_rate is smaller, so that training errors are converged, the sub-sampling (subsamples) is used for controlling the sample proportion of the data set participating in fitting, the variance of the whole model can be effectively reduced when the sample proportion is set to be smaller than 1, the overfitting is prevented, the maximum depth (max_depth) of the decision tree, the minimum sample number (min_samples_slit) required by internal node subdivision and the minimum sample number (min_samples_leaf) contained by leaf nodes are used for controlling the complexity of each tree, the specific value depends on data distribution, if the value is too large, the model structure is complex, the overfitting is easy to cause, and otherwise the underfitting is easy to cause.

In the embodiment, calculating the fitness value of each badger, determining a global optimal solution, selecting the first half of excellent badger individuals, and reserving the position information and fitness value information of the optimal badger individuals;

calculating a convergence factor value according to formula (2-4);

Judging whether the random number is smaller than 0.5 or not, and updating the position of the badger by utilizing a segmentation optimal neighborhood strategy;

if the random parameter r is less than 0.5, updating the position information of the badger individual according to the formula (2-9) when the random parameter r is less than 0.5, if the random parameter r is more than or equal to 0.5, updating the position information of the badger individual according to the formula (2-10) when the random parameter r is more than or equal to 0.5, judging whether the maximum iteration number is reached, outputting the individual position and the fitness value of the optimal middle group when the maximum iteration number is met, and if the maximum iteration number is not met, re-executing the step S3, and updating the position of the badger individual again.

The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims

1. A gyroscope group fault prediction method based on IHBA-RF, characterized by comprising the following steps:

S1, collect data of gyroscope group pin signal;

S2. Perform dimension reduction and normalization preprocessing on the collected data through KPCA algorithm, and select pin signal data with high correlation as the model input data set;

Divide the model input dataset into a test dataset and a training dataset;

S3. Improvements to the honey badger optimization algorithm HBA, including the introduction of Sine chaos mapping in the HBA population initialization phase to calculate fitness; the introduction of segmented optimal neighborhood strategy in the HBA mining phase and honey-seeking phase to enhance global search capabilities;

The specific steps include:

S301, introduce Sine chaos mapping in the HBA population initialization phase,

, (2-1)

, (2-2)

Where: and is the chaotic number generated by the sinusoidal chaotic mapping, the value of 𝜇 is 0.99, is the position of the i-th honey badger, and are the lower and upper boundaries of the search respectively;

S302, calculate the density factor α and the intensity factor I; the intensity is related to the concentration of the prey and the distance between the prey and the honey badger, and I is the smell intensity of the prey; if the smell is high, the movement speed is fast, and vice versa. The update formula is as follows:

, (2-3)

Where I is the intensity factor, S is the source intensity or concentrated intensity, Indicates the distance between the prey and the current honey badger individual, For the prey location, is the position of the i-th honey badger, is the position of the i+1th honey badger;

Density factor update: The density factor α controls the time-varying randomization to ensure a smooth transition from digging to honey collection. The formula is as follows:

, (2-4)

Where α is the density factor, Represents the maximum number of iterations, ;

S303, introducing the segmented optimal neighborhood strategy in the HBA mining stage and the honey-seeking stage to enhance the global search capability is specifically: setting a direction correction coefficient F to change the search direction, so as to take advantage of a large number of opportunities to allow the search individuals to strictly scan the search space,

During the digging phase, the honey badger performs a heart-shaped motion, which can be simulated by equation (2-5): (2-5)

In the formula, This is the updated individual position of the honey badger; For the prey location, is a random parameter, It represents the ability of honey badger to obtain food, I is the intensity factor, α is the density factor, is the distance between the prey and the current honey badger individual, Represents three different random numbers between [0,1], F is the direction correction coefficient, see the specific expression:

, (2-6)

In the honey-seeking stage, the formula for the honey badger following the honey guide badger to reach the hive is:

, (2-7)

In the formula, is the updated individual position of the honey badger, For the prey location, is a random parameter, F is a direction correction coefficient, and α is a density factor, is a random number between [0,1], based on the distance information Honey badger in prey position Search is performed nearby. In this phase, the search is affected by the search behavior α which changes with iterations;

At this time, the segmented optimal neighborhood strategy is introduced in the mining stage and the honey-seeking stage to enhance the global search capability, so as to deal with the situation where there are better individuals near the optimal individual. The segmented nonlinear decreasing weight is introduced in the neighborhood parameter, which is defined as follows: , (2-8)

In the formula, is the piecewise nonlinear decreasing parameter, δ is the neighborhood parameter, is a random number uniformly distributed in the interval [0,1];

are nonlinear decreasing coefficients, representing three different downward trends. is the current iteration number, is the maximum number of iterations;

The position update of the mining phase after the improved segmented optimal neighborhood strategy is as follows:

, (2-9)

In the formula, is the piecewise nonlinear decreasing parameter, For prey location; is the updated position of the honey badger, F is the direction correction coefficient, It represents the ability of honey badger to obtain food, I is the intensity factor, α is the density factor, is the distance between the prey and the current honey badger individual, Represents three different random numbers between [0,1];

The position update of the honey-seeking phase after the segmented optimal neighborhood strategy is improved is as follows: , (2-10)

In the formula, is the piecewise nonlinear decreasing parameter; For prey location; is the updated position of the honey badger, F is the direction correction coefficient, is a random parameter in [0,1], α is the density factor, is the distance between the i-th honey badger and its prey;

S4, using the improved honey badger optimization algorithm IHBA to optimize the parameters of the random forest algorithm RF, building an IHBA-RF fault diagnosis model, and training the training data set;

S5. Input the test data set into the IHBA-RF prediction model, perform fault diagnosis on the gyroscope group, and output the prediction results.

2. A gyroscope group fault prediction method based on IHBA-RF according to claim 1: In step S2, the collected data is preprocessed by dimension reduction and normalization using the KPCA algorithm, specifically:

S201, construct a decision matrix: suppose n sample data with m dimensions, and obtain an n*m decision matrix X ,

,(1-1)

in is a vector, n is a positive integer;

Map the matrix X to a d-dimensional high-dimensional space. The mapping relationship is as follows:

,(1-2)

Where Rm is the low-dimensional space before mapping, ^Rd is the high-dimensional space after mapping, ^and the new kernel matrix after mapping is:

,(1-3)

S202. Center the kernel matrix: return the sample data to the zero vector. After dimensionality increase, the covariance matrix is:

,(1-4)

Where C is a d-dimensional matrix; for The transposed matrix of is the i-th mapping function, ;

S203, eigenvalue solution: According to the eigenvalue solution formula:

,(1-5)

In the formula, p is the eigenvector, is the characteristic coefficient. Substitute formula (1-4) into formula (1-5) and omit the coefficient to obtain formula (1-6):

,(1-6)

Divide both sides of formula (1-6) by , we get formula (1-7)

;

In the formula, The linear combination of is represented by the eigenvector:

,(1-8)

In the formula, is an n-dimensional column vector. Substitute formula (1-8) into formula (1-6) and multiply both sides by Get formula (1-9)

, (1-9)

make is an n-dimensional symmetric semi-positive definite matrix, , then the above formula can be simplified to:

,(1-10)

The non-zero eigenvalues obtained by solving formula (1-10) are equivalent to the non-zero eigenvalues of formula (1-5). If p in formula (1-5) is normalized, we have:

, (1-11)

S204, projecting the data processed in step S203 onto the selected principal component to form a new low-dimensional representation, obtaining the projection t of the data onto the feature vector p , and obtaining the reduced-dimensional data:

, (1-12)

S205, using a normalization formula to normalize the obtained dimension reduction data, the formula is:

, (1-13)

In the formula, represents the normalized data. represents the normalized data, Represents the maximum value of the normalized data, Indicates the minimum value of the normalized data.

3. A gyroscope group fault prediction method based on IHBA-RF according to claim 1: said step S4 uses an improved honey badger optimization algorithm IHBA to perform parameter optimization on the random forest algorithm RF, specifically: judging whether the random parameter r is less than 0.5, and using a segmented optimal neighborhood strategy to update the position of the honey badger;

If the random parameter r < 0.5 , it is the digging stage, and the location information of the honey badger is updated according to formula (2-9); if the random parameter r ≥ 0.5 , it is the honey-seeking stage, and the location information of the honey badger is updated according to formula (2-10);

Determine whether the maximum number of iterations has been reached. If so, output the individual position and fitness value of the optimal group; if not , re-execute S3 and update the individual position of the honey badger again.