IARO-GBDT-based observation and aiming system fault prediction method

Abstract

The invention relates to the technical field of artificial intelligence fault diagnosis, and discloses an observation and aiming system fault prediction method based on IARO-GBDT, and the method comprises the steps: carrying out the preprocessing of collected data through an LPP algorithm, and screening the high-evaluation pin signal data as a model input data set; the artificial rabbit optimization algorithm ARO is improved, elite reverse learning is introduced in the ARO population initialization stage, and the fitness is calculated; an adaptive weight factor is introduced in an ARO circuitous foraging stage; t-distribution disturbance is introduced in the ARO random hiding stage to enhance the global search capability; and carrying out parameter optimization on the gradient boosting decision tree GBDT by adopting an improved artificial rabbit optimization algorithm IARO, constructing an IARO-GBDT fault diagnosis model, carrying out fault diagnosis on the observation and aiming system, and outputting a prediction result. According to the method, main parameters of the gradient boosting tree GBDT are optimized through the improved artificial rabbit optimization algorithm (IARO), the defect of blindness of parameter selection in the training process is overcome, and the prediction precision of the classification prediction model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence fault diagnosis, and in particular to a method for predicting faults of an observation and aiming system based on IARO-GBDT. Background Art

[0002] The sighting and aiming system is an important component of the fire control system. It consists of gyroscope sensors, signal processing components, execution components, operation components, and auxiliary components. In the fire control system, the sighting and aiming system is used to accurately control the identification and tracking of targets, so as to improve the efficiency and accuracy of aiming, shorten the target acquisition time, and enhance the hit rate of artillery while on the move. The sighting and aiming box plays a key role in battlefield operations. Frequent use also increases the failure rate. Failure prediction can improve battlefield operational mobility and firepower output, and maximize combat capabilities.

[0003] Currently, various AI-based algorithms have been widely applied in the field of fault diagnosis. However, the arbitrariness of kernel function selection and the limitations of large-scale training in Support Vector Machines lead to a lack of accuracy in prediction results. Expert systems also suffer from difficulties in knowledge acquisition and limitations in knowledge base storage, which cannot guarantee the efficiency and correctness of prediction results. Fault tree analysis is computationally intensive, cumbersome, and consumes excessive memory, which also slows down the operation speed in fault prediction. Compared with other algorithms, Gradient Boosting Decision Tree (GBDT) has the advantages of interpretability and accuracy, perfectly solving the problem of complex internal mechanisms. Gradient Boosting Decision Tree has strong generalization and expressive capabilities, which also makes the fault prediction results more accurate. Through ensemble learning, it has improved the limitations of other machine learning algorithms in the field of fault diagnosis to a certain extent, such as weak generalization ability. Summary of the Invention

[0004] In view of the shortcomings of the fault analysis algorithms in the prior art, such as large calculation scale, high complexity, tedious work, and lack of accuracy in prediction results, the technical solution adopted by the present invention is: a method for predicting faults of sighting and aiming systems based on IARO-GBDT, characterized by comprising the following steps: S1. Acquire data from the pin signals of the observation and aiming system; S2. Preprocess the collected data using the LPP algorithm and select high-rated pin signal data as the model input data set; Divide the model input dataset into a test dataset and a training dataset; S3. Improvements to the Artificial Rabbit Optimization (ARO) algorithm, including the introduction of elite reverse learning to calculate fitness during the ARO population initialization phase; the introduction of adaptive weighting factors during the ARO foraging phase; and the introduction of t-distribution perturbations during the ARO random hiding phase to enhance global search capabilities. S4. The improved Artificial Rabbit Optimization (IARO) algorithm is used to optimize the parameters of the Gradient Boosting Decision Tree (GBDT) and construct the IARO-GBDT fault diagnosis model, which is then trained on the training dataset. S5. Input the test dataset into the IARO-GBDT prediction model to perform fault diagnosis on the observation system and output the prediction results.

[0005] In step S2, the LPP algorithm is used to reduce the dimensionality of the collected raw data, transforming the original high-dimensional dataset... Dimensionality reduction is performed to extract features and obtain a low-dimensional dataset. ,in is a sample point in a high-dimensional dataset, is a sample point in the low-dimensional dataset, K is the number of data point features, N is the total number of data points, k To reduce the dimension of low-dimensional data samples, the specific steps include: S201, calculate each data point The correlation distances with other data points are sorted in ascending order, and the correlation coefficient is calculated using the following formula: , (1-1) , (1-2) , (1-3) Where: is the relevant distance; is the relevant distance parameter; is the average value of the data set; For the i data points, For the j data points; S202, S202, calculation data concentration and Rank-order distance between : , (1-4) Where: For The nth nearest neighbor data point in the relevant distance nearest neighbor list, for exist The position in the list of neighbors of the relevant distance, i.e. for The number of nearest neighbors of ; The final correlation Rank-order distance is: , (1-5) Where: is the final correlation Rank-order distance, For data points Its critical point The asymmetric distance, for exist The position and sum of the neighbor list of related distances exist The minimum value of the position in the list of related distance neighbors; S203. Calculate the distance between each two data points based on the correlation rank-order distance, determine the neighbors of each data point, and construct an adjacency graph. Determine the weight of each edge of the adjacency graph based on the heat kernel function. : , (1-6) S204. Solve the projection matrix 𝑃 to obtain low-dimensional data: , (1-7) Where P is the projection matrix; is the transposed matrix of the projection matrix; D is a diagonal matrix, that is ; Mapped by Laplace; X is the initial position matrix; for X The transposed matrix of , then Equation (1-7) transforms the optimal solution problem into a generalized eigenvalue problem according to the Lagrange multiplier method: in for The minimum eigenvalue of , find the projection matrix composed of the eigenvectors of the first h minimum eigenvalues P ,but ,Depend on Get low-dimensional projection data.

[0006] Step S3 includes the following steps: S301, introduce reference elite reverse learning in the ARO population initialization stage to calculate fitness; The initialization position is described as: , (2-1) Where: For the i The initial position of the rabbit; and are the upper and lower limits of the search range respectively; r is a random number in the range [0,1]; n is the size of the rabbit population; The formula for introducing elite reverse learning is as follows: , (2-2) Where, To learn the position of individuals in reverse; According to the fitness value, the better solution is selected as the initial individual to form the initial population, and the fitness function is selected Therefore, when the fitness value of the reverse learning individual is less than the fitness value of the original individual, the position of the reverse learning individual is retained. Otherwise, the original position of the individual is retained. , the expression of the artificial rabbit position retained after combining the initial solution and the reverse solution is: , (2-3) Where, Initialize the position after improvement; is the fitness value of the reverse learning individual; is the fitness value of the original individual; S302. Introducing an adaptive weighting factor during the ARO foraging phase: Rabbits usually do not forage near their own nests. They forage near other rabbit nests to prevent their nests from being discovered by predators and increase their survival probability. The position is updated as follows: , (2-4) Where: For the i Rabbit t +1 iteration position; For the j Rabbit t Iteration position; R is the circuitous foraging operator, which is used to simulate the circuitous foraging characteristics of rabbits; For the i Rabbit t Iteration position; is a random number between 0 and 1; To obey the standard normal distribution; At this time, the adaptive weight factor is introduced , which improves the quality of individual positions, enables other individuals to converge to the optimal position faster, and speeds up the convergence. Based on the characteristics of artificial rabbits, the formula for adaptive weights is as follows: , (2-5) Where: is the adaptive weight factor; t represents the maximum number of iterations; t represents the current number of iterations. The improved rabbit position is updated as follows: , (2-6) Where: For the improved i The rabbit in t +1 iteration position; To improve the previous i Rabbit t +1 iteration position; S303. Introducing t-distribution perturbation in the ARO random hiding phase to enhance global search capability: To escape predators, rabbits usually dig several different burrows around their nests for hiding. In each iteration, the rabbit always generates d burrows along each dimension of the search space; the burrow generation is as follows: , (2-7) Where: For the i Rabbit j cave locations; H To hide the parameters, during the iteration H It decreases linearly from 1 to 1 / T, where g is an integer of 0 or 1; Rabbits are often hunted and attacked by predators. In order to survive, the rabbit randomly chooses a hole from all the holes to hide in to avoid being preyed on. The updated position is: , (2-8) Where: For the i The rabbit in t +1 iteration position; for t The cave location is randomly selected in the iteration; A random number between 0 and 1; At this point, by introducing a t-distribution perturbation strategy and using the t-distribution mutation operator to perturb the individual's position, the degree-of-freedom parameters of the t-distribution mutation operator will change with the number of iterations, thereby guiding the artificial rabbit to better evolve towards the optimal position and effectively improving the convergence speed of the algorithm. Therefore, the rabbit's position is now updated as follows: (2-9) Where, The final position of the rabbit after updating, is a random number generated by the t-distribution in the cth dimension, Let c be the optimal position of the rabbit in dimension c, where c is the spatial dimension.

[0007] In the ARO algorithm, the rabbit always tends to frequently forage in the early stages of the iteration and frequently hide randomly in the later stages of the iteration. This search mechanism depends on the rabbit's energy, which gradually shrinks over time. An energy factor is set to simulate the automatic transition from the exploration phase to the exploitation phase. The energy factor is defined as follows: (2-10) Where: Let be the energy factor for the t-th iteration; is a random number between 0 and 1; t represents the maximum number of iterations; t represents the current number of iterations.

[0008] Step S4 employs an improved Artificial Rabbit Optimization (IARO) algorithm to optimize the parameters of the Gradient Boosting Decision Tree (GBDT), specifically by determining the energy factor. Determine the individual's current stage: If energy factor If the value is greater than 1, it indicates a roundabout foraging phase. An adaptive weighting factor is introduced, and the location information of the individual rabbit is updated according to formula (2-6). If the random parameter... If ≤1, it is the random hiding stage. The t-distribution strategy is used to perturb and select a cave as the hiding place. The position information of the individual rabbits is updated according to formula (2-9). Determine if the maximum number of iterations has been reached. If it has, output the position and fitness value of the individual in the optimal population. If not, re-execute S3 and update the position of the individual rabbits again.

[0009] Compared with the prior art, the present invention has the following beneficial technical effects and advantages: 1. Compared with other comprehensive evaluation algorithms (hierarchical analysis method, grey relational analysis, etc.), the LPP algorithm is more suitable for manifold learning, especially for high-dimensional data. This method can find the optimal linear approximation of the characteristic function of the Laplace-Beltrami operation on the manifold to obtain the optimal projection direction, which can better meet the needs of dimensionality reduction and reduce dependence on experience, thereby improving the reliability and effect of the algorithm. 2. Compared to Support Vector Machines (SVM) and decision trees, Gradient Boosting Trees (GBDT) exhibits significant advantages in various tasks. First, it boasts high prediction accuracy, making it particularly suitable for complex datasets. Second, GBDT naturally handles missing values ​​without additional imputation, reducing the complexity of data preprocessing. Furthermore, it provides feature importance assessment, aiding in understanding model decisions. Finally, GBDT's flexibility lies in its support for multiple loss functions, enabling it to adapt to different types of prediction tasks and greatly expanding its application scope. 3. The main parameters of the gradient boosting tree (GBDT) are optimized by using the improved artificial rabbit optimization algorithm (IARO), which makes up for the blindness of parameter selection during the training process and improves the prediction accuracy of the classification prediction model. Compared with the decision tree, support vector machine, and GBDT algorithm experiments, IARO-GBDT has higher prediction accuracy and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a schematic diagram of the main framework of the observation and aiming system fault prediction method based on IARO-GBDT of the present invention. Figure 2 This is the flow chart of the LPP algorithm of the present invention. DETAILED DESCRIPTION

[0011] The present invention will now be described in detail with reference to the accompanying drawings, but the scope of protection of the present invention is not limited by the drawings.

[0012] This invention provides a fault prediction method for observation and aiming systems based on IARO-GBDT, such as... Figure 1 As shown: The following steps are included: S1. Collect data from the pin signals of the observation and aiming system. The research object is a certain type of observation and aiming system. The data values ​​of the pin signals of the power module of the observation and aiming system are collected through the equipment test bench as the initial data for the experiment.

[0013] S2. Preprocess the collected data using the LPP algorithm and select high-rated pin signal data as the model input data set; Furthermore, the LPP algorithm is used to reduce the dimensionality of the collected raw data, transforming the original high-dimensional dataset... Dimensionality reduction is performed to extract features and obtain a low-dimensional dataset. ,in is a sample point in a high-dimensional dataset, is a sample point in the low-dimensional dataset, K is the number of data point features, N is the total number of data points, k To reduce the dimension of low-dimensional data samples, the specific steps include: S201, calculate each data point The correlation distances with other data points are sorted in ascending order, and the correlation coefficient is calculated using the following formula: , (1-1) , (1-2) , (1-3) Where: is the relevant distance; is the relevant distance parameter; is the average value of the data set; For the i data points, For the j data points; S202, calculation data concentration and Rank-order distance between : , (1-4) Where: For The nth nearest neighbor data point in the relevant distance nearest neighbor list, for exist The position in the list of neighbors of the relevant distance, i.e. for The number of nearest neighbors of ; The final correlation Rank-order distance is: , (1-5) Where: is the final correlation Rank-order distance, For data points Its critical point The asymmetric distance is the asymmetric distance between a data point and its critical point. for exist The position and sum of the neighbor list of related distances exist The minimum value of the position in the list of related distance neighbors;

[0014] S203. Calculate the distance between each two data points based on the correlation rank-order distance, determine the neighbors of each data point, and construct an adjacency graph. Determine the weight of each edge of the adjacency graph based on the heat kernel function. : , (1-6) S204. Solve the projection matrix \(P\) to obtain the low-dimensional data: , (1-7) where \(P\) is the projection matrix; is the transpose matrix of the projection matrix; D is a diagonal matrix, that is ; is the Laplacian eigenmap; X is the initial position matrix; is X the transpose matrix of; Introduce the constraint condition , then Equation (1-7) transforms the optimal solution problem into a generalized eigenvalue problem according to the Lagrange multiplier method: where is the smallest eigenvalue of, and the eigenvectors corresponding to the first \(h\) smallest eigenvalues form the projection matrix P , then , from the low-dimensional projection data is obtained.

[0015] As Figure 2 shown, the specific process of S2 data processing is as follows: Input: Original data set , target dimension \(k \ll K\), initial neighborhood parameter \(k\).

[0016] Output: \(k\)-dimensional vector \(Y\) after dimensionality reduction of data \(X\).

[0017] (1) Calculate the correlation distance between each data point and other data points according to Equations (1-1), (1-2) and (1-3), and sort them in ascending order; (2) Calculate the symmetric and normalized correlation Rank-order distance according to Equations (1-4) and (1-5); (3) Select the nearest neighbor points for each data point according to the correlation Rank-order distance obtained in the previous step, and construct a neighborhood graph; (4) Assign weights to the neighborhood graph according to Equation (1-6); (5) Solve the projection matrix \(P\) according to Equation (1-7), and calculate the low-dimensional projection data by Equation .

[0018] Improve the artificial rabbit optimization algorithm ARO, including introducing elite opposition-based learning in the ARO population initialization stage to calculate the fitness; introducing an adaptive weight factor in the ARO meandering foraging stage; introducing a t-distribution perturbation in the ARO random hiding stage to enhance the global search ability; specifically: S301, introduce reference elite reverse learning in the ARO population initialization stage to calculate fitness; The initialization position is described as: , (2-1) Where: For the i The initial position of the rabbit; and are the upper and lower limits of the search range respectively; r is a random number in the range [0,1]; n is the size of the rabbit population; This random population initialization method results in an uneven population distribution, affecting subsequent iterative optimization. Elite individuals in the artificial rabbit population contain more critical information than other individuals. By constructing reverse solutions for these elite individuals during the initialization phase and leveraging information from high-fitness individuals, the population is guided toward a more optimal solution. This significantly improves the population's initial fitness, promotes algorithm convergence, and reduces early random search time.

[0019] The formula for introducing elite reverse learning is as follows: , (2-2) Where, To learn the position of individuals in reverse; According to the fitness value, the better solution is selected as the initial individual to form the initial population, and the fitness function is selected Therefore, when the fitness value of the reverse learning individual is less than the fitness value of the original individual, the position of the reverse learning individual is retained. Otherwise, the original position of the individual is retained. , the expression of the artificial rabbit position retained after combining the initial solution and the reverse solution is: , (2-3) Where, Initialize the position after improvement; is the fitness value of the reverse learning individual; This represents the original fitness value of the individual. S302. Introducing an adaptive weighting factor during the ARO foraging phase: Rabbits usually do not forage near their own nests. They forage near other rabbit nests to prevent their nests from being discovered by predators and increase their survival probability. The position is updated as follows: , (2-4) Where: For the i Rabbit t +1 iteration position; For the j Rabbitt Iteration position; R is the circuitous foraging operator, which is used to simulate the circuitous foraging characteristics of rabbits; For the i Rabbit t Iteration position; is a random number between 0 and 1; To obey the standard normal distribution; At this point, the algorithm has the problem of being easily trapped in local optimality and poor convergence speed, so the adaptive weight factor is introduced. The weight changes with the number of iterations, effectively reducing the possibility of getting stuck in local optima and thus enhancing the algorithm's search capability. Furthermore, the introduction of adaptive weights improves the quality of individual positions, enabling other individuals to converge to the optimal position more quickly, accelerating the convergence speed. Based on the characteristics of the artificial rabbit, the formula for the adaptive weights is as follows: , (2-5) Where: is the adaptive weight factor; t represents the maximum number of iterations; t represents the current number of iterations. The improved rabbit position is updated as follows: , (2-6) Where: For the improved i The rabbit in t +1 iteration position; To improve the previous i Rabbit t +1 iteration position;

[0020] S303. Introducing t-distribution perturbation in the ARO random hiding phase to enhance global search capability: To escape predators, rabbits usually dig several different burrows around their nests for hiding. In each iteration, the rabbit always generates d burrows along each dimension of the search space; the burrow generation is as follows: , (2-7) Where: For the i Rabbit j cave locations; H To hide the parameters, during the iteration H It decreases linearly from 1 to 1 / T, where g is an integer of 0 or 1; Rabbits are often hunted and attacked by predators. In order to survive, the rabbit randomly chooses a hole from all the holes to hide in to avoid being preyed on. The updated position is: , (2-8) Where: For the i The rabbit in t +1 iteration position; for t The cave location is randomly selected in the iteration; A random number between 0 and 1; At this point, by introducing a t-distribution perturbation strategy and using the t-distribution mutation operator to perturb the individual's position, the degree-of-freedom parameters of the t-distribution mutation operator will change with the number of iterations, thereby guiding the artificial rabbit to better evolve towards the optimal position and effectively improving the convergence speed of the algorithm. Therefore, the rabbit's position is now updated as follows: (2-9) Where, The final position of the rabbit after updating, For random numbers generated in the c-th dimension of the t-distribution, Let c be the optimal position of the rabbit in dimension c, where c is the spatial dimension.

[0021] In the ARO algorithm, the rabbit always tends to frequently forage in the early stages of the iteration and frequently hide randomly in the later stages of the iteration. This search mechanism depends on the rabbit's energy, which gradually shrinks over time. An energy factor is set to simulate the automatic transition from the exploration phase to the exploitation phase. The energy factor is defined as follows: (2-10) Where: Let be the energy factor for the t-th iteration; is a random number between 0 and 1; t represents the maximum number of iterations; t represents the current number of iterations.

[0022] S4. Use the improved artificial rabbit optimization algorithm IARO to optimize the parameters of the gradient boosting decision tree GBDT, build the IARO-GBDT fault diagnosis model, and train the training data set; specifically: Constructing a gradient boosted decision tree (GBDT) first sets the input sample set , maximum number of iterations (number of base learners), loss function Output regression tree model .

[0023] Step 1: Initialization ; ; Let the loss function be squared loss: .

[0024] Step 2: Calculate the residuals of the loss function in the current model. The value of ; The residual obtained in a) Taking the new true value of the sample, we obtain the m-th regression tree, and the corresponding leaf node region is... Indicates the number of leaf nodes.

[0025] right The leaf nodes calculate the best fit value: ; yes Minimum square loss; Update the regression tree model ; .

[0026] Step 3: Get the regression tree model: .

[0027] Step 1 is the initialization process, which calculates the minimum value of the loss function and uses it as the root node; Step 2 a) calculates the residual estimate through the negative gradient of the loss function in the current model; Step 2 b) fits the residual approximation to establish the mth regression tree; Step 2 c) calculates the loss function to reach the minimum value; Step 2 d) updates the regression tree model; The third step is to obtain the final regression tree model.

[0028] Improved algorithm optimization parameters: The parameters involved in GBDT algorithm modeling are mainly the learning rate (learning_rate), which is used to control the step size of parameter updates during learning. If the step size is too large, the learning process may diverge. Conversely, it will cause the model to iterate too many times, and the learning time will increase significantly. The maximum number of iterations (n_estimators) represents the number of base learners and interacts with the learning_rate. When the learning_rate is small, the number of iterations needs to be increased to converge the training error. Subsampling (subsample) is used to control the proportion of data set samples involved in fitting. When it is set to less than 1, it can effectively reduce the variance of the overall model and prevent overfitting. The maximum depth of the decision tree (max_depth), the minimum number of samples required for internal node re-division (min_samples_split) and the minimum number of samples contained in the leaf node (min_samples_leaf) are all used to control the complexity of each tree. The specific value depends on the data distribution. If the value is too large, the model structure will be complicated and prone to overfitting. Conversely, it will easily lead to underfitting. Step size factor If the loss function is excessively minimized, a step size factor can be added to prevent overfitting. Step size factor The value range is , under the same learning effect, if A smaller value might require more iterations, while an excessively large value would skip the optimal point.

[0029] Determine the energy factor Determine the individual's current stage: If energy factor >1, it is the roundabout foraging stage, by introducing the adaptive weight factor, the position information of the rabbit individual is updated according to formula (2-6); if the random parameter If ≤1, it is the random hiding stage. The t-distribution strategy is used to perturb and select a cave as the hiding place. The position information of the individual rabbits is updated according to formula (2-9). Determine if the maximum number of iterations has been reached. If it has, output the position and fitness value of the individual in the optimal population. If not, re-execute S3 and update the position of the individual rabbits again.

[0030] S5. Input the test data set into the IARO-GBDT prediction model, perform fault diagnosis on the sighting system, and output the prediction results, which can effectively predict the fault of the pin signal of the power module of the sighting system.

[0031] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting faults in an observation and aiming system based on IARO-GBDT, characterized in that: The following steps are involved: S1, collect data of the pin signal of the sighting system; S2. Preprocess the collected data using the LPP algorithm and select the pin signal data with high evaluation as the model input data set; Divide the model input dataset into a test dataset and a training dataset; S3. Improvements to the artificial rabbit optimization algorithm ARO, including the introduction of elite reverse learning to calculate fitness in the ARO population initialization phase; the introduction of adaptive weight factors in the ARO foraging phase; and the introduction of t-distribution perturbations in the ARO random hiding phase to enhance global search capabilities. S4. Use the improved artificial rabbit optimization algorithm IARO to optimize the parameters of the gradient boosting decision tree GBDT, build the IARO-GBDT fault diagnosis model, and train the training data set; S5. Input the test data set into the IARO-GBDT prediction model, perform fault diagnosis on the sighting system, and output the prediction results.

2. The method for predicting a fault of an observation and aiming system based on IARO-GBDT according to claim 1, characterized in that: In step S2, the LPP algorithm is used to reduce the dimension of the collected original data, and the original high-dimensional data set is Perform dimensionality reduction to extract features and obtain a low-dimensional data set ,in is a sample point in a high-dimensional data set, is a sample point in the low-dimensional data set, K is the number of data point features, N is the total number of data points, k is the dimension of low-dimensional data samples; specifically includes the following steps: S201, calculate each data point The correlation distance with other data points is sorted in order of size. The correlation coefficient calculation formula is: ，(1-1) ，(1-2) ，(1-3) Where: is the relevant distance; is the relevant distance parameter; is the average value of the data set; For the i data points, For the j data points; S202, calculation data concentration and Rank-order distance between : ，(1-4) Where: For The 𝑚th nearest neighbor data point in the nearest neighbor list of the relevant distance, for exist The position in the list of neighbors of the relevant distance, i.e. for The number of nearest neighbors of ; The final correlation Rank-order distance is: ，(1-5) Where: is the final correlation Rank-order distance, For data points Its critical point The asymmetric distance for exist The position and sum of the neighbor list of related distances exist The minimum value among the positions in the list of neighboring points of the relevant distance; S203, calculate the distance between every two data points based on the correlation Rank-order distance, determine the neighbors of each data point, and construct an adjacency graph, and determine the weight of each edge of the adjacency graph based on the heat kernel function : ，(1-6) S204. Solve the projection matrix 𝑃 to obtain low-dimensional data: ，(1-7) Where P is the projection matrix; is the transposed matrix of the projection matrix; D is a diagonal matrix, that is ; For Laplace's reflection; X is the initial position matrix; for X The transposed matrix of , then formula (1-7) transforms the optimal solution problem into the problem of finding generalized eigenvalues ​​according to the Lagrange multiplier method: in for The minimum eigenvalue of , find the projection matrix composed of the eigenvectors of the first h minimum eigenvalues P ,but ,Depend on Get low-dimensional projection data.

3. The method for predicting faults of an observation and aiming system based on IARO-GBDT according to claim 1, characterized in that: Step S3 includes the following steps: S301, introduce reference elite reverse learning in the ARO population initialization stage to calculate fitness; The initialization position is described as: ，(2-1) Where: For the i The initial position of the rabbit; and are the upper and lower limits of the search range respectively; r is a random number in the range [0,1]; n is the size of the rabbit population; The formula for introducing elite reverse learning is as follows: ，(2-2) In the formula, To reversely learn the position of individuals; According to the fitness value, a better solution is selected as the initial individual to form the initial population, and the fitness function is selected Therefore, when the fitness value of the reverse learning individual is less than the fitness value of the original individual, the position of the reverse learning individual is retained. Otherwise, the original position of the individual is retained. , the expression of the artificial rabbit position retained after combining the initial solution and the reverse solution is: ，(2-3) In the formula, Initialize the position for improvement; is the fitness value of the reverse learning individual; is the fitness value of the original individual; S302, introducing an adaptive weight factor in the ARO roundabout foraging stage: Rabbits usually do not forage near their own nests, they always forage near other rabbit nests to prevent rabbit nests from being discovered by predators and increase the probability of survival. The position is updated as follows: ，(2-4) Where: For the i Rabbit t +1 iteration position; For the j Rabbit t Iteration position; R is the circuitous foraging operator, which is used to simulate the circuitous foraging characteristics of rabbits; For the i Rabbit t Iteration position; is a random number between 0 and 1; It follows the standard normal distribution; At this time, the adaptive weight factor is introduced , which improves the quality of individual positions, enables other individuals to converge to the optimal position faster, and speeds up the convergence. Based on the characteristics of artificial rabbits, the formula for adaptive weights is as follows: ，(2-5) Where: is the adaptive weight factor; is the maximum number of iterations; t is the current number of iterations; The improved bunny position is updated as follows: ，(2-6) Where: For the improved i The rabbit is in t +1 iteration position; To improve the previous i Rabbit t +1 iteration position; S303. Introducing t-distribution perturbation in the ARO random hiding stage to enhance global search capability: To escape predators, rabbits usually dig several different caves around their nests to hide. In each iteration, rabbits always generate d caves along each dimension of the search space; the caves are generated as: ，(2-7) Where: For the i Rabbit j cave locations; H To hide the parameters, during the iteration H It decreases linearly from 1 to 1 / T, where g is an integer of 0 or 1; Rabbits are often hunted and attacked by predators. In order to survive, the rabbit randomly chooses a hole from all the holes to hide in order to avoid being preyed on. The position is updated as follows: ，(2-8) Where: For the i The rabbit is in t +1 iteration position; for t The cave location is randomly selected in the iteration; A random number between 0 and 1; At this time, by introducing the t-distribution perturbation strategy, the t-distribution mutation operator is used to perturb the individual position. The degree of freedom parameter of the t-distribution mutation operator will change with the number of iterations, thereby guiding the artificial rabbit to evolve to the optimal position better and effectively improving the convergence speed of the algorithm. At this time, the position of the rabbit is updated as follows: ，(2-9) In the formula, The final position of the rabbit after update. is a random number generated by t-distribution in the cth dimension, is the optimal position of the rabbit in c dimension, where c is the spatial dimension.

4. The method for predicting faults of an observation and aiming system based on IARO-GBDT according to claim 3 is characterized in that: In the ARO algorithm, the rabbit always tends to frequently forage in the initial stage of the iteration, and frequently hides randomly in the later stage of the iteration. This search mechanism depends on the energy of the rabbit, which gradually shrinks over time. An energy factor is set to simulate the automatic transition from the exploration stage to the development stage. The definition of the energy factor is as follows: ，(2-10) Where: is the energy factor of the tth iteration; is a random number between 0 and 1; is the maximum number of iterations; t is the current number of iterations.

5. The method for predicting faults of an observation and aiming system based on IARO-GBDT according to claim 4 is characterized in that: The step S4 uses the improved artificial rabbit optimization algorithm IARO to optimize the parameters of the gradient boosting decision tree GBDT, specifically: determine the energy factor Determine the current stage of the individual: If the energy factor >1, it is the roundabout foraging stage. By introducing the adaptive weight factor, the position information of the rabbit individual is updated according to formula (2-6); if the random parameter ≤1, it is the random hiding stage, using the t-distribution strategy to perturb and select a cave as hiding, and update the position information of the rabbit individual according to formula (2-9); Determine whether the maximum number of iterations has been reached. If so, output the individual positions and fitness values ​​of the optimal group. If not, re-execute S3 and update the individual positions of the rabbits again.

Images