A cross-connect cable fault diagnosis method and system based on improved Bayesian optimization XGBoost

Abstract

The present invention proposes a cross-connected cable fault diagnosis method and system based on improved Bayesian optimization XGBoost, aiming to solve the problems of insufficient accuracy and low efficiency of existing cable fault detection methods in complex environments. First, the Bayesian optimization algorithm is improved. For the Gaussian process in the prior function, the Marton kernel function and the square exponential kernel function are multiplied and compounded as the kernel function. The improved Markov chain Monte Carlo method is used to optimize the parameters of the kernel function. Secondly, for the acquisition function, the adjustment parameters in the traditional confidence boundary strategy are improved. Finally, the improved Bayesian optimization algorithm is used to intelligently search and optimize the hyperparameters of XGBoost, which effectively improves the prediction performance and generalization ability of the model, and realizes efficient and accurate diagnosis of cross-connected cable faults. The implementation of the present invention will help to improve the automation level of cable fault processing, reduce operation and maintenance costs, and has important practical application value and promotion prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault diagnosis, and in particular to a cross-connection cable fault diagnosis method and system based on improved Bayesian optimization XGBoost. Background Art

[0002] In modern power systems, cross-connect cables are a crucial component of power transmission, and their reliability is directly linked to the safe and stable operation of the power grid. However, due to the long-term operation of cables in complex and changing environments, they are subject to various external factors such as soil moisture, temperature fluctuations, mechanical damage, and aging, leading to frequent cable failures. Rapidly and accurately diagnosing and locating cable faults is crucial for reducing power outages, ensuring power supply reliability, and lowering maintenance costs.

[0003] Traditional cable fault detection methods, such as the bridge method and low-voltage pulse reflection method, can detect faults to a certain extent. However, they often suffer from low detection accuracy and complex operations when faced with complex and changing fault types. In recent years, XGBoost, as an efficient gradient boosting algorithm, has performed well in data classification and regression prediction and has been widely used in various fault diagnosis fields.

[0004] However, XGBoost has numerous hyperparameters, and efficiently determining the optimal parameter combination is key to improving model predictive performance. Bayesian optimization, as an advanced optimization method, has become an effective means of optimizing complex machine learning models by constructing a probabilistic model of the objective function and efficiently searching for the optimal parameter combination based on this model.

[0005] For example, patent application CN117434388A discloses a method for identifying fault patterns in cross-connected high-voltage cable sheath grounding faults. This method selects the factor with the greatest impact on the fault as a characteristic indicator of the input sample and uses a Bayesian approach to determine hyperparameters. This method adaptively adjusts the hyperparameters during network training to achieve optimal values, thereby improving the generalization capability of the neural network. The patent mentions the use of high-voltage pulse emission and high-energy audio emission as detection units. However, different cable faults may require signals of different frequencies or intensities for detection, making it difficult to accurately detect certain specific faults. Under varying environmental conditions (such as extreme temperature, humidity, and soil conditions), the performance of the detection unit may be affected, resulting in reduced accuracy in detection results. The threshold setting may need to be adjusted based on actual conditions to ensure accurate and reliable fault identification. However, the selection of the threshold may be affected by various factors, such as cable type and operating conditions, and therefore may involve certain difficulties and uncertainties. As cable operating conditions change (such as aging and environmental factors), the threshold may need to be regularly updated and adjusted to ensure continued effectiveness in fault identification.

[0006] Patent publication number CN117630562A discloses a superconducting cable fault detection method based on the combined characteristics of current amplitude and phase. This method uses a machine learning algorithm to construct a classifier model and trains the classifier based on data from a sample set based on the combined characteristics of current amplitude and phase, resulting in a mature classifier model. Multiple test points are selected for the superconducting cable under test, and testing is performed at each test point to construct a classifier based on the combined characteristics of current amplitude and phase. The combined characteristics of each test point are input into the mature classifier model, and fault detection of the superconducting cable under test is performed based on the output of the classifier model. This method, based on the combined characteristics of current amplitude and phase, offers improved accuracy for quench detection and can pinpoint the location of the quench fault when a superconducting cable quenches. However, this method primarily focuses on quench detection based on current amplitude and phase, and may not cover all types of superconducting cable faults, such as mechanical damage or insulation aging. The operating environment of superconducting cables is complex and variable, with factors such as temperature, humidity, and electromagnetic interference potentially affecting detection results. While this patent uses the combined characteristics of current amplitude and phase for detection, it does not explicitly address how to eliminate or reduce the impact of these environmental factors. Therefore, the fault detection categories and accuracy need to be improved. Summary of the Invention

[0007] Purpose of the invention: In view of the limitations of traditional fault detection methods in the face of complex fault types and the insufficient accuracy and low efficiency of existing cable fault detection methods in complex environments, the present invention provides a cross-connected cable fault diagnosis method based on improved Bayesian optimization XGBoost. The present invention also discloses a cross-connected cable fault diagnosis system based on improved Bayesian optimization XGBoost.

[0008] Technical solution: The present invention provides a cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost, the method comprising:

[0009] The collected circulation samples are used to construct a multidimensional feature vector I. Circulation sample data is constructed based on multiple multidimensional feature vectors I. The circulation sample data is divided into a training set and a test set according to a certain ratio. The circulation samples are obtained from the monitoring background by installing an online monitoring device for sheath circulation at the cable line circulation monitoring point. Then, all the measured circulation amplitudes are used to construct a multidimensional feature vector I.

[0010] Setting the range of XGBoost model hyperparameters, and pre-training the XGBoost model using an improved Bayesian optimization algorithm based on the training set to find the optimal hyperparameter combination;

[0011] The improved Bayesian optimization algorithm includes optimizing the parameters in the prior function and optimizing the acquisition function, wherein optimizing the parameters in the prior function includes: selecting a Gaussian process to model the objective function, the function obeying the distribution of the objective function includes a mean function and a kernel function, the kernel function is a composite kernel function obtained by multiplying two single kernel functions, a Marten kernel function and a square exponential kernel function, and optimizing one or more parameters corresponding to the composite kernel function using an improved Markov chain Monte Carlo method based on flower pollination, calculating the distribution of the parameters to be optimized, and then calculating the corresponding expected values, thereby obtaining the optimized parameter values; the optimization of the acquisition function includes: optimizing and improving the adjustment parameters therein using an adaptive confidence boundary strategy;

[0012] The optimal hyperparameter combination is used as the parameters in the final model, and the training set is reused to train the XGBoost model;

[0013] The trained XGBoost model is used to predict the fault type of the test set, output the prediction results and analyze the diagnostic effect.

[0014] Further, including:

[0015] The Gaussian Process (GP) is selected to model the objective function. The functions that obey the distribution of the objective function include a mean function and a kernel function, specifically including:

[0016] The Gaussian process GP is used to model the objective function. GP is regarded as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function that obeys its distribution consists of a mean function and a kernel function, namely:

[0017] f(x)~GP(m(x),k(x,x′));

[0018] Where, the mean function m(x) = E[f(x)], and the kernel function k(x,x′) = E[(f(x)-m(x))(f(x′)-m(x′))] are used to indicate the similarity between two data sets. For the hyperparameter optimization process, x is a series of hyperparameter combinations, and f(x) is the final evaluation index score. Therefore, without loss of generality, let m(x) = 0.

[0019] If the output value on the training set is f=(f1,f2,...,f n ) T , the predicted function value f(x * )=f * , then the following joint distribution exists in the properties of the Gaussian process:

[0020]

[0021] Where K is the Gram matrix k(X,X), k * ={k(x * ,x1),k(x * ,x2),...,k(x * ,x n )} T ,

[0022] But the objective function f(x) and the observed value y=(y1,y2,...,y n ) T There is usually an error between them, which is represented by noise∈, that is:

[0023] y=f(x)+∈;

[0024] And assume that the noise ∈ i independent and identically distributed, and Then the prior of the kernel function becomes, the prior of the kernel function, that is, after considering the noise, the prior of the kernel function needs to be adjusted to include the influence of noise:

[0025]

[0026] Then we get y and f * The joint distribution of is:

[0027]

[0028] Where I is the identity matrix, and denote the variance of the objective function and the variance of the noise, respectively.

[0029] The variance of the objective function is part of the kernel function in the Gaussian process, which measures the degree of variation of the objective function between different points in the absence of observation noise. This variance together with the kernel function defines the smoothness of the objective function. The value means that the objective function changes more drastically in the input space, while smaller A value of means that the objective function is smoother. In the context of Gaussian processes, Often works together with the kernel's scale parameter to control the smoothness and variability of the function.

[0030] The variance of the noise represents the random error or noise in the observed data and describes the distribution of random errors. In Gaussian process regression, the noise variance is an important hyperparameter that affects how well the model fits the observed data. If the noise variance is set too large, the model may be too smooth and fail to capture the true changes in the objective function. If it is set too small, the model may overfit the noise in the observed data.

[0031] According to the properties of multivariate normal distribution, we get:

[0032] f * |x * ,y~N(μ(x * ),σ 2 (x * ));

[0033] μ(x * ) is the final predicted value, σ 2 (x * ) is the prediction variance,

[0034] The traditional Marton kernel function and the square exponential kernel function are multiplied and compounded to obtain a more accurate fitting result. The Marton-square exponential composite function is:

[0035]

[0036] Where r = xx′ represents the Euclidean distance between sample points, l is the scale parameter, and α is the smoothing parameter.

[0037] The scale parameter controls the smoothness and speed of change of the function in the input space. It affects the rate at which the function value changes with the input points. A smaller scale parameter value means that the function changes more quickly in the input space, that is, the function is more sensitive to small changes in the input points, which may cause the model to overfit the data. A larger scale parameter value means that the function changes more slowly and smoothly, which helps the model generalize to unseen data, but may cause the model to underfit the data.

[0038] Further, including:

[0039] One or more parameters corresponding to the composite kernel function are optimized using an improved Markov chain Monte Carlo method based on flower pollination, including:

[0040] Arbitrarily select the initial parameter θ0 to be optimized and execute the loop i=1,2,...,k times;

[0041] Let the value of the parameter to be optimized for the current number of visits be θ i-1 =θ, randomly select the next candidate parameter value θ′ according to the proposed distribution;

[0042] According to the flower pollination FPA, when the conversion probability between global optimization and local optimization p>0.8, the next candidate parameter θ′ is calculated globally as:

[0043]

[0044] Where, is the sample generated by the mth Markov chain at the i-th iteration, is the optimal sampling point among the currently generated samples, L is the search step length, and its calculation formula is as follows:

[0045]

[0046] In the formula, u=i -λ , Γ is the standard gamma function, λ is usually 1-5, v ~ N(0,1);

[0047] If p≤0.8, the next candidate parameter θ′ is:

[0048]

[0049] Where ξ is a random value uniformly distributed on [0,1], and are the samples generated by the m-th Markov chain at the j-th and k-th iterations, respectively. and are equivalent to random solutions in the model.

[0050] The flower pollination algorithm is an optimization algorithm inspired by the pollination process of flowers in nature. It consists of two main phases: global pollination and local pollination. Global pollination simulates the process by which pollen is spread over a large area by pollinators such as insects. In global pollination, pollen can travel great distances, which facilitates the pollination and reproduction of the fittest individuals. Global pollination is simulated in the algorithm using Lévy flights, a type of random walk with a heavy-tailed distribution. This means that there is a relatively high probability of large strides during the random walk, thus achieving global search. Local pollination simulates the process of self-pollination, in which pollen is spread between physically close flowers. Local pollination is achieved through local random walks, which search within the neighborhood of the current solution. This helps the algorithm conduct a refined search near the found high-quality solution. The flower pollination algorithm introduces a conversion probability p to balance the weights of global and local searches, thereby achieving a balance between global and local searches.

[0051] Using the updated acceptance probability, we can get the posterior samples θ1, θ2..., θ k ;

[0052] Therefore, the expected value of the distribution of the parameters to be optimized is expressed as:

[0053]

[0054] Further, including:

[0055] The acceptance probability of the update is expressed as:

[0056]

[0057] Where p(·) is the given prior distribution and q(·) is the given proposed distribution;

[0058] Randomly select u from uniform distribution, u∈(0,1);

[0059] If u≤α(θ,θ′), then accept the transfer θ i =θ′, otherwise, do not accept the transfer, and θ i =θ i-1 After the loop is executed, we can get the posterior samples θ1, θ2..., θ k .

[0060] Further, including:

[0061] The adaptive confidence boundary strategy is used to optimize and improve the adjustment parameters, specifically including:

[0062] The acquisition function adopts an adaptive confidence boundary strategy to sample in unknown areas or explore near the currently considered global optimal solution, which is expressed as:

[0063] AUCB(x)=μ(x)+εσ(x);

[0064] In the formula, the adaptive adjustment parameter s=|y max +D max | / (y max +D max ), y max Indicates the maximum value in the current fitting curve, D max It represents the maximum value in the current observation set, and s is the coefficient of difference, which is used to further represent y max With D max degree of difference.

[0065] In the context of Bayesian optimization, a fitted curve typically refers to a prediction curve obtained by modeling the objective function using a Gaussian process (GP). This curve represents the predicted behavior of the objective function in the input space, including its mean and uncertainty. The Gaussian process estimates the value of the objective function at unobserved points using existing observed data points, resulting in a smooth curve. This curve is called the fitted curve.

[0066] Among them, the observation set: The observation set refers to the set of parameter points and their corresponding objective function values ​​that have been evaluated during the optimization process. These data points are used to build a Gaussian process model, that is, to learn the behavior of the objective function through these points. During the iterative process of Bayesian optimization, new observation points are continuously added to this set to update and improve the fitting curve. In the adaptive confidence boundary strategy, the maximum value of the fitting curve and the maximum value in the observation set are used to calculate the difference coefficient, which in turn affects the value of the adaptive adjustment parameter. This parameter determines the tendency of the acquisition function between exploration and development. When y max and D max When they are close, the acquisition function tends to explore; when they are far away, the acquisition function tends to exploit.

[0067] Further, including:

[0068] The adaptive adjustment parameters and y max 、D max The relationships include:

[0069] When y max With D max When it approaches infinity, ε tends to 1, and the acquisition function based on the adaptive confidence boundary strategy is more inclined to the exploration state;

[0070] When y max With D max When it is infinitely far away, ε tends to 0, and the acquisition function is more inclined to the development state.

[0071] In a second aspect, the present invention further provides a cross-connect cable fault diagnosis method system based on improved Bayesian optimization XGBoost, the system comprising:

[0072] An acquisition module is used to construct a multidimensional feature vector I using the collected circulation samples, construct circulation sample data based on multiple multidimensional feature vectors I, and divide the circulation sample data into a training set and a test set according to a certain ratio. The circulation samples are obtained from the monitoring background by installing an online monitoring device for sheath circulation at the cable line circulation monitoring point, and then constructing a multidimensional feature vector I using all the measured circulation amplitudes;

[0073] A model pre-training module is used to set the range of XGBoost model hyperparameters and pre-train the XGBoost model based on the training set using an improved Bayesian optimization algorithm to find the optimal hyperparameter combination;

[0074] The improved Bayesian optimization algorithm includes optimizing the parameters in the prior function and optimizing the acquisition function, wherein optimizing the parameters in the prior function includes: selecting a Gaussian process to model the objective function, the function obeying the distribution of the objective function includes a mean function and a kernel function, the kernel function is a composite kernel function obtained by multiplying two single kernel functions, a Marten kernel function and a square exponential kernel function, and optimizing one or more parameters corresponding to the composite kernel function using an improved Markov chain Monte Carlo method based on flower pollination, calculating the distribution of the parameters to be optimized, and then calculating the corresponding expected values, thereby obtaining the optimized parameter values; the optimization of the acquisition function includes: optimizing and improving the adjustment parameters therein using an adaptive confidence boundary strategy;

[0075] The model training module is used to use the obtained optimal hyperparameter combination as the parameters in the final model and re-use the training set to train the XGBoost model;

[0076] The prediction module is used to use the trained XGBoost model to predict the fault type of the test set, output the prediction results and analyze the diagnosis effect.

[0077] Further, including:

[0078] In the model pre-training module, a Gaussian process is used to model the objective function. The functions that obey the distribution of the objective function include a mean function and a kernel function, specifically including:

[0079] The Gaussian process GP is used to model the objective function. GP is regarded as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function that obeys its distribution consists of a mean function and a kernel function, namely:

[0080] f(x)~GP(m(x),k(x,x′));

[0081] Where, the mean function m(x) = E[f(x)], and the kernel function k(x,x′) = E[(f(x)-m(x))(f(x′)-m(x′))] are used to indicate the similarity between two data sets. For the hyperparameter optimization process, x is a series of hyperparameter combinations, and f(x) is the final evaluation index score. Therefore, without loss of generality, let m(x) = 0.

[0082] If the output value on the training set is f=(f1,f2,...,f n ) T , the predicted function value f(x * )=f * , then the following joint distribution exists in the properties of the Gaussian process:

[0083]

[0084] Where K is the Gram matrix k(X,X), k * ={k(x * ,x1),k(x * ,x2),...,k(x * ,x n )} T ,

[0085] But the objective function f(x) and the observed value y=(y1,y2,...,y n ) T There is usually an error between them, which is represented by noise∈, that is:

[0086] y=f(x)+∈;

[0087] And assume that the noise ∈ i independent and identically distributed, and Then the prior of the kernel function becomes:

[0088]

[0089] Then we get y and f * The joint distribution of is:

[0090]

[0091] Where I is the identity matrix, and denote the variance of the objective function and the variance of the noise, respectively.

[0092] According to the properties of multivariate normal distribution, we get:

[0093] f * |x * ,y~N(μ(x * ),σ 2 (x * ));

[0094] μ(x * ) is the final predicted value, σ 2 (x * ) is the prediction variance,

[0095] The traditional Marton kernel function and the square exponential kernel function are multiplied and compounded to obtain a more accurate fitting result. The Marton-square exponential composite function is:

[0096]

[0097] Where r = xx′ represents the Euclidean distance between sample points, l is the scale parameter, and α is the smoothing parameter.

[0098] Further, including:

[0099] One or more parameters corresponding to the composite kernel function are optimized using an improved Markov chain Monte Carlo method based on flower pollination, including:

[0100] Arbitrarily select the initial parameter θ0 to be optimized and execute the loop i=1,2,...,k times;

[0101] Let the value of the parameter to be optimized for the current number of visits be θ i-1 =θ, randomly select the next candidate parameter value θ′ according to the proposed distribution;

[0102] According to the flower pollination FPA, when the conversion probability between global optimization and local optimization p>0.8, the next candidate parameter θ′ is calculated globally as:

[0103]

[0104] Where, is the sample generated by the mth Markov chain at the i-th iteration, is the optimal sampling point among the currently generated samples, L is the search step length, and its calculation formula is as follows:

[0105]

[0106] In the formula, u=i -λ , Γ is the standard gamma function, λ is usually 1-5, v ~ N(0,1);

[0107] If p≤0.8, the next candidate parameter θ′ is:

[0108]

[0109] Where ξ is a random value uniformly distributed on [0,1], and are the samples generated by the m-th Markov chain at the j-th and k-th iterations, respectively. and are equivalent to random solutions in the model;

[0110] Using the updated acceptance probability, we can get the posterior samples θ1, θ2..., θ k ;

[0111] Therefore, the expected value of the distribution of the parameters to be optimized is expressed as:

[0112]

[0113] Further, including:

[0114] The acceptance probability of the update is expressed as:

[0115]

[0116] Where p(·) is the given prior distribution and q(·) is the given proposed distribution;

[0117] Randomly select u from uniform distribution, u∈(0,1);

[0118] If u≤α(θ,θ′), then accept the transfer θ i =θ′, otherwise, do not accept the transfer, and θ i =θ i-1 After the loop is executed, we can get the posterior samples θ1, θ2..., θ k .

[0119] Further, including:

[0120] The adaptive confidence boundary strategy is used to optimize and improve the adjustment parameters, specifically including:

[0121] The acquisition function adopts an adaptive confidence boundary strategy to sample in unknown areas or explore near the currently considered global optimal solution, which is expressed as:

[0122] AUCB(x)=μ(x)+εσ(x);

[0123] In the formula, the adaptive adjustment parameter s=|y max +D max | / (y max +D max ), y max Indicates the maximum value in the current fitting curve, D max It represents the maximum value in the current observation set, and s is the coefficient of difference, which is used to further represent y max With D max degree of difference.

[0124] In a third aspect, the present invention further provides a cross-connection cable fault diagnosis device based on improved Bayesian optimization XGBoost, wherein the cross-connection cable fault diagnosis device based on improved Bayesian optimization XGBoost comprises: a memory, a processor, and a cross-connection cable fault diagnosis program based on improved Bayesian optimization XGBoost stored in the memory and executable on the processor, wherein when the fault diagnosis program is executed by the processor, the steps of the cross-connection cable fault diagnosis method based on improved Bayesian optimization XGBoost as described above are implemented.

[0125] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0126] The present invention uses an improved Bayesian algorithm to optimize the hyperparameters of XGBoost, obtains an optimized XGBoost model, and uses the trained XGBoost model to diagnose faults. Specifically, first, the present invention uses a Gaussian process to model the objective function, and the function that obeys its distribution is composed of a mean function and a kernel function. The present invention first abandons the use of a single kernel function and optimizes the improved kernel function, that is, the traditional Marten kernel function and the square exponential kernel function, two more typical single kernel functions, are multiplied and compounded to obtain a more accurate fitting result; the optimization method used is the Markov chain Monte Carlo method (MCMC) improved based on flower pollination (FPA); secondly, the acquisition function in the Bayesian optimization algorithm is optimized, thereby effectively improving the accuracy and efficiency of model diagnosis.

[0127] Compared with other mainstream classification models, XGBoost has better fault diagnosis effect, and combined with the improved Bayesian algorithm to optimize XGBoost's hyperparameters, the accuracy and efficiency of fault detection can be improved.

[0128] In summary, the present application first improves the Bayesian optimization algorithm. For the Gaussian process in the prior function, the Marton kernel function and the square exponential kernel function are multiplied and compounded as the kernel function. The improved Markov chain Monte Carlo method is used to optimize the parameters of the kernel function. Secondly, for the acquisition function, the adjustment parameters in the traditional confidence boundary strategy are improved. Finally, the improved Bayesian optimization algorithm is used to intelligently search and optimize the hyperparameters of XGBoost, which effectively improves the prediction performance and generalization ability of the model, and realizes efficient and accurate diagnosis of various faults of cross-connected cables. The implementation of the present invention will help to improve the automation level of cable fault processing, reduce operation and maintenance costs, have important practical application value and promotion prospects, and are of great significance to ensuring the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0129] Figure 1 This is a flow chart of a cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost according to an embodiment of the present invention;

[0130] Figure 2 The present invention provides a flowchart of a method for diagnosing a cross-connect cable fault according to an embodiment of the present invention. DETAILED DESCRIPTION

[0131] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention and not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0132] like Figure 1 As shown, the present invention provides a cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost, the method comprising:

[0133] S1 uses the collected circulation samples to construct a multidimensional feature vector I, constructs circulation sample data based on multiple multidimensional feature vectors I, and divides the circulation sample data into a training set and a test set according to a certain ratio. The circulation samples are obtained from the monitoring background by installing an online monitoring device for sheath circulation at the cable line circulation monitoring point, and then constructs a multidimensional feature vector I with all the measured circulation amplitudes;

[0134] S2 sets the range of XGBoost model hyperparameters, and pre-trains the XGBoost model based on the training set using an improved Bayesian optimization algorithm to find the optimal hyperparameter combination;

[0135] The improved Bayesian optimization algorithm includes optimizing the parameters in the prior function and optimizing the acquisition function, wherein optimizing the parameters in the prior function includes: selecting a Gaussian process to model the objective function, the function obeying the distribution of the objective function includes a mean function and a kernel function, the kernel function is a composite kernel function obtained by multiplying two single kernel functions, a Marten kernel function and a square exponential kernel function, and optimizing one or more parameters corresponding to the composite kernel function using an improved Markov chain Monte Carlo method based on flower pollination, calculating the distribution of the parameters to be optimized, and then calculating the corresponding expected values, thereby obtaining the optimized parameter values; the optimization of the acquisition function includes: optimizing and improving the adjustment parameters therein using an adaptive confidence boundary strategy;

[0136] S3 uses the optimal hyperparameter combination as the parameters in the final model and re-uses the training set to train the XGBoost model;

[0137] S4 uses the trained XGBoost model to predict the fault type of the test set, outputs the prediction results and analyzes the diagnosis effect.

[0138] In the above-mentioned cross-connection cable fault diagnosis method based on improved Bayesian optimization XGBoost, in step 2, the construction of the multidimensional feature vector is achieved by installing an online monitoring device for sheath circulation at the cable line circulation monitoring point, obtaining relevant circulation data from the monitoring background, and measuring the circulation on the left and right sides of the metal sheath lead-out line in the cross-connection box at the same time, and constructing a multidimensional feature vector with all the measured circulation amplitudes.

[0139] In the aforementioned cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost, in step 2, the XGBoost model uses a classification regression tree as the base learner, and the objective function includes a loss function and a regularization term. The regularization term is used to penalize model complexity to avoid overfitting. The overall prediction performance is improved by integrating the prediction results of a series of CART models. The Bayesian optimization algorithm is an optimization algorithm based on probability distribution. Specifically, a prior distribution model of the objective function is assumed, and then the model is continuously optimized using sample data to ultimately obtain a posterior distribution model of the objective function. For the prior function, the objective function is modeled by multiplying the Marten kernel function and the squared exponential kernel function together as the kernel function of a Gaussian process. GP can be viewed as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function obeying its distribution consists of a mean function and a kernel function. The kernel function is used to represent the degree of similarity between two data points. The acquisition function adopts an adaptive confidence bound strategy to sample in unknown areas or explore near the currently considered global optimal solution.

[0140] In the above-mentioned cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost, in step 2, the acquisition function adopts an adaptive confidence margin strategy. When the adaptive adjustment parameter approaches 1, the acquisition function based on the adaptive confidence margin strategy tends to be more in the exploration state; when it approaches 0, the acquisition function tends to be more in the exploitation state.

[0141] Specifically, in order to implement the above method, this embodiment provides the following specific implementation methods, such as Figure 2 As shown, this method can effectively diagnose the fault of the cross-connect cable. The following are detailed descriptions:

[0142] The first step is to use the collected circulation sample data to construct a multidimensional feature vector I, and divide the sample data into a training set and a test set according to a certain ratio (such as 4:1).

[0143] The construction of the multidimensional feature vector in the first step of the method is to install an online monitoring device for the sheath circulation at the cable line circulation monitoring point, obtain relevant circulation data from the monitoring background, and measure the circulation on both sides of the metal sheath lead-out line in the cross-connection box at the same time in order to effectively distinguish the fault type, and construct a multidimensional feature vector I with all the measured circulation amplitudes.

[0144] The second step is to set the range of XGBoost model hyperparameters, use the improved Bayesian optimization algorithm to pre-train the XGBoost model, and find the optimal hyperparameter combination by continuously optimizing the acquisition function of the Gaussian process.

[0145] The Extreme Gradient Boosting (XGBoost) algorithm in the second step uses the Classification and Regression Tree (CART) as the base learner. It improves the overall prediction effect by integrating the prediction results of a series of CARTs. The classification principle of XGBoost is explained in detail below:

[0146] Assume that the training data set D = {(x i ,y i )}(x i ∈R m ,y i ∈R), where x i is a feature vector with m dimensions, y i is the sample label. For cross-connection system fault diagnosis, x i That is, the characteristic vector of each sheath circulation, y i is the corresponding fault category, consisting of K CART trees, then the XGBoost model is:

[0147]

[0148] F={f(x)=ω q(x)},q:R m →T,ω∈R T (2)

[0149] Where, is the final prediction value of the XGBoost model for the i-th sample; f k is the kth tree. Each CART tree maps each sample to a leaf node according to the sample characteristics and classification rules. Each leaf node corresponds to a weight score ω, which is the predicted value of the sample falling on the corresponding leaf node. q is the structure vector corresponding to each CART tree, ω q(x) is the predicted value of CART for sample x; F is the set formed by K CART trees; T is the total number of leaf nodes; for sample x i , when all CART trees have completed the classification, the x i The weights of the leaf nodes are added together to get the sample x i The final predicted value of .

[0150] In order to train each CART tree well and make the XGBoost model have excellent performance, a regularization term (penalty term) is added to the XGBoost objective function. The specific function is as follows:

[0151]

[0152]

[0153] The objective function above consists of two parts. The first term is the loss function, a differentiable convex function that represents the error between the model's predicted value and the true value. The second term is the regularization term, which penalizes model complexity to prevent overfitting. φ is the set of functions in the model. Here, γ and λ are two parameters that control complexity; larger values ​​of these two parameters reduce overfitting. γ suppresses model complexity by controlling the number of leaf nodes. The second term on the right is the L2 regularization term, and λ reduces complexity by controlling the weight fraction of leaf nodes.

[0154] XGBoost starts to optimize from the first sub-model during model training, finds the optimal sub-model under the current conditions, and then optimizes the subsequent sub-models in sequence. When training the t-th sub-model f t When t-1 sub-models are known, the goal of the t-th round of optimization is to find an optimal f t , so that the objective function value is minimized, then the new model after the tth round of training is the t-1th round model plus the optimal f tThe predicted value during model training and the objective function L of the t-th round optimization (t) It can be expressed as follows:

[0155]

[0156] Where, is the predicted value of the model after the tth round of training.

[0157] Perform a second-order Taylor expansion on the objective function, ignoring the constant term, and the objective function formula is:

[0158]

[0159] f t (x i ) is essentially a tree model, so x i After mapping, it will definitely fall on a node in the tree. Define a sample set I of leaf node j j ,Right now

[0160] I j ={i|q(x i )=j}(7)

[0161] Rewrite the objective function as

[0162]

[0163] For a fixed tree structure, the optimal weight of its leaf node j is

[0164]

[0165] The leaf node weight depends on the first-order and second-order gradient statistics and the L2 regularization coefficient λ. After finding the optimal weight, it is easy to get the optimal objective function value

[0166]

[0167] The above formula is a scoring function for evaluating the quality of a tree. The smaller the value, the better the tree model. Therefore, by calculating the score of each tree model, the optimal tree model can be selected. XGBoost starts from the root node, calculates the score before and after the node split, and traverses all the split points of all features to find the best split point and the optimal feature.

[0168] The present invention improves the Bayesian Optimization (BO) algorithm. First, the algorithm is an optimization algorithm based on probability distribution, that is, the prior distribution model of the objective function is assumed in advance, and then the model is continuously optimized using sample data, and finally the posterior distribution model of the objective function is obtained. Its main process is the construction of the prior function and the acquisition function.

[0169] The prior function in the improved Bayesian optimization algorithm uses Gaussian Process (GP) to model the objective function. GP can be regarded as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function that obeys its distribution consists of a mean function and a kernel function, that is,

[0170] f(x)~GP(m(x),k(x,x′))(11)

[0171] Where m(x) = E[f(x)], and k(x, x′) = E[(f(x) - m(x))(f(x′) - m(x′))] are the mean function and kernel function, respectively. For hyperparameter optimization, x represents a series of hyperparameter combinations, and f(x) represents the final evaluation metric score. Without loss of generality, let m(x) = 0.

[0172] If the output value on the training set is f=(f1,f2,...,f n ) T , the predicted function value f(x * )=f * , then the following joint distribution exists in the properties of the Gaussian process:

[0173]

[0174] Where K is the Gram matrix k(X,X), k * ={k(x * ,x1),k(x * ,x2),...,k(x * ,x n )} T ,

[0175] But the objective function f(x) and the observed value y=(y1,y2,...,y n ) T There is usually a certain error between them, which is represented by noise∈, that is:

[0176] y=f(x)+∈(13)

[0177] And assume that the noise ∈ i independent and identically distributed, and Then the prior of the kernel function becomes:

[0178]

[0179] Then we get y and f * The joint distribution of is:

[0180]

[0181] Where I is the identity matrix, and denote the variance of the objective function and the variance of the noise, respectively.

[0182] According to the properties of multivariate normal distribution, we can get:

[0183] f * |x * ,y~N(μ(x * ),σ 2 (x * ))(16)

[0184] μ(x * ) is the final predicted value, σ 2 (x * ) is the prediction variance,

[0185] The Gaussian process observes real data points on the unknown target function and uses a kernel function to measure the relationship between any two observation points, thereby inferring the distribution of the unknown target function. Compared to single kernel functions, composite kernel functions have more characteristics and can better fit different types of data. The traditional Marton kernel function and the squared exponential kernel function, two more representative single kernel functions, are multiplied together to form a composite function to obtain a more accurate fitting result. The Marton-squared exponential composite function is:

[0186]

[0187] Where r = xx′ represents the Euclidean distance between sample points, l is the scale parameter, and α is the smoothing parameter.

[0188] In this embodiment, the parameters of the kernel function, such as the scale parameter, the objective function variance, and the noise variance, are optimized using the Markov chain Monte Carlo method (MCMC) improved based on flower pollination (FPA). The distribution of the parameters is obtained by the acceptance-rejection method, and then their expected values ​​are calculated to obtain the optimized parameter values.

[0189] The acceptance probability is:

[0190]

[0191] Where p(·) is the given prior distribution and q(·) is the given proposed distribution.

[0192] The specific steps for optimizing the parameter θ are:

[0193] (1) Arbitrarily select the initial parameter θ0;

[0194] (2) Loop execution for i=1,2,...,k:

[0195] Let the currently accessed parameter value θ i-1 =θ, randomly extract the next candidate parameter value θ′ according to the proposed distribution.

[0196] According to the basic principle of FPA, when the conversion probability between global optimization and local optimization p>0.8, the next candidate parameter θ′ is calculated globally as:

[0197]

[0198] Where, is the sample generated by the mth Markov chain at the i-th iteration, is the optimal sampling point among the currently generated samples, L is the search step length, and its calculation formula is as follows:

[0199]

[0200] In the formula, u=i -λ , Γ is the standard gamma function, λ is usually 1-5, v~N(0,1).

[0201] If p≤0.8, the next candidate parameter θ′ is:

[0202]

[0203] Where ξ is a random value uniformly distributed on [0,1], and are the samples generated by the m-th Markov chain at the j-th and k-th iterations, respectively. and are equivalent to random solutions in the model.

[0204] Update acceptance probability:

[0205]

[0206] Randomly draw u from uniform distribution, u∈(0,1). If u≤α(θ,θ′), then accept the transfer θ i =θ′, otherwise the transfer is not accepted, θ i =θ i-1 After the loop is executed, we can get the posterior samples θ1, θ2..., θ k .

[0207] (3) Calculate the expected value of the parameter distribution as:

[0208]

[0209] The improved Markov chain Monte Carlo method can increase the efficiency of optimizing the candidate parameters θ′ and improve the acceptance rate.

[0210] At the same time, the present invention also improves the acquisition function. Specifically, the acquisition function in the improved Bayesian optimization algorithm adopts the adaptive upper confidence bound strategy (The Adaptive Upper Confidence Bbound Algorithm, AUCB). By improving the adjustment parameters in the traditional confidence bound strategy, the improved adjustment parameters have the ability to self-adapt and update during the calculation process. Its acquisition function is expressed as:

[0211] AUCB(x)=μ(x)+εσ(x); (24)

[0212] In the formula, the adaptive adjustment parameter s=|y max +D max | / (y max +D max ).y max Indicates the maximum value in the current fitting curve, D max It represents the maximum value in the current observation set, and s is the coefficient of difference, which is used to further represent y max With D max degree of difference.

[0213] When y max With D max When it is infinitely close, ε tends to 1. At this time, the acquisition function based on the adaptive confidence boundary strategy is more inclined to the exploration state; when y max With D max When it is infinitely far away, ε tends to 0, and the acquisition function is more inclined to the development state.

[0214] The third step is to use the optimal hyperparameter combination obtained in the second step as the parameters in the final model, and re-use the training set to train the XGBoost model.

[0215] Step 4: Use the trained XGBoost model to predict the fault type of the test set, output the prediction results and analyze the diagnostic effect.

[0216] This paper proposes a cross-connected high-voltage cable fault diagnosis method based on improved Bayesian optimization (XGBoost). The implementation principles of the XGBoost algorithm and the improved Bayesian optimization algorithm are detailed. First, based on engineering practice and fault circulation characteristics, a multidimensional feature vector based on the circulating current amplitude at the monitoring point is constructed. The improved Bayesian algorithm is then used to optimize the XGBoost hyperparameters to obtain an optimized XGBoost model. Finally, the trained XGBoost model is used to diagnose the fault. Consequently, this diagnostic model can effectively identify various fault types and significantly improve the accuracy of model diagnosis.

[0217] Secondly, the present invention also provides a cross-connect cable fault diagnosis method system based on improved Bayesian optimization XGBoost, which includes:

[0218] The acquisition module is used to construct a multidimensional feature vector I using the collected circulation sample data, and divide the sample data into a training set and a test set according to a certain ratio. The circulation sample data is obtained from the monitoring background by installing an online monitoring device for sheath circulation at the cable line circulation monitoring point, and then constructing a multidimensional feature vector I with all the measured circulation amplitudes;

[0219] A model pre-training module is used to set the range of XGBoost model hyperparameters and pre-train the XGBoost model based on the training set using an improved Bayesian optimization algorithm to find the optimal hyperparameter combination;

[0220] The improved Bayesian optimization algorithm includes optimizing the parameters in the prior function and optimizing the acquisition function, wherein optimizing the parameters in the prior function includes: selecting a Gaussian process to model the objective function, the function obeying the distribution of the objective function includes a mean function and a kernel function, the kernel function is a composite kernel function obtained by multiplying two single kernel functions, a Marten kernel function and a square exponential kernel function, and optimizing one or more parameters corresponding to the composite kernel function using an improved Markov chain Monte Carlo method based on flower pollination, calculating the distribution of the parameters to be optimized, and then calculating the corresponding expected values, thereby obtaining the optimized parameter values; the optimization of the acquisition function includes: optimizing and improving the adjustment parameters therein using an adaptive confidence boundary strategy;

[0221] The model training module is used to use the obtained optimal hyperparameter combination as the parameters in the final model and re-use the training set to train the XGBoost model;

[0222] The prediction module is used to use the trained XGBoost model to predict the fault type of the test set, output the prediction results and analyze the diagnosis effect.

[0223] Further, including:

[0224] In the model pre-training module, a Gaussian process is used to model the objective function. The functions that obey the distribution of the objective function include a mean function and a kernel function, specifically including:

[0225] The Gaussian process GP is used to model the objective function. GP is regarded as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function that obeys its distribution consists of a mean function and a kernel function, namely:

[0226] f(x)~GP(m(x),k(x,x′));

[0227] Where, the mean function m(x) = E[f(x)], and the kernel function k(x,x′) = E[(f(x)-m(x))(f(x′)-m(x′))] are used to indicate the similarity between two data sets. For the hyperparameter optimization process, x is a series of hyperparameter combinations, and f(x) is the final evaluation index score. Therefore, without loss of generality, let m(x) = 0.

[0228] If the output value on the training set is f=(f1,f2,...,f n ) T , the predicted function value f(x * )=f * , then the following joint distribution exists in the properties of the Gaussian process:

[0229]

[0230] Where K is the Gram matrix k(X,X), k * ={k(x * ,x1),k(x * ,x2),...,k(x * ,x n )} T ,

[0231] But the objective function f(x) and the observed value y=(y1,y2,...,y n ) T There is usually an error between them, which is represented by noise∈, that is:

[0232] y=f(x)+∈;

[0233] And assume that the noise ∈ i independent and identically distributed, and Then the prior of the kernel function becomes:

[0234]

[0235] Then we get y and f * The joint distribution of is:

[0236]

[0237] Where I is the identity matrix, and denote the variance of the objective function and the variance of the noise, respectively.

[0238] According to the properties of multivariate normal distribution, we get:

[0239] f * |x * ,y~N(μ(x * ),σ 2 (x * ));

[0240] μ(x * ) is the final predicted value, σ 2 (x * ) is the prediction variance,

[0241] The traditional Marton kernel function and the square exponential kernel function are multiplied and compounded to obtain a more accurate fitting result. The Marton-square exponential composite function is:

[0242]

[0243] Where r = xx′ represents the Euclidean distance between sample points, l is the scale parameter, and α is the smoothing parameter.

[0244] Further, including:

[0245] One or more parameters corresponding to the composite kernel function are optimized using an improved Markov chain Monte Carlo method based on flower pollination, including:

[0246] Arbitrarily select the initial parameter θ0 to be optimized and execute the loop i=1,2,...,k times;

[0247] Let the value of the parameter to be optimized for the current number of visits be θ i-1 =θ, randomly select the next candidate parameter value θ′ according to the proposed distribution;

[0248] According to the flower pollination FPA, when the conversion probability between global optimization and local optimization p>0.8, the next candidate parameter θ′ is calculated globally as:

[0249]

[0250] Where, is the sample generated by the mth Markov chain at the i-th iteration, is the optimal sampling point among the currently generated samples, L is the search step length, and its calculation formula is as follows:

[0251]

[0252] In the formula, u=i -λ , Γ is the standard gamma function, λ is usually 1-5, v ~ N(0,1);

[0253] If p≤0.8, the next candidate parameter θ′ is:

[0254]

[0255] Where ξ is a random value uniformly distributed on [0,1], and are the samples generated by the m-th Markov chain at the j-th and k-th iterations, respectively. and are equivalent to random solutions in the model;

[0256] Using the updated acceptance probability, we can get the posterior samples θ1, θ2..., θ k ;

[0257] Therefore, the expected value of the distribution of the parameters to be optimized is expressed as:

[0258]

[0259] Further, including:

[0260] The acceptance probability of the update is expressed as:

[0261]

[0262] Where p(·) is the given prior distribution and q(·) is the given proposed distribution;

[0263] Randomly select u from uniform distribution, u∈(0,1);

[0264] If u≤α(θ,θ′), then accept the transfer θ i =θ′, otherwise, do not accept the transfer, and θ i =θ i-1 After the loop is executed, we can get the posterior samples θ1, θ2..., θ k .

[0265] Further, including:

[0266] The adaptive confidence boundary strategy is used to optimize and improve the adjustment parameters, specifically including:

[0267] The acquisition function adopts an adaptive confidence boundary strategy to sample in unknown areas or explore near the currently considered global optimal solution, which is expressed as:

[0268] AUCB(x)=μ(x)+εσ(x);

[0269] In the formula, the adaptive adjustment parameter s=|y max +D max | / (y max +D max ), y max Indicates the maximum value in the current fitting curve, D max It represents the maximum value in the current observation set, and s is the coefficient of difference, which is used to further represent y max With D max degree of difference.

[0270] Finally, the present invention also provides a cross-connection cable fault diagnosis device based on improved Bayesian optimization XGBoost, and the cross-connection cable fault diagnosis device based on improved Bayesian optimization XGBoost includes: a memory, a processor, and a cross-connection cable fault diagnosis program based on improved Bayesian optimization XGBoost stored in the memory and capable of running on the processor. When the fault diagnosis program is executed by the processor, the steps of the cross-connection cable fault diagnosis method based on improved Bayesian optimization XGBoost as described above are implemented.

[0271] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0272] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0273] Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if such changes and modifications of the embodiments of the present invention fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost, characterized in that: The method includes: The collected circulation samples are used to construct a multidimensional feature vector. Multiple multidimensional feature vectors I constitute the entire circulation sample data, and the circulation sample data is divided into a training set and a test set according to a certain ratio; Setting the range of XGBoost model hyperparameters, and pre-training the XGBoost model using an improved Bayesian optimization algorithm based on the training set to find the optimal hyperparameter combination; The improved Bayesian optimization algorithm includes optimizing the parameters in the prior function and optimizing the acquisition function, wherein optimizing the parameters in the prior function includes: selecting a Gaussian process to model the objective function, the function obeying the distribution of the objective function includes a mean function and a kernel function, the kernel function is a composite kernel function obtained by multiplying two single kernel functions, a Marten kernel function and a square exponential kernel function, and optimizing one or more parameters corresponding to the composite kernel function using an improved Markov chain Monte Carlo method based on flower pollination, calculating the distribution of the parameters to be optimized, and then calculating the corresponding expected values, thereby obtaining the optimized parameter values; the optimization of the acquisition function includes: optimizing and improving the adjustment parameters therein using an adaptive confidence boundary strategy; The optimal hyperparameter combination is used as the parameters in the final model, and the training set is reused to train the XGBoost model; The trained XGBoost model is used to predict the fault type of the test set, output the prediction results and analyze the diagnostic effect.

2. The cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost according to claim 1 is characterized in that: The Gaussian process is used to model the objective function. The functions that obey the distribution of the objective function include a mean function and a kernel function, specifically including: The Gaussian process GP is used to model the objective function. GP is regarded as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function that obeys its distribution consists of a mean function and a kernel function, namely: f(x)~GP(m(x),k(x,x′)); Where, the mean function m(x) = E[f(x)], and the kernel function k(x,x′) = E[(f(x)-m(x))(f(x′)-m(x′))] are used to indicate the similarity between two data sets. For the hyperparameter optimization process, x is a series of hyperparameter combinations, and f(x) is the final evaluation index score. Therefore, let m(x) = 0. If the output value on the training set is f=(f1,f2,...,f n ) T , the predicted function value f(x * )=f * , then the following joint distribution exists in the properties of the Gaussian process: Where K is the Gram matrix k(X,X), k * ={k(x * ,x1),k(x * ,x2),...,k(x * ,x n )} T , But the objective function f(x) and the observed value y=(y1,y2,...,y n ) T There is an error between them, which is represented by noise∈, that is: y=f(x)+∈; And assume that the noise ∈ is independent and identically distributed, and Then the prior of the kernel function becomes: Then we get y and f * The joint distribution of is: Where I is the identity matrix, and denote the variance of the objective function and the variance of the noise, respectively. According to the properties of multivariate normal distribution, we get: f * |x * ,y~N(μ(x * ),s 2 (x * )); μ(x * ) is the final predicted value, σ 2 (x * ) is the prediction variance, The traditional Marton kernel function and the square exponential kernel function are multiplied and compounded to obtain a more accurate fitting result. The Marton-square exponential composite function is: Where r = xx′ represents the Euclidean distance between sample points, l is the scale parameter, and α is the smoothing parameter.

3. The cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost according to claim 2 is characterized in that: One or more parameters corresponding to the composite kernel function are optimized using an improved Markov chain Monte Carlo method based on flower pollination, including: Arbitrarily select the initial parameter θ0 to be optimized and execute the loop i=1,2,...,k times; Let the value of the parameter to be optimized for the current number of visits be θ i-1 =θ, randomly select the next candidate parameter value θ′ according to the proposed distribution; According to the flower pollination FPA, when the conversion probability between global optimization and local optimization p>0.8, the next candidate parameter θ′ is calculated globally as: Where, is the sample generated by the mth Markov chain at the i-th iteration, is the optimal sampling point among the currently generated samples, L is the search step length, and its calculation formula is as follows: In the formula, u=i -λ , Γ is the standard gamma function, λ is usually 1-5, v ~ N(0,1); If p≤0.8, the next candidate parameter θ′ is: Where ξ is a random value uniformly distributed on [0,1], and are the samples generated by the m-th Markov chain at the j-th and k-th iterations, respectively. and are equivalent to random solutions in the model; Using the updated acceptance probability, we get the posterior samples θ1, θ2..., θ k ; Therefore, the expected value of the distribution of the parameters to be optimized is expressed as:

4. The cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost according to claim 3 is characterized in that: The acceptance probability of the update is expressed as: Where p(·) is the given prior distribution and q(·) is the given proposed distribution; Randomly select u from uniform distribution, u∈(0,1); If u≤α(θ,θ′), then accept the transfer θ i =θ′, otherwise, do not accept the transfer, and θ i =θ i-1 After the loop is executed, we can get the posterior samples θ1, θ2..., θ k .

5. The cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost according to claim 4 is characterized in that: The adaptive confidence boundary strategy is used to optimize and improve the adjustment parameters, specifically including: The acquisition function adopts an adaptive confidence boundary strategy to sample in unknown areas or explore near the currently considered global optimal solution, which is expressed as: AUCB(x)=μ(x)+εσ(x); In the formula, the adaptive adjustment parameter s=|y max +D max | / (y max +D max ), y max Indicates the maximum value in the current fitting curve, D max It represents the maximum value in the current observation set, and s is the coefficient of difference, which is used to further represent y max With D max The observation set refers to the set of adjustment parameters and their corresponding objective function values ​​that have been evaluated during the optimization process. In the context of Bayesian optimization, the fitting curve refers to the prediction curve obtained after modeling the objective function using a Gaussian process.

6. The cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost according to claim 5 is characterized in that: The adaptive adjustment parameters and y max 、D max The relationships include: When y max With D max When it approaches infinity, ε tends to 1, and the acquisition function based on the adaptive confidence boundary strategy is more inclined to the exploration state; When y max With D max When it is infinitely far away, ε tends to 0, and the acquisition function is more inclined to the development state.

7. A cross-connect cable fault diagnosis system based on improved Bayesian optimization XGBoost, characterized in that: The system includes: The acquisition module is used to construct a multidimensional feature vector using the collected circulation samples. Multiple multidimensional feature vectors I constitute the entire circulation sample data, and the circulation sample data is divided into a training set and a test set according to a certain ratio; A model pre-training module is used to set the range of XGBoost model hyperparameters and pre-train the XGBoost model based on the training set using an improved Bayesian optimization algorithm to find the optimal hyperparameter combination; The improved Bayesian optimization algorithm includes optimizing the parameters in the prior function and optimizing the acquisition function, wherein optimizing the parameters in the prior function includes: selecting a Gaussian process to model the objective function, the function obeying the distribution of the objective function includes a mean function and a kernel function, the kernel function is a composite kernel function obtained by multiplying two single kernel functions, a Marten kernel function and a square exponential kernel function, and optimizing one or more parameters corresponding to the composite kernel function using an improved Markov chain Monte Carlo method based on flower pollination, calculating the distribution of the parameters to be optimized, and then calculating the corresponding expected values, thereby obtaining the optimized parameter values; the optimization of the acquisition function includes: optimizing and improving the adjustment parameters therein using an adaptive confidence boundary strategy; The model training module is used to use the obtained optimal hyperparameter combination as the parameters in the final model and re-use the training set to train the XGBoost model; The prediction module is used to use the trained XGBoost model to predict the fault type of the test set, output the prediction results and analyze the diagnosis effect.

8. The cross-connect cable fault diagnosis system based on improved Bayesian optimization XGBoost according to claim 7, characterized in that: In the model pre-training module, a Gaussian process is used to model the objective function. The functions that obey the distribution of the objective function include a mean function and a kernel function, specifically including: The Gaussian process GP is used to model the objective function. GP is regarded as a generalized extension of the multivariate Gaussian distribution in infinite dimensions. The function that obeys its distribution consists of a mean function and a kernel function, namely: f(x)~GP(m(x),k(x,x′)); Where, the mean function m(x) = E[f(x)], and the kernel function k(x,x′) = E[(f(x)-m(x))(f(x′)-m(x′))] are used to indicate the similarity between two data sets. For the hyperparameter optimization process, x is a series of hyperparameter combinations, and f(x) is the final evaluation index score. Therefore, let m(x) = 0. If the output value on the training set is f=(f1,f2,...,f n ) T , the predicted function value f(x * )=f * , then the following joint distribution exists in the properties of the Gaussian process: Where K is the Gram matrix k(X,X), k * ={k(x * ,x1),k(x * ,x2),...,k(x * ,x n )} T , But the objective function f(x) and the observed value y=(y1,y2,...,y n ) T There is an error between them, which is represented by noise∈, that is: y=f(x)+∈; And assume that the noise ∈ is independent and identically distributed, and Then the prior of the kernel function becomes: Then we get y and f * The joint distribution of is: Where I is the identity matrix, and denote the variance of the objective function and the variance of the noise, respectively. According to the properties of multivariate normal distribution, we get: f * |x * ,y~N(μ(x * ),s 2 (x * )); μ(x * ) is the final predicted value, σ 2 (x * ) is the prediction variance, The traditional Marton kernel function and the square exponential kernel function are multiplied and compounded to obtain a more accurate fitting result. The Marton-square exponential composite function is: Where r = xx′ represents the Euclidean distance between sample points, l is the scale parameter, and α is the smoothing parameter.

9. The cross-connect cable fault diagnosis system based on improved Bayesian optimization XGBoost according to claim 8, characterized in that: One or more parameters corresponding to the composite kernel function are optimized using an improved Markov chain Monte Carlo method based on flower pollination, including: Arbitrarily select the initial parameter θ0 to be optimized and execute the loop i=1,2,...,k times; Let the value of the parameter to be optimized for the current number of visits be θ i-1 =θ, randomly select the next candidate parameter value θ′ according to the proposed distribution; According to the flower pollination FPA, when the conversion probability between global optimization and local optimization p>0.8, the next candidate parameter θ′ is calculated globally as: Where, is the sample generated by the mth Markov chain at the i-th iteration, is the optimal sampling point among the currently generated samples, L is the search step length, and its calculation formula is as follows: In the formula, u=i -λ , Γ is the standard gamma function, λ is usually 1-5, v ~ N(0,1); If p≤0.8, the next candidate parameter θ′ is: Where ξ is a random value uniformly distributed on [0,1], and are the samples generated by the m-th Markov chain at the j-th and k-th iterations, respectively. and are equivalent to random solutions in the model; Using the updated acceptance probability, we get the posterior samples θ1, θ2..., θ k ; Therefore, the expected value of the distribution of the parameters to be optimized is expressed as:

10. The cross-connect cable fault diagnosis system based on improved Bayesian optimization XGBoost according to claim 9, characterized in that: The acceptance probability of the update is expressed as: Where p(·) is the given prior distribution and q(·) is the given proposed distribution; Randomly select u from uniform distribution, u∈(0,1); If u≤α(θ,θ′), then accept the transfer θ i =θ′, otherwise, do not accept the transfer, and θ i =θ i-1 After the loop is executed, we can get the posterior samples θ1, θ2..., θ k .

11. The cross-connect cable fault diagnosis system based on improved Bayesian optimization XGBoost according to claim 10, characterized in that: The adaptive confidence boundary strategy is used to optimize and improve the adjustment parameters, specifically including: The acquisition function adopts an adaptive confidence boundary strategy to sample in unknown areas or explore near the currently considered global optimal solution, which is expressed as: AUCB(x)=μ(x)+εσ(x); In the formula, the adaptive adjustment parameter s=|y max +D max | / (y max +D max ), y max Indicates the maximum value in the current fitting curve, D max It represents the maximum value in the current observation set, and s is the coefficient of difference, which is used to further represent y max With D max The observation set refers to the set of adjustment parameters and their corresponding objective function values ​​that have been evaluated during the optimization process. In the context of Bayesian optimization, the fitting curve refers to the prediction curve obtained after modeling the objective function using a Gaussian process.

12. A cross-connect cable fault diagnosis device based on improved Bayesian optimization XGBoost, characterized in that: The cross-connect cable fault diagnosis device based on improved Bayesian optimization XGBoost includes: a memory, a processor, and a cross-connect cable fault diagnosis program stored in the memory and capable of running on the processor based on improved Bayesian optimization XGBoost. When the fault diagnosis program is executed by the processor, the steps of the cross-connect cable fault diagnosis method based on improved Bayesian optimization XGBoost as described in any one of claims 1 to 6 are implemented.

Images