Fault detection method for cloud data center based on IFA-BP-Adaboost

Abstract

The present invention discloses a cloud data center fault detection method based on IFA-BP-Adaboost, including: collecting cloud data center system fault feature information, obtaining a monitoring data set; performing feature extraction on the monitoring data set to obtain a training sample set; determining the BP neural network structure, initializing the training sample weight distribution; using the IFA algorithm to calculate the initial optimal weight and threshold of the neural network, and constructing several IFA-BP neural networks; training the neural network constructed in the previous step, and calculating the weight of the network; dynamically constructing a cloud data center fault detection model based on IFA-BP-AdaBoost according to the calculated weight; performing feature extraction on the monitoring data collected by the target system, inputting it into the trained fault detection model, and judging whether a fault occurs. The method of the present invention can perform fault detection without affecting the operation of the system, and make a fault response in a timely manner when a fault occurs.

Description

Technical Field

[0001] The present invention belongs to the technical field of cloud fault detection, and in particular to a cloud data center fault detection method based on IFA-BP-Adaboost. Background Art

[0002] Cloud computing technology plays an increasingly important role in large-scale data processing and massive computing resource management. It can improve the reliability of enterprises' external services, reduce enterprise operating costs, and provide convenience for the development of Internet-related enterprises. However, the cloud computing platform is huge in scale and complex in structure, and its reliability faces huge challenges. At the same time, there are a large number of malicious attacks that may cause failures in the system and various software and hardware on it. Therefore, it is necessary for cloud service providers to apply fault detection technology in cloud systems and respond promptly when failures occur.

[0003] At present, there are some research results on fault detection algorithms for cloud platforms at home and abroad, which are mainly divided into two categories: fault detection based on machine learning and fault detection based on statistics. The fault detection method based on machine learning first builds a model for the target system and compares the system behavior with the benchmark to determine whether the system has a fault. For example, Pannu et al. proposed a fault detection method in a cloud computing environment, using a Bayesian model to detect faulty nodes, marking abnormal points through system administrators, and using the marked data to build a decision tree for fault prediction. However, the decision tree has an over-learning problem, which has a great impact on the detection results. Ma Yong et al. proposed a cloud system fault detection method based on Pytorch and a bidirectional GRU network. After data preprocessing, embedded technology is used to enable the neural network to extract relevant data and use these features for further processing. Finally, the Adam optimization algorithm is used to train the neural network. Chirag N.Modi proposed a cloud computing fault detection method that combines Snort and Bayesian networks. Although these methods can effectively detect and predict faults in cloud environments, the detection method based on machine learning has difficulties in the selection of kernel functions and parameters, which has a great impact on the accuracy of the detection results. The detection method based on statistics assumes that the distribution of data is known a priori in many cases, and the dynamic operation of the system is described by a probability model to obtain the detection situation. For example, Karan Mitra et al. proposed a Bayesian system for cloud performance diagnosis and prediction. The system uses a Bayesian network to simulate the uncertainty and complex relationship between related factors. However, due to the complexity of the Bayesian network establishment process and the need to consider the network structure and sample integrity, it cannot be well applied in large cloud data centers. Wang Tao et al. proposed an adaptive fault detection method for cloud data centers. This method uses principal component analysis to analyze the operating status of the system, uses cosine similarity to calculate the degree of abnormality of the system, and finally establishes a reliability model based on exponential distribution to predict failure time and adjust the monitoring cycle. However, the detection method based on statistics often relies heavily on certain probability distribution assumptions and cannot adapt well to the complex and changing environment of cloud computing.

[0004] In summary, existing fault detection methods in cloud environments still have problems such as some machine learning-based methods have difficulty in selecting some parameters and statistics-based methods often depend to a large extent on assumptions about probability distribution. Summary of the invention

[0005] The purpose of the present invention is to provide a cloud data center fault detection method based on IFA-BP-Ababoost to address the problems existing in the prior art, which can improve the accuracy of fault detection and further improve the reliability of the cloud system.

[0006] The technical solution to achieve the purpose of the present invention is: a cloud data center fault detection method based on IFA-BP-Adaboost, the method comprising the following steps:

[0007] Step 1: monitor the operating status of the cloud data center system, collect fault feature information, and obtain a monitoring data set;

[0008] Step 2: Extract features from the monitoring data set using principal component analysis to obtain a training sample set;

[0009] Step 3, determine the structure of the BP neural network for cloud data center fault detection and initialize the weight distribution of training samples;

[0010] Step 4, using the improved firefly algorithm IFA to calculate and optimize the initial optimal weights and thresholds of the BP neural network, and constructing several IFA-BP neural networks;

[0011] Step 5, using the training sample set to train each IFA-BP neural network, and calculating the weight of each IFA-BP neural network according to the Adaboost algorithm;

[0012] Step 6, dynamically constructing a cloud data center fault detection model based on IFA-BP-AdaBoost according to the weights of each IFA-BP neural network;

[0013] Step 7: Extract features from the monitoring data collected by the target system and input them into the trained IFA-BP-AdaBoost cloud data center fault detection model to analyze and determine whether a fault occurs.

[0014] Furthermore, the fault feature information F in step 1 is a four-tuple:

[0015] F=(F cpu ,F Networks ,F Storage ,F I / O ) (1)

[0016] Among them, F cpu represents the computing resource attributes of the cloud data center, F cpu =<x1,x2,x3> , x1 represents the CPU usage at the user level, x2 represents the CPU usage at the system level, and x3 represents the idle CPU percentage; F Networks represents the network resource attributes of the cloud data center, F Networks =<x4,x5,x6,x7> , x4 represents the number of packets with errors when receiving, x5 represents the number of packets received per second, x6 represents the number of packets sent per second, and x7 represents the number of packets with errors when sending; F Storage represents the storage resource attributes of the cloud data center, FStorage = <x8,x9,x 10 , x 11 >, x8 represents the current memory usage, x9 represents the available free memory, x 10 Indicates the memory already used, x 11 Indicates the number of bytes swapped from disk or virtual memory to main memory per second; F I / o Indicates the I / O resource attributes of the cloud data center, F I / o = <x 12 , x 13 , x 14 >, x 12 Indicates the total number of times data is requested from the disk device per second, x 13 Indicates the number of read requests to the disk device per second, x 14 Indicates the number of write requests to the disk device per second;

[0017] The monitoring data set in step 1 is KF, KF = {F1, F2, ..., F i , …, F n}, n is the number of samples in the monitoring data set, i = 1, ..., n, F i is the fault feature information corresponding to the i-th sample.

[0018] Furthermore, step 3 determines the structure of the BP neural network for cloud data center fault detection and initializes the weight distribution of the training samples, specifically including:

[0019] Step 3-1, using BP neural network to perform fault detection on the cloud data center, the BP neural network adopts a three-layer BP neural network, including an input layer, a hidden layer, and an output layer;

[0020] Step 3-2, initialization of training data weights: Initialize the weight distribution of the training sample set Train to (w 11 , w 12 ,…,w 1n ),in N represents the number of sample data, Train = {(f1, l1), ..., (f i , l i ),…,(f n , l n )}(i=1,2,…,n),f i , f i is the instance data, l i ∈{1, -1} is the label, 1 means normal and -1 means fault.

[0021] Furthermore, for step 4, the corresponding relationship between the firefly behavior and the BP neural network optimization problem is shown in Table 1:

[0022] Table 1 Correspondence between firefly behavior and BP neural network optimization problem

[0023]

[0024] The specific steps of optimization are as follows:

[0025] Step 4.1, initialize the basic parameters of IFA, set the firefly population number Num, each firefly represents a candidate solution for the initial weights and thresholds of the BP neural network; set the maximum attraction of fireflies β0, β0 represents the attraction of fireflies at the maximum fluorescence brightness; set the light intensity absorption coefficient γ, which reflects the characteristics of fluorescence weakening due to the influence of distance and propagation medium, step size factor α, maximum number of iterations G max , maximum and minimum w max and the minimum weight w min ,The BP neural network error function is used as the fitness function. The smaller the fitness value, the better the quality of the BP neural network;

[0026] Step 4.2, respectively encoding the threshold of the output layer, the threshold of the hidden layer, the weights of the output layer and the hidden layer, and the weights of the hidden layer and the input layer of the BP neural network constructed in step 3 to form firefly individuals;

[0027] Step 4.3, randomly initialize the position of each firefly, that is, initialize the BP neural network weights and thresholds, calculate the individual fitness value of the firefly, and use the fitness value as the maximum fluorescence brightness I0 of each firefly;

[0028] Step 4.4, calculate the fluorescence brightness I of the firefly individual according to formula (2):

[0029]

[0030]

[0031] Among them, r ij is the distance between fireflies i and j, x ik , x jk are the kth components of fireflies i and j in space, respectively, and d is the dimension of space;

[0032] The mutual attraction β of individual fireflies is calculated as formula (4). The mutual attraction β determines the distance that the fireflies move.

[0033]

[0034] Where β0 is the attraction at the position of maximum fluorescence brightness, that is, the attraction at r = 0;

[0035] Step 4.5, if the fluorescence brightness of firefly j is greater than that of firefly i, then firefly i is attracted to firefly j and moves toward firefly j, that is, the BP neural network under the firefly j value scheme is better than the BP neural network under the firefly i value scheme, and the position of firefly i is updated according to formula (5);

[0036] x i (t+1)=wx i (t)+β(x j (t)-x i (t))+α(rand-1 / 2) (5)

[0037] Among them, x i (t+1) represents the position of firefly i after t+1 moves, α(rand-1 / 2) is the random perturbation term, α is the step size factor, rand is a random number uniformly distributed on [0, 1], and w is the inertia weight;

[0038] Step 4.6, after the position of the firefly individual is updated, recalculate the fitness value of each firefly individual, sort the fitness values ​​of the firefly individuals in ascending order, find the firefly individual with the smallest fitness value, that is, find the optimal initial weight and threshold value scheme of the BP neural network under this iteration, and then determine whether the current number of iterations has reached the maximum number of iterations G max Or the set search accuracy, if it is reached, go to step 4.7, otherwise go to step 4.4;

[0039] Step 4.7, taking the firefly individual with the smallest fitness value as the brightest firefly individual, that is, the optimal value scheme for the initial weight and threshold of the BP neural network;

[0040] Finally, the IFA optimization results are substituted into the BP neural network, the initial weights and thresholds are determined, and M IFA-BP neural networks are initialized.

[0041] Furthermore, IFA is used to calculate the initial optimal weights and thresholds of the neural network. After several iterations, all fireflies are close to the optimal point, the distance between fireflies gradually decreases, and the mutual attraction β also increases accordingly. This may cause the fireflies to move too far and miss the optimal point, resulting in the phenomenon of fireflies oscillating near the optimal point. In order to avoid this situation, the dynamic update formula of the inertia weight w is proposed as shown in formula (6):

[0042]

[0043] Among them, w max and w min are the maximum and minimum weights respectively, t is the current iteration number, G maxis the maximum number of iterations. By adjusting w(t), in the early stages of the search process, this method can maintain a high inertia weight to enhance global search and avoid falling into the local optimum too early. In the later stages, reducing the inertia weight can quickly transition to local search, improve local search performance, and avoid too fast position update speed, which leads to repeated oscillations and affects the accurate and rapid movement of individual fireflies to the optimal point.

[0044] Furthermore, in step 5, each IFA-BP neural network is trained using the training sample set, and the weight of each IFA-BP neural network is calculated according to the Adaboost algorithm, which specifically includes:

[0045] Step 5.1, use the data with weight distribution to train the IFA-BP neural network, and use the trained IFA-BP neural network as a weak classifier, denoted as g m (x), m = 1, ..., M, training data in the current weak classifier g m The error rate err under (x) m As shown in formula (7):

[0046]

[0047] Among them, w mi is the weight corresponding to the i-th training data, f i is the instance data, l i is the label corresponding to the instance data;

[0048] Step 5.2: Calculate the weak classifier g according to formula (8): m The weight of (x) is a m :

[0049]

[0050] Step 5.3: Update the weight distribution of the training data set according to formula (9):

[0051]

[0052] Among them, Z m is the normalization factor;

[0053]

[0054] Furthermore, in step 6, a cloud data center fault detection model based on IFA-BP-AdaBoost is dynamically constructed according to the weights of each IFA-BP neural network, as follows:

[0055] Each IFA-BP neural network weak classifier is transformed into mThe combined IFA-BP-AdaBoost strong classifier G is calculated as shown in formula (11). The obtained G is the cloud data center fault detection model based on IFA-BP-AdaBoost.

[0056]

[0057] Furthermore, in step 7, feature extraction is performed on the monitoring data collected by the target system and input into the trained IFA-BP-AdaBoost cloud data center fault detection model to analyze and determine whether a fault occurs. The specific steps are as follows:

[0058] Step 7.1: Use the extracted fault features as the BP neural network weak classifier g in step 5.1 m (x) (m=1, ..., M) input;

[0059] Step 7.2, and according to step 6 Merge the results of each weak classifier;

[0060] Step 7.3, determine whether a fault occurs based on the result of G. If the result is 1, it means it is normal, and if the result is -1, it means a fault.

[0061] Compared with the prior art, the present invention has the following significant advantages: a cloud data center fault detection model based on IFA-BP-Adaboost is proposed, the monitoring data is feature extracted by principal component analysis, the BP neural network is optimized by IFA to obtain the initial optimal weights and thresholds, and multiple IFA-BP neural networks are reasonably combined by weights using the Adaboost algorithm to form a high-precision fault detection model. This method can effectively overcome the defects of unstable output and overfitting of the BP neural network and improve the detection accuracy.

[0062] The present invention is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of the cloud data center fault detection method based on IFA-BP-Adaboost of the present invention.

[0064] Figure 2 This is a framework diagram of the cloud data center detection model based on IFA-BP-Adaboost of the present invention.

[0065] Figure 3 This is a diagram of the verification results of an embodiment of the present invention.

[0066] Figure 4 4 is a comparison chart of the detection accuracy experimental results of the embodiments of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0068] In one embodiment, the present invention is further verified and described. In this embodiment, a cloud data center is built based on OpenStack, with a total of 4 nodes, 1 control node, 2 computing nodes and 1 storage node. The Linux monitoring tool package sysstat is used to record monitoring data every 10 seconds on the built cloud data center.

[0069] refer to Figure 1 The cloud data center fault detection method based on the improved firefly algorithm to optimize BP-Adaboost of the present invention comprises the following steps:

[0070] Step 1: monitor the operating status of the cloud data center system, collect fault feature information, and obtain a monitoring data set; the fault feature information F is a four-tuple:

[0071] F=(F cpu ,F Networks ,F Storage ,F I / O ) (12)

[0072] Among them, F cpu represents the computing resource attributes of the cloud data center, F cpu =<x1,x2,x3> , x1 represents the CPU usage at the user level, x2 represents the CPU usage at the system level, and x3 represents the idle CPU percentage; F Networks represents the network resource attributes of the cloud data center, F Networks =<x4,x5,x6,x7> , x4 represents the number of packets with errors when receiving, x5 represents the number of packets received per second, x6 represents the number of packets sent per second, and x7 represents the number of packets with errors when sending; F Storage represents the storage resource attributes of the cloud data center, F Storage = <x8,x9,x 10 , x 11 >, x8 represents the current memory usage, x9 represents the available free memory, x 10 Indicates the memory already used, x 11 Indicates the number of bytes swapped from disk or virtual memory to main memory per second; F I / o Indicates the I / O resource attributes of the cloud data center, F I / o = <x 12 , x13 , x 14 >, x 12 Indicates the total number of times data is requested from the disk device per second, x 13 Indicates the number of read requests to the disk device per second, x 14 Indicates the number of write requests to the disk device per second;

[0073] The monitoring data set in step 1 is KF, KF = {F1, F2, ..., F i , …, F n}, n is the number of samples in the monitoring data set, i = 1, ..., n, F i is the fault feature information corresponding to the ith sample. In this embodiment, 8000 pieces of data are used in common, that is, n=8000, of which 4000 pieces are normal data samples and 4000 pieces are abnormal data samples.

[0074] Step 2: Use principal component analysis to extract features from the monitoring data set and obtain the sample data set Data = {(f1, l1), ..., (f i , l i ),…,(f n , l n )}, where l i ∈{1, -1} is the label, 1 indicates normal, -1 indicates fault, the n samples are randomly shuffled, 70% of the sample data set Data is taken as the training data set Train, and the remaining 30% is the test data set Test training sample set.

[0075] Step 3: determine the structure of the BP neural network for cloud data center fault detection and initialize the weight distribution of the training samples; specifically, include:

[0076] Step 3-1, using BP neural network to perform fault detection on the cloud data center, the BP neural network adopts a three-layer BP neural network, including an input layer, a hidden layer, and an output layer;

[0077] Step 3-2, initialization of training data weights: Initialize the weight distribution of the training sample set Train as (w 11 , w 12 ,…,w 1n ),in N represents the number of sample data, where N=5600.

[0078] Step 4: Use the improved firefly algorithm IFA to calculate the initial optimal weights and thresholds of the BP neural network and construct several IFA-BP neural networks. The specific steps are as follows:

[0079] Step 4.1, initialize the basic parameters of IFA, set the number of firefly population Num, each firefly represents a candidate solution of the initial weight and threshold of the BP neural network; set the maximum attraction of fireflies β0, β0 represents the attraction of fireflies at the maximum fluorescence brightness; set the light intensity absorption coefficient γ = 1, which reflects the characteristic that fluorescence is weakened by distance and propagation medium, step factor α = 0.2, and the maximum number of iterations G max , the maximum weight w max and the minimum weight w min , BP neural network error function is used as the fitness function;

[0080] Step 4.2, respectively encoding the threshold of the output layer, the threshold of the hidden layer, the weights of the output layer and the hidden layer, and the weights of the hidden layer and the input layer of the BP neural network constructed in step 3 to form firefly individuals;

[0081] Step 4.3, randomly initialize the position of each firefly, that is, initialize the BP neural network weights and thresholds, calculate the individual fitness value of the firefly, and use the fitness value as the maximum fluorescence brightness I0 of each firefly;

[0082] Step 4.4, calculate the fluorescence brightness I of the firefly individual according to formula (13):

[0083]

[0084]

[0085] where r ij is the distance between fireflies i and j, x ik , x jk are the kth components of fireflies i and j in space, respectively, and d is the dimension of space;

[0086] The mutual attraction β of individual fireflies is calculated as shown in formula (15). The mutual attraction β determines the distance that the fireflies move;

[0087]

[0088] Where β0 is the attraction at the position of maximum fluorescence brightness, that is, the attraction at r = 0;

[0089] Step 4.5, if the fluorescence brightness of firefly j is greater than that of firefly i, then firefly i is attracted to firefly j and moves toward firefly j, that is, the BP neural network under the firefly j value scheme is better than the BP neural network under the firefly i value scheme, and the position of firefly i is updated according to formula (16);

[0090] x i (t+1)=wxi (t)+β(x j (t)-x i (t))+α(rand-1 / 2) (16)

[0091] Among them, x i (t+1) represents the position of firefly i after t+1 moves, α(rand-1 / 2) is the random perturbation term, α is the step size factor, rand is a random number uniformly distributed on [0, 1], w is the inertia weight, and the calculation formula is:

[0092]

[0093] Among them, w max and w min are the maximum and minimum weights respectively, t is the current iteration number, G max is the maximum number of iterations.

[0094] Step 4.6, after the position of the firefly individual is updated, recalculate the fitness value of each firefly individual, sort the fitness values ​​of the firefly individuals in ascending order, find the firefly individual with the smallest fitness value, that is, find the optimal initial weight and threshold value scheme of the BP neural network under this iteration, and then determine whether the current number of iterations has reached the maximum number of iterations G max Or the set search accuracy, if it is reached, go to step 4.7, otherwise go to step 4.4;

[0095] Step 4.7, taking the firefly individual with the smallest fitness value as the brightest firefly individual, that is, the optimal value scheme for the initial weight and threshold of the BP neural network;

[0096] Finally, the IFA optimization results are substituted into the BP neural network, the initial weights and thresholds are determined, and M IFA-BP neural networks are initialized.

[0097] Step 5: Use the training sample set to train each IFA-BP neural network, and calculate the weight of each IFA-BP neural network according to the Adaboost algorithm; specifically including:

[0098] Step 5.1, use the data with weight distribution to train the IFA-BP neural network, and use the trained IFA-BP neural network as a weak classifier, denoted as g m (x), m = 1, ..., M, training data in the current weak classifier g m The error rate err under (x) m As shown in formula (18):

[0099]

[0100] Among them, wmi is the weight corresponding to the i-th training data, f i is the instance data, l i is the label corresponding to the instance data;

[0101] Step 5.2: Calculate the weak classifier g according to formula (19): m The weight of (x) is a m :

[0102]

[0103] Step 5.3, update the weight distribution of the training data set according to formula (20):

[0104]

[0105] Among them, Z m is the normalization factor;

[0106]

[0107] Step 6: dynamically construct a cloud data center fault detection model based on IFA-BP-AdaBoost according to the weights of each IFA-BP neural network; the details are as follows:

[0108] Each IFA-BP neural network weak classifier is transformed into m The combined IFA-BP-AdaBoost strong classifier G is calculated as shown in formula (22). The obtained G is the cloud data center fault detection model based on IFA-BP-AdaBoost;

[0109]

[0110] Figure 2 This is a framework diagram of the cloud data center fault detection model based on IFA-BP-Adaboost.

[0111] This example is programmed under the MATLAB R2019a platform, and the verification results are as follows Figure 3 As shown in Figure 2, the detection accuracy comparison of BP, BP-Adaboost and IFA-BP-Adaboost is as follows: Figure 4 shown.

[0112] It can be seen from the simulation results that the classification accuracy has been significantly improved after using AdaBoost to integrate the BP neural network optimized by IFA.

[0113] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A cloud data center fault detection method based on IFA-BP-Adaboost, characterized in that: The method comprises the following steps: Step 1: monitor the operating status of the cloud data center system, collect fault feature information, and obtain a monitoring data set; The fault feature information F is a four-tuple: F=(F cpu ,F Networks ,F Storage ,F I / O )(1) Among them, F cpu represents the computing resource attributes of the cloud data center, F cpu =<x1,x2,x3> , x1 represents the CPU usage at the user level, x2 represents the CPU usage at the system level, and x3 represents the idle CPU percentage; F Networks represents the network resource attributes of the cloud data center, F Networks =<x4,x5,x6,x7> , x4 represents the number of packets with errors when receiving, x5 represents the number of packets received per second, x6 represents the number of packets sent per second, and x7 represents the number of packets with errors when sending; F Storage represents the storage resource attributes of the cloud data center, F Storage = <x8,x9,x 10 ,x 11 >, x8 represents the current memory usage, x9 represents the available free memory, x 10 Indicates the memory already used, x 11 Indicates the number of bytes swapped from disk or virtual memory to main memory per second; F I / o Indicates the I / O resource attributes of the cloud data center, F I / o = <x 12 ,x 13 ,x 14 >, x 12 Indicates the total number of times data is requested from the disk device per second, x 13 Indicates the number of read requests to the disk device per second, x 14 Indicates the number of write requests to the disk device per second Step 2: Extract features from the monitoring data set using principal component analysis to obtain a training sample set; Step 3, determine the structure of the BP neural network for cloud data center fault detection and initialize the weight distribution of training samples; Step 4, using the improved firefly algorithm IFA to calculate the initial optimal weights and thresholds of the BP neural network, and constructing several IFA-BP neural networks; The corresponding relationship between firefly behavior and BP neural network optimization problem is shown in Table 1: Table 1 Correspondence between firefly behavior and BP neural network optimization problem The specific steps of optimization are as follows: Step 4.1, initialize the basic parameters of IFA, set the firefly population number Num, each firefly represents a candidate solution for the initial weights and thresholds of the BP neural network; set the maximum attraction of fireflies β0, β0 represents the attraction of fireflies at the maximum fluorescence brightness; set the light intensity absorption coefficient γ, which reflects the characteristics of fluorescence weakening due to the influence of distance and propagation medium, step size factor α, maximum number of iterations G max , maximum and minimum w max and the minimum weight w min ,The BP neural network error function is used as the fitness function. The smaller the fitness value, the better the quality of the BP neural network; Step 4.2, respectively encoding the threshold of the output layer, the threshold of the hidden layer, the weights of the output layer and the hidden layer, and the weights of the hidden layer and the input layer of the BP neural network constructed in step 3 to form firefly individuals; Step 4.3, randomly initialize the position of each firefly, that is, initialize the BP neural network weights and thresholds, calculate the individual fitness value of the firefly, and use the fitness value as the maximum fluorescence brightness I0 of each firefly; Step 4.4, calculate the fluorescence brightness I of the firefly individual according to formula (2): Among them, r ij is the distance between fireflies i and j, x ik ,x jk are the kth components of fireflies i and j in space, respectively, and d is the dimension of space; The mutual attraction β of individual fireflies is calculated as formula (4). The mutual attraction β determines the distance that the fireflies move. Where β0 is the attraction at the position of maximum fluorescence brightness, that is, the attraction at r = 0; Step 4.5, if the fluorescence brightness of firefly j is greater than that of firefly i, then firefly i is attracted to firefly j and moves toward firefly j, that is, the BP neural network under the firefly j value scheme is better than the BP neural network under the firefly i value scheme, and the position of firefly i is updated according to formula (5); x i (t+1)=wx i (t)+β(x j (t)-x i (t))+α(rand-1 / 2) (5) Among them, x i (t+1) represents the position of firefly i after t+1 moves, α(rand-1 / 2) is the random perturbation term, α is the step size factor, rand is a random number uniformly distributed on [0,1], and w is the inertia weight; Step 4.6, after the position of the firefly individual is updated, recalculate the fitness value of each firefly individual, sort the fitness values ​​of the firefly individuals in ascending order, find the firefly individual with the smallest fitness value, that is, find the optimal initial weight and threshold value scheme of the BP neural network under this iteration, and then determine whether the current number of iterations has reached the maximum number of iterations G max Or the set search accuracy, if it is reached, go to step 4.7, otherwise go to step 4.4; Step 4.7, taking the firefly individual with the smallest fitness value as the brightest firefly individual, that is, the optimal value scheme for the initial weight and threshold of the BP neural network; Finally, the IFA optimization results are substituted into the BP neural network, the initial weights and thresholds are determined, and M IFA-BP neural networks are initialized; Step 5: Use the training sample set to train each IFA-BP neural network, and calculate the weight of each IFA-BP neural network according to the Adaboost algorithm; specifically including: Step 5.1, use the data with weight distribution to train the IFA-BP neural network, and use the trained IFA-BP neural network as a weak classifier, denoted as g m (x), m = 1, ..., M, training data in the current weak classifier g m The error rate err under (x) m As shown in formula (6): Among them, w mi is the weight corresponding to the i-th training data, f i is the instance data, l i is the label corresponding to the instance data; Step 5.2: Calculate the weak classifier g according to formula (7): m The weight of (x) is a m : Step 5.3: Update the weight distribution of the training data set according to formula (8): Among them, Z m is the normalization factor; Step 6: dynamically construct a cloud data center fault detection model based on IFA-BP-AdaBoost according to the weights of each IFA-BP neural network; the details are as follows: Each IFA-BP neural network weak classifier is transformed into m The combined IFA-BP-AdaBoost strong classifier G is calculated as shown in formula (10). The obtained G is the cloud data center fault detection model based on IFA-BP-AdaBoost. Step 7: Extract features from the monitoring data collected by the target system and input them into the trained IFA-BP-AdaBoost cloud data center fault detection model to analyze and determine whether a fault occurs.

2. The cloud data center fault detection method based on IFA-BP-Adaboost according to claim 1, characterized in that: The monitoring data set in step 1 is KF, KF = {F1, F2, ..., F i ,…,F n }, n is the number of samples in the monitoring data set, i = 1, ..., n, F i is the fault feature information corresponding to the i-th sample.

3. The cloud data center fault detection method based on IFA-BP-Adaboost according to claim 1, characterized in that: Step 3 determines the structure of the BP neural network for cloud data center fault detection and initializes the weight distribution of the training samples, specifically including: Step 3-1, using BP neural network to perform fault detection on the cloud data center, the BP neural network adopts a three-layer BP neural network, including an input layer, a hidden layer, and an output layer; Step 3-2, initialization of training data weights: Initialize the weight distribution of the training sample set Train to (w 11 ,w 12 ,…,w 1N ),in N represents the number of sample data, Train={(f1,l1),…,(f i ,l i ),…,(f N ,l N )}(i=1,2,…,N),f i is the instance data, l i ∈{1,-1} is the label, 1 means normal and -1 means fault.

4. The cloud data center fault detection method based on IFA-BP-Adaboost according to claim 1, characterized in that: The dynamic update formula of inertia weight w is as follows: Among them, w max and w min are the maximum and minimum weights respectively, t is the current iteration number, G max is the maximum number of iterations.

5. The cloud data center fault detection method based on IFA-BP-Adaboost according to claim 1, characterized in that: In step 7, feature extraction is performed on the monitoring data collected by the target system and input into the trained IFA-BP-AdaBoost cloud data center fault detection model to analyze and determine whether a fault occurs. The specific steps are as follows: Step 7.1: Use the extracted fault features as the BP neural network weak classifier g in step 5.1 m (x), m = 1, ..., M input; Step 7.2, and according to step 6 Merge the results of each weak classifier; Step 7.3, determine whether a fault occurs based on the result of G. If the result is 1, it means it is normal, and if the result is -1, it means a fault.

Images